Halide  20.0.0
Halide compiler and libraries
Func.h
Go to the documentation of this file.
1 #ifndef HALIDE_FUNC_H
2 #define HALIDE_FUNC_H
3 
4 /** \file
5  *
6  * Defines Func - the front-end handle on a halide function, and related classes.
7  */
8 
9 #include "Argument.h"
10 #include "Expr.h"
11 #include "JITModule.h"
12 #include "Module.h"
13 #include "Param.h"
14 #include "Pipeline.h"
15 #include "RDom.h"
16 #include "Target.h"
17 #include "Tuple.h"
18 #include "Var.h"
19 
20 #include <map>
21 #include <utility>
22 
23 namespace Halide {
24 
25 class OutputImageParam;
26 
27 /** A class that can represent Vars or RVars. Used for reorder calls
28  * which can accept a mix of either. */
29 struct VarOrRVar {
30  VarOrRVar(const std::string &n, bool r)
31  : var(n), rvar(n), is_rvar(r) {
32  }
33  VarOrRVar(const Var &v)
34  : var(v), is_rvar(false) {
35  }
36  VarOrRVar(const RVar &r)
37  : rvar(r), is_rvar(true) {
38  }
39  VarOrRVar(const RDom &r)
40  : rvar(RVar(r)), is_rvar(true) {
41  }
42  template<int N>
44  : var(u), is_rvar(false) {
45  }
46 
47  const std::string &name() const {
48  if (is_rvar) {
49  return rvar.name();
50  } else {
51  return var.name();
52  }
53  }
54 
57  bool is_rvar;
58 };
59 
60 class ImageParam;
61 
62 namespace Internal {
63 class Function;
64 struct Split;
65 struct StorageDim;
66 } // namespace Internal
67 
68 /** A single definition of a Func. May be a pure or update definition. */
69 class Stage {
70  /** Reference to the Function this stage (or definition) belongs to. */
71  Internal::Function function;
72  Internal::Definition definition;
73  /** Indicate which stage the definition belongs to (0 for initial
74  * definition, 1 for first update, etc.). */
75  size_t stage_index;
76  /** Pure Vars of the Function (from the init definition). */
77  std::vector<Var> dim_vars;
78 
79  void set_dim_type(const VarOrRVar &var, Internal::ForType t);
80  void set_dim_device_api(const VarOrRVar &var, DeviceAPI device_api);
81  void split(const std::string &old, const std::string &outer, const std::string &inner,
82  const Expr &factor, bool exact, TailStrategy tail);
83  void remove(const std::string &var);
84  Stage &purify(const VarOrRVar &old_name, const VarOrRVar &new_name);
85 
86  const std::vector<Internal::StorageDim> &storage_dims() const {
87  return function.schedule().storage_dims();
88  }
89 
90  Stage &compute_with(LoopLevel loop_level, const std::map<std::string, LoopAlignStrategy> &align);
91 
92 public:
93  Stage(Internal::Function f, Internal::Definition d, size_t stage_index)
94  : function(std::move(f)), definition(std::move(d)), stage_index(stage_index) {
95  internal_assert(definition.defined());
96 
97  dim_vars.reserve(function.args().size());
98  for (const auto &arg : function.args()) {
99  dim_vars.emplace_back(arg);
100  }
101  internal_assert(definition.args().size() == dim_vars.size());
102  }
103 
104  /** Return the current StageSchedule associated with this Stage. For
105  * introspection only: to modify schedule, use the Func interface. */
107  return definition.schedule();
108  }
109 
110  /** Return a string describing the current var list taking into
111  * account all the splits, reorders, and tiles. */
112  std::string dump_argument_list() const;
113 
114  /** Return the name of this stage, e.g. "f.update(2)" */
115  std::string name() const;
116 
117  /** Calling rfactor() on an associative update definition a Func will split
118  * the update into an intermediate which computes the partial results and
119  * replaces the current update definition with a new definition which merges
120  * the partial results. If called on a init/pure definition, this will
121  * throw an error. rfactor() will automatically infer the associative reduction
122  * operator and identity of the operator. If it can't prove the operation
123  * is associative or if it cannot find an identity for that operator, this
124  * will throw an error. In addition, commutativity of the operator is required
125  * if rfactor() is called on the inner dimension but excluding the outer
126  * dimensions.
127  *
128  * rfactor() takes as input 'preserved', which is a list of <RVar, Var> pairs.
129  * The rvars not listed in 'preserved' are removed from the original Func and
130  * are lifted to the intermediate Func. The remaining rvars (the ones in
131  * 'preserved') are made pure in the intermediate Func. The intermediate Func's
132  * update definition inherits all scheduling directives (e.g. split,fuse, etc.)
133  * applied to the original Func's update definition. The loop order of the
134  * intermediate Func's update definition is the same as the original, although
135  * the RVars in 'preserved' are replaced by the new pure Vars. The loop order of the
136  * intermediate Func's init definition from innermost to outermost is the args'
137  * order of the original Func's init definition followed by the new pure Vars.
138  *
139  * The intermediate Func also inherits storage order from the original Func
140  * with the new pure Vars added to the outermost.
141  *
142  * For example, f.update(0).rfactor({{r.y, u}}) would rewrite a pipeline like this:
143  \code
144  f(x, y) = 0;
145  f(x, y) += g(r.x, r.y);
146  \endcode
147  * into a pipeline like this:
148  \code
149  f_intm(x, y, u) = 0;
150  f_intm(x, y, u) += g(r.x, u);
151 
152  f(x, y) = 0;
153  f(x, y) += f_intm(x, y, r.y);
154  \endcode
155  *
156  * This has a variety of uses. You can use it to split computation of an associative reduction:
157  \code
158  f(x, y) = 10;
159  RDom r(0, 96);
160  f(x, y) = max(f(x, y), g(x, y, r.x));
161  f.update(0).split(r.x, rxo, rxi, 8).reorder(y, x).parallel(x);
162  f.update(0).rfactor({{rxo, u}}).compute_root().parallel(u).update(0).parallel(u);
163  \endcode
164  *
165  *, which is equivalent to:
166  \code
167  parallel for u = 0 to 11:
168  for y:
169  for x:
170  f_intm(x, y, u) = -inf
171  parallel for x:
172  for y:
173  parallel for u = 0 to 11:
174  for rxi = 0 to 7:
175  f_intm(x, y, u) = max(f_intm(x, y, u), g(8*u + rxi))
176  for y:
177  for x:
178  f(x, y) = 10
179  parallel for x:
180  for y:
181  for rxo = 0 to 11:
182  f(x, y) = max(f(x, y), f_intm(x, y, rxo))
183  \endcode
184  *
185  */
186  // @{
187  Func rfactor(std::vector<std::pair<RVar, Var>> preserved);
188  Func rfactor(const RVar &r, const Var &v);
189  // @}
190 
191  /** Schedule the iteration over this stage to be fused with another
192  * stage 's' from outermost loop to a given LoopLevel. 'this' stage will
193  * be computed AFTER 's' in the innermost fused dimension. There should not
194  * be any dependencies between those two fused stages. If either of the
195  * stages being fused is a stage of an extern Func, this will throw an error.
196  *
197  * Note that the two stages that are fused together should have the same
198  * exact schedule from the outermost to the innermost fused dimension, and
199  * the stage we are calling compute_with on should not have specializations,
200  * e.g. f2.compute_with(f1, x) is allowed only if f2 has no specializations.
201  *
202  * Also, if a producer is desired to be computed at the fused loop level,
203  * the function passed to the compute_at() needs to be the "parent". Consider
204  * the following code:
205  \code
206  input(x, y) = x + y;
207  f(x, y) = input(x, y);
208  f(x, y) += 5;
209  g(x, y) = x - y;
210  g(x, y) += 10;
211  f.compute_with(g, y);
212  f.update().compute_with(g.update(), y);
213  \endcode
214  *
215  * To compute 'input' at the fused loop level at dimension y, we specify
216  * input.compute_at(g, y) instead of input.compute_at(f, y) since 'g' is
217  * the "parent" for this fused loop (i.e. 'g' is computed first before 'f'
218  * is computed). On the other hand, to compute 'input' at the innermost
219  * dimension of 'f', we specify input.compute_at(f, x) instead of
220  * input.compute_at(g, x) since the x dimension of 'f' is not fused
221  * (only the y dimension is).
222  *
223  * Given the constraints, this has a variety of uses. Consider the
224  * following code:
225  \code
226  f(x, y) = x + y;
227  g(x, y) = x - y;
228  h(x, y) = f(x, y) + g(x, y);
229  f.compute_root();
230  g.compute_root();
231  f.split(x, xo, xi, 8);
232  g.split(x, xo, xi, 8);
233  g.compute_with(f, xo);
234  \endcode
235  *
236  * This is equivalent to:
237  \code
238  for y:
239  for xo:
240  for xi:
241  f(8*xo + xi) = (8*xo + xi) + y
242  for xi:
243  g(8*xo + xi) = (8*xo + xi) - y
244  for y:
245  for x:
246  h(x, y) = f(x, y) + g(x, y)
247  \endcode
248  *
249  * The size of the dimensions of the stages computed_with do not have
250  * to match. Consider the following code where 'g' is half the size of 'f':
251  \code
252  Image<int> f_im(size, size), g_im(size/2, size/2);
253  input(x, y) = x + y;
254  f(x, y) = input(x, y);
255  g(x, y) = input(2*x, 2*y);
256  g.compute_with(f, y);
257  input.compute_at(f, y);
258  Pipeline({f, g}).realize({f_im, g_im});
259  \endcode
260  *
261  * This is equivalent to:
262  \code
263  for y = 0 to size-1:
264  for x = 0 to size-1:
265  input(x, y) = x + y;
266  for x = 0 to size-1:
267  f(x, y) = input(x, y)
268  for x = 0 to size/2-1:
269  if (y < size/2-1):
270  g(x, y) = input(2*x, 2*y)
271  \endcode
272  *
273  * 'align' specifies how the loop iteration of each dimension of the
274  * two stages being fused should be aligned in the fused loop nests
275  * (see LoopAlignStrategy for options). Consider the following loop nests:
276  \code
277  for z = f_min_z to f_max_z:
278  for y = f_min_y to f_max_y:
279  for x = f_min_x to f_max_x:
280  f(x, y, z) = x + y + z
281  for z = g_min_z to g_max_z:
282  for y = g_min_y to g_max_y:
283  for x = g_min_x to g_max_x:
284  g(x, y, z) = x - y - z
285  \endcode
286  *
287  * If no alignment strategy is specified, the following loop nest will be
288  * generated:
289  \code
290  for z = min(f_min_z, g_min_z) to max(f_max_z, g_max_z):
291  for y = min(f_min_y, g_min_y) to max(f_max_y, g_max_y):
292  for x = f_min_x to f_max_x:
293  if (f_min_z <= z <= f_max_z):
294  if (f_min_y <= y <= f_max_y):
295  f(x, y, z) = x + y + z
296  for x = g_min_x to g_max_x:
297  if (g_min_z <= z <= g_max_z):
298  if (g_min_y <= y <= g_max_y):
299  g(x, y, z) = x - y - z
300  \endcode
301  *
302  * Instead, these alignment strategies:
303  \code
304  g.compute_with(f, y, {{z, LoopAlignStrategy::AlignStart}, {y, LoopAlignStrategy::AlignEnd}});
305  \endcode
306  * will produce the following loop nest:
307  \code
308  f_loop_min_z = f_min_z
309  f_loop_max_z = max(f_max_z, (f_min_z - g_min_z) + g_max_z)
310  for z = f_min_z to f_loop_max_z:
311  f_loop_min_y = min(f_min_y, (f_max_y - g_max_y) + g_min_y)
312  f_loop_max_y = f_max_y
313  for y = f_loop_min_y to f_loop_max_y:
314  for x = f_min_x to f_max_x:
315  if (f_loop_min_z <= z <= f_loop_max_z):
316  if (f_loop_min_y <= y <= f_loop_max_y):
317  f(x, y, z) = x + y + z
318  for x = g_min_x to g_max_x:
319  g_shift_z = g_min_z - f_loop_min_z
320  g_shift_y = g_max_y - f_loop_max_y
321  if (g_min_z <= (z + g_shift_z) <= g_max_z):
322  if (g_min_y <= (y + g_shift_y) <= g_max_y):
323  g(x, y + g_shift_y, z + g_shift_z) = x - (y + g_shift_y) - (z + g_shift_z)
324  \endcode
325  *
326  * LoopAlignStrategy::AlignStart on dimension z will shift the loop iteration
327  * of 'g' at dimension z so that its starting value matches that of 'f'.
328  * Likewise, LoopAlignStrategy::AlignEnd on dimension y will shift the loop
329  * iteration of 'g' at dimension y so that its end value matches that of 'f'.
330  */
331  // @{
332  Stage &compute_with(LoopLevel loop_level, const std::vector<std::pair<VarOrRVar, LoopAlignStrategy>> &align);
334  Stage &compute_with(const Stage &s, const VarOrRVar &var, const std::vector<std::pair<VarOrRVar, LoopAlignStrategy>> &align);
336  // @}
337 
338  /** Scheduling calls that control how the domain of this stage is
339  * traversed. See the documentation for Func for the meanings. */
340  // @{
341 
342  Stage &split(const VarOrRVar &old, const VarOrRVar &outer, const VarOrRVar &inner, const Expr &factor, TailStrategy tail = TailStrategy::Auto);
343  Stage &fuse(const VarOrRVar &inner, const VarOrRVar &outer, const VarOrRVar &fused);
344  Stage &serial(const VarOrRVar &var);
345  Stage &parallel(const VarOrRVar &var);
346  Stage &vectorize(const VarOrRVar &var);
347  Stage &unroll(const VarOrRVar &var);
348  Stage &parallel(const VarOrRVar &var, const Expr &task_size, TailStrategy tail = TailStrategy::Auto);
349  Stage &vectorize(const VarOrRVar &var, const Expr &factor, TailStrategy tail = TailStrategy::Auto);
350  Stage &unroll(const VarOrRVar &var, const Expr &factor, TailStrategy tail = TailStrategy::Auto);
351  Stage &partition(const VarOrRVar &var, Partition partition_policy);
353  Stage &never_partition(const std::vector<VarOrRVar> &vars);
355  Stage &always_partition(const std::vector<VarOrRVar> &vars);
356 
357  Stage &tile(const VarOrRVar &x, const VarOrRVar &y,
358  const VarOrRVar &xo, const VarOrRVar &yo,
359  const VarOrRVar &xi, const VarOrRVar &yi, const Expr &xfactor, const Expr &yfactor,
361  Stage &tile(const VarOrRVar &x, const VarOrRVar &y,
362  const VarOrRVar &xi, const VarOrRVar &yi,
363  const Expr &xfactor, const Expr &yfactor,
365  Stage &tile(const std::vector<VarOrRVar> &previous,
366  const std::vector<VarOrRVar> &outers,
367  const std::vector<VarOrRVar> &inners,
368  const std::vector<Expr> &factors,
369  const std::vector<TailStrategy> &tails);
370  Stage &tile(const std::vector<VarOrRVar> &previous,
371  const std::vector<VarOrRVar> &outers,
372  const std::vector<VarOrRVar> &inners,
373  const std::vector<Expr> &factors,
375  Stage &tile(const std::vector<VarOrRVar> &previous,
376  const std::vector<VarOrRVar> &inners,
377  const std::vector<Expr> &factors,
379  Stage &reorder(const std::vector<VarOrRVar> &vars);
380 
381  template<typename... Args>
382  HALIDE_NO_USER_CODE_INLINE typename std::enable_if<Internal::all_are_convertible<VarOrRVar, Args...>::value, Stage &>::type
383  reorder(const VarOrRVar &x, const VarOrRVar &y, Args &&...args) {
384  std::vector<VarOrRVar> collected_args{x, y, std::forward<Args>(args)...};
385  return reorder(collected_args);
386  }
387 
388  template<typename... Args>
389  HALIDE_NO_USER_CODE_INLINE typename std::enable_if<Internal::all_are_convertible<VarOrRVar, Args...>::value, Stage &>::type
390  never_partition(const VarOrRVar &x, Args &&...args) {
391  std::vector<VarOrRVar> collected_args{x, std::forward<Args>(args)...};
392  return never_partition(collected_args);
393  }
394 
395  template<typename... Args>
396  HALIDE_NO_USER_CODE_INLINE typename std::enable_if<Internal::all_are_convertible<VarOrRVar, Args...>::value, Stage &>::type
397  always_partition(const VarOrRVar &x, Args &&...args) {
398  std::vector<VarOrRVar> collected_args{x, std::forward<Args>(args)...};
399  return always_partition(collected_args);
400  }
401 
402  Stage &rename(const VarOrRVar &old_name, const VarOrRVar &new_name);
403  Stage specialize(const Expr &condition);
404  void specialize_fail(const std::string &message);
405 
406  Stage &gpu_threads(const VarOrRVar &thread_x, DeviceAPI device_api = DeviceAPI::Default_GPU);
407  Stage &gpu_threads(const VarOrRVar &thread_x, const VarOrRVar &thread_y, DeviceAPI device_api = DeviceAPI::Default_GPU);
408  Stage &gpu_threads(const VarOrRVar &thread_x, const VarOrRVar &thread_y, const VarOrRVar &thread_z, DeviceAPI device_api = DeviceAPI::Default_GPU);
409 
410  Stage &gpu_lanes(const VarOrRVar &thread_x, DeviceAPI device_api = DeviceAPI::Default_GPU);
411 
413 
415  Stage &gpu_blocks(const VarOrRVar &block_x, const VarOrRVar &block_y, DeviceAPI device_api = DeviceAPI::Default_GPU);
416  Stage &gpu_blocks(const VarOrRVar &block_x, const VarOrRVar &block_y, const VarOrRVar &block_z, DeviceAPI device_api = DeviceAPI::Default_GPU);
417 
418  Stage &gpu(const VarOrRVar &block_x, const VarOrRVar &thread_x, DeviceAPI device_api = DeviceAPI::Default_GPU);
419  Stage &gpu(const VarOrRVar &block_x, const VarOrRVar &block_y,
420  const VarOrRVar &thread_x, const VarOrRVar &thread_y,
421  DeviceAPI device_api = DeviceAPI::Default_GPU);
422  Stage &gpu(const VarOrRVar &block_x, const VarOrRVar &block_y, const VarOrRVar &block_z,
423  const VarOrRVar &thread_x, const VarOrRVar &thread_y, const VarOrRVar &thread_z,
424  DeviceAPI device_api = DeviceAPI::Default_GPU);
425 
426  Stage &gpu_tile(const VarOrRVar &x, const VarOrRVar &bx, const VarOrRVar &tx, const Expr &x_size,
428  DeviceAPI device_api = DeviceAPI::Default_GPU);
429 
430  Stage &gpu_tile(const VarOrRVar &x, const VarOrRVar &tx, const Expr &x_size,
432  DeviceAPI device_api = DeviceAPI::Default_GPU);
433  Stage &gpu_tile(const VarOrRVar &x, const VarOrRVar &y,
434  const VarOrRVar &bx, const VarOrRVar &by,
435  const VarOrRVar &tx, const VarOrRVar &ty,
436  const Expr &x_size, const Expr &y_size,
438  DeviceAPI device_api = DeviceAPI::Default_GPU);
439 
440  Stage &gpu_tile(const VarOrRVar &x, const VarOrRVar &y,
441  const VarOrRVar &tx, const VarOrRVar &ty,
442  const Expr &x_size, const Expr &y_size,
444  DeviceAPI device_api = DeviceAPI::Default_GPU);
445 
446  Stage &gpu_tile(const VarOrRVar &x, const VarOrRVar &y, const VarOrRVar &z,
447  const VarOrRVar &bx, const VarOrRVar &by, const VarOrRVar &bz,
448  const VarOrRVar &tx, const VarOrRVar &ty, const VarOrRVar &tz,
449  const Expr &x_size, const Expr &y_size, const Expr &z_size,
451  DeviceAPI device_api = DeviceAPI::Default_GPU);
452  Stage &gpu_tile(const VarOrRVar &x, const VarOrRVar &y, const VarOrRVar &z,
453  const VarOrRVar &tx, const VarOrRVar &ty, const VarOrRVar &tz,
454  const Expr &x_size, const Expr &y_size, const Expr &z_size,
456  DeviceAPI device_api = DeviceAPI::Default_GPU);
457 
459  Stage &atomic(bool override_associativity_test = false);
460 
462 
463  Stage &prefetch(const Func &f, const VarOrRVar &at, const VarOrRVar &from, Expr offset = 1,
465  Stage &prefetch(const Parameter &param, const VarOrRVar &at, const VarOrRVar &from, Expr offset = 1,
467  template<typename T>
468  Stage &prefetch(const T &image, const VarOrRVar &at, const VarOrRVar &from, Expr offset = 1,
470  return prefetch(image.parameter(), at, from, std::move(offset), strategy);
471  }
472  // @}
473 
474  /** Assert that this stage has intentionally been given no schedule, and
475  * suppress the warning about unscheduled update definitions that would
476  * otherwise fire. This counts as a schedule, so calling this twice on the
477  * same Stage will fail the assertion. */
478  void unscheduled();
479 };
480 
481 // For backwards compatibility, keep the ScheduleHandle name.
483 
484 class FuncTupleElementRef;
485 
486 /** A fragment of front-end syntax of the form f(x, y, z), where x, y,
487  * z are Vars or Exprs. If could be the left hand side of a definition or
488  * an update definition, or it could be a call to a function. We don't know
489  * until we see how this object gets used.
490  */
491 class FuncRef {
492  Internal::Function func;
493  int implicit_placeholder_pos;
494  int implicit_count;
495  std::vector<Expr> args;
496  std::vector<Expr> args_with_implicit_vars(const std::vector<Expr> &e) const;
497 
498  /** Helper for function update by Tuple. If the function does not
499  * already have a pure definition, init_val will be used as RHS of
500  * each tuple element in the initial function definition. */
501  template<typename BinaryOp>
502  Stage func_ref_update(const Tuple &e, int init_val);
503 
504  /** Helper for function update by Expr. If the function does not
505  * already have a pure definition, init_val will be used as RHS in
506  * the initial function definition. */
507  template<typename BinaryOp>
508  Stage func_ref_update(Expr e, int init_val);
509 
510 public:
511  FuncRef(const Internal::Function &, const std::vector<Expr> &,
512  int placeholder_pos = -1, int count = 0);
513  FuncRef(Internal::Function, const std::vector<Var> &,
514  int placeholder_pos = -1, int count = 0);
515 
516  /** Use this as the left-hand-side of a definition or an update definition
517  * (see \ref RDom).
518  */
519  Stage operator=(const Expr &);
520 
521  /** Use this as the left-hand-side of a definition or an update definition
522  * for a Func with multiple outputs. */
524 
525  /** Define a stage that adds the given expression to this Func. If the
526  * expression refers to some RDom, this performs a sum reduction of the
527  * expression over the domain. If the function does not already have a
528  * pure definition, this sets it to zero.
529  */
530  // @{
534  // @}
535 
536  /** Define a stage that adds the negative of the given expression to this
537  * Func. If the expression refers to some RDom, this performs a sum reduction
538  * of the negative of the expression over the domain. If the function does
539  * not already have a pure definition, this sets it to zero.
540  */
541  // @{
545  // @}
546 
547  /** Define a stage that multiplies this Func by the given expression. If the
548  * expression refers to some RDom, this performs a product reduction of the
549  * expression over the domain. If the function does not already have a pure
550  * definition, this sets it to 1.
551  */
552  // @{
556  // @}
557 
558  /** Define a stage that divides this Func by the given expression.
559  * If the expression refers to some RDom, this performs a product
560  * reduction of the inverse of the expression over the domain. If the
561  * function does not already have a pure definition, this sets it to 1.
562  */
563  // @{
567  // @}
568 
569  /* Override the usual assignment operator, so that
570  * f(x, y) = g(x, y) defines f.
571  */
573 
574  /** Use this as a call to the function, and not the left-hand-side
575  * of a definition. Only works for single-output Funcs. */
576  operator Expr() const;
577 
578  /** When a FuncRef refers to a function that provides multiple
579  * outputs, you can access each output as an Expr using
580  * operator[].
581  */
583 
584  /** How many outputs does the function this refers to produce. */
585  size_t size() const;
586 
587  /** What function is this calling? */
588  Internal::Function function() const {
589  return func;
590  }
591 };
592 
593 /** Explicit overloads of min and max for FuncRef. These exist to
594  * disambiguate calls to min on FuncRefs when a user has pulled both
595  * Halide::min and std::min into their namespace. */
596 // @{
597 inline Expr min(const FuncRef &a, const FuncRef &b) {
598  return min(Expr(a), Expr(b));
599 }
600 inline Expr max(const FuncRef &a, const FuncRef &b) {
601  return max(Expr(a), Expr(b));
602 }
603 // @}
604 
605 /** A fragment of front-end syntax of the form f(x, y, z)[index], where x, y,
606  * z are Vars or Exprs. If could be the left hand side of an update
607  * definition, or it could be a call to a function. We don't know
608  * until we see how this object gets used.
609  */
611  FuncRef func_ref;
612  std::vector<Expr> args; // args to the function
613  int idx; // Index to function outputs
614 
615  /** Helper function that generates a Tuple where element at 'idx' is set
616  * to 'e' and the rests are undef. */
617  Tuple values_with_undefs(const Expr &e) const;
618 
619 public:
620  FuncTupleElementRef(const FuncRef &ref, const std::vector<Expr> &args, int idx);
621 
622  /** Use this as the left-hand-side of an update definition of Tuple
623  * component 'idx' of a Func (see \ref RDom). The function must
624  * already have an initial definition.
625  */
626  Stage operator=(const Expr &e);
627 
628  /** Define a stage that adds the given expression to Tuple component 'idx'
629  * of this Func. The other Tuple components are unchanged. If the expression
630  * refers to some RDom, this performs a sum reduction of the expression over
631  * the domain. The function must already have an initial definition.
632  */
633  Stage operator+=(const Expr &e);
634 
635  /** Define a stage that adds the negative of the given expression to Tuple
636  * component 'idx' of this Func. The other Tuple components are unchanged.
637  * If the expression refers to some RDom, this performs a sum reduction of
638  * the negative of the expression over the domain. The function must already
639  * have an initial definition.
640  */
641  Stage operator-=(const Expr &e);
642 
643  /** Define a stage that multiplies Tuple component 'idx' of this Func by
644  * the given expression. The other Tuple components are unchanged. If the
645  * expression refers to some RDom, this performs a product reduction of
646  * the expression over the domain. The function must already have an
647  * initial definition.
648  */
649  Stage operator*=(const Expr &e);
650 
651  /** Define a stage that divides Tuple component 'idx' of this Func by
652  * the given expression. The other Tuple components are unchanged.
653  * If the expression refers to some RDom, this performs a product
654  * reduction of the inverse of the expression over the domain. The function
655  * must already have an initial definition.
656  */
657  Stage operator/=(const Expr &e);
658 
659  /* Override the usual assignment operator, so that
660  * f(x, y)[index] = g(x, y) defines f.
661  */
663 
664  /** Use this as a call to Tuple component 'idx' of a Func, and not the
665  * left-hand-side of a definition. */
666  operator Expr() const;
667 
668  /** What function is this calling? */
669  Internal::Function function() const {
670  return func_ref.function();
671  }
672 
673  /** Return index to the function outputs. */
674  int index() const {
675  return idx;
676  }
677 };
678 
679 namespace Internal {
680 class IRMutator;
681 } // namespace Internal
682 
683 /** Helper class for identifying purpose of an Expr passed to memoize.
684  */
685 class EvictionKey {
686 protected:
688  friend class Func;
689 
690 public:
691  explicit EvictionKey(const Expr &expr = Expr())
692  : key(expr) {
693  }
694 };
695 
696 /** A halide function. This class represents one stage in a Halide
697  * pipeline, and is the unit by which we schedule things. By default
698  * they are aggressively inlined, so you are encouraged to make lots
699  * of little functions, rather than storing things in Exprs. */
700 class Func {
701 
702  /** A handle on the internal halide function that this
703  * represents */
704  Internal::Function func;
705 
706  /** When you make a reference to this function with fewer
707  * arguments than it has dimensions, the argument list is bulked
708  * up with 'implicit' vars with canonical names. This lets you
709  * pass around partially applied Halide functions. */
710  // @{
711  std::pair<int, int> add_implicit_vars(std::vector<Var> &) const;
712  std::pair<int, int> add_implicit_vars(std::vector<Expr> &) const;
713  // @}
714 
715  /** The imaging pipeline that outputs this Func alone. */
716  Pipeline pipeline_;
717 
718  /** Get the imaging pipeline that outputs this Func alone,
719  * creating it (and freezing the Func) if necessary. */
720  Pipeline pipeline();
721 
722  // Helper function for recursive reordering support
723  Func &reorder_storage(const std::vector<Var> &dims, size_t start);
724 
725  void invalidate_cache();
726 
727 public:
728  /** Declare a new undefined function with the given name */
729  explicit Func(const std::string &name);
730 
731  /** Declare a new undefined function with the given name.
732  * The function will be constrained to represent Exprs of required_type.
733  * If required_dims is not AnyDims, the function will be constrained to exactly
734  * that many dimensions. */
735  explicit Func(const Type &required_type, int required_dims, const std::string &name);
736 
737  /** Declare a new undefined function with the given name.
738  * If required_types is not empty, the function will be constrained to represent
739  * Tuples of the same arity and types. (If required_types is empty, there is no constraint.)
740  * If required_dims is not AnyDims, the function will be constrained to exactly
741  * that many dimensions. */
742  explicit Func(const std::vector<Type> &required_types, int required_dims, const std::string &name);
743 
744  /** Declare a new undefined function with an
745  * automatically-generated unique name */
746  Func();
747 
748  /** Declare a new function with an automatically-generated unique
749  * name, and define it to return the given expression (which may
750  * not contain free variables). */
751  explicit Func(const Expr &e);
752 
753  /** Construct a new Func to wrap an existing, already-define
754  * Function object. */
756 
757  /** Construct a new Func to wrap a Buffer. */
758  template<typename T, int Dims>
760  : Func() {
761  (*this)(_) = im(_);
762  }
763 
764  /** Evaluate this function over some rectangular domain and return
765  * the resulting buffer or buffers. Performs compilation if the
766  * Func has not previously been realized and compile_jit has not
767  * been called. If the final stage of the pipeline is on the GPU,
768  * data is copied back to the host before being returned. The
769  * returned Realization should probably be instantly converted to
770  * a Buffer class of the appropriate type. That is, do this:
771  *
772  \code
773  f(x) = sin(x);
774  Buffer<float> im = f.realize(...);
775  \endcode
776  *
777  * If your Func has multiple values, because you defined it using
778  * a Tuple, then casting the result of a realize call to a buffer
779  * or image will produce a run-time error. Instead you should do the
780  * following:
781  *
782  \code
783  f(x) = Tuple(x, sin(x));
784  Realization r = f.realize(...);
785  Buffer<int> im0 = r[0];
786  Buffer<float> im1 = r[1];
787  \endcode
788  *
789  * In Halide formal arguments of a computation are specified using
790  * Param<T> and ImageParam objects in the expressions defining the
791  * computation. Note that this method is not thread-safe, in that
792  * Param<T> and ImageParam are globals shared by all threads; to call
793  * jitted code in a thread-safe manner, use compile_to_callable() instead.
794  *
795  \code
796  Param<int32> p(42);
797  ImageParam img(Int(32), 1);
798  f(x) = img(x) + p;
799 
800  Buffer<int32_t) arg_img(10, 10);
801  <fill in arg_img...>
802 
803  Target t = get_jit_target_from_environment();
804  Buffer<int32_t> result = f.realize({10, 10}, t);
805  \endcode
806  *
807  * Alternatively, an initializer list can be used
808  * directly in the realize call to pass this information:
809  *
810  \code
811  Param<int32> p(42);
812  ImageParam img(Int(32), 1);
813  f(x) = img(x) + p;
814 
815  Buffer<int32_t) arg_img(10, 10);
816  <fill in arg_img...>
817 
818  Target t = get_jit_target_from_environment();
819  Buffer<int32_t> result = f.realize({10, 10}, t, { { p, 17 }, { img, arg_img } });
820  \endcode
821  *
822  * If the Func cannot be realized into a buffer of the given size
823  * due to scheduling constraints on scattering update definitions,
824  * it will be realized into a larger buffer of the minimum size
825  * possible, and a cropped view at the requested size will be
826  * returned. It is thus not safe to assume the returned buffers
827  * are contiguous in memory. This behavior can be disabled with
828  * the NoBoundsQuery target flag, in which case an error about
829  * writing out of bounds on the output buffer will trigger
830  * instead.
831  *
832  */
833  Realization realize(std::vector<int32_t> sizes = {}, const Target &target = Target());
834 
835  /** Same as above, but takes a custom user-provided context to be
836  * passed to runtime functions. This can be used to pass state to
837  * runtime overrides in a thread-safe manner. A nullptr context is
838  * legal, and is equivalent to calling the variant of realize
839  * that does not take a context. */
841  std::vector<int32_t> sizes = {},
842  const Target &target = Target());
843 
844  /** Evaluate this function into an existing allocated buffer or
845  * buffers. If the buffer is also one of the arguments to the
846  * function, strange things may happen, as the pipeline isn't
847  * necessarily safe to run in-place. If you pass multiple buffers,
848  * they must have matching sizes. This form of realize does *not*
849  * automatically copy data back from the GPU. */
851 
852  /** Same as above, but takes a custom user-provided context to be
853  * passed to runtime functions. This can be used to pass state to
854  * runtime overrides in a thread-safe manner. A nullptr context is
855  * legal, and is equivalent to calling the variant of realize
856  * that does not take a context. */
857  void realize(JITUserContext *context,
859  const Target &target = Target());
860 
861  /** For a given size of output, or a given output buffer,
862  * determine the bounds required of all unbound ImageParams
863  * referenced. Communicates the result by allocating new buffers
864  * of the appropriate size and binding them to the unbound
865  * ImageParams.
866  */
867  // @{
868  void infer_input_bounds(const std::vector<int32_t> &sizes,
869  const Target &target = get_jit_target_from_environment());
871  const Target &target = get_jit_target_from_environment());
872  // @}
873 
874  /** Versions of infer_input_bounds that take a custom user context
875  * to pass to runtime functions. */
876  // @{
878  const std::vector<int32_t> &sizes,
879  const Target &target = get_jit_target_from_environment());
882  const Target &target = get_jit_target_from_environment());
883  // @}
884  /** Statically compile this function to llvm bitcode, with the
885  * given filename (which should probably end in .bc), type
886  * signature, and C function name (which defaults to the same name
887  * as this halide function */
888  //@{
889  void compile_to_bitcode(const std::string &filename, const std::vector<Argument> &, const std::string &fn_name,
890  const Target &target = get_target_from_environment());
891  void compile_to_bitcode(const std::string &filename, const std::vector<Argument> &,
892  const Target &target = get_target_from_environment());
893  // @}
894 
895  /** Statically compile this function to llvm assembly, with the
896  * given filename (which should probably end in .ll), type
897  * signature, and C function name (which defaults to the same name
898  * as this halide function */
899  //@{
900  void compile_to_llvm_assembly(const std::string &filename, const std::vector<Argument> &, const std::string &fn_name,
901  const Target &target = get_target_from_environment());
902  void compile_to_llvm_assembly(const std::string &filename, const std::vector<Argument> &,
903  const Target &target = get_target_from_environment());
904  // @}
905 
906  /** Statically compile this function to an object file, with the
907  * given filename (which should probably end in .o or .obj), type
908  * signature, and C function name (which defaults to the same name
909  * as this halide function. You probably don't want to use this
910  * directly; call compile_to_static_library or compile_to_file instead. */
911  //@{
912  void compile_to_object(const std::string &filename, const std::vector<Argument> &, const std::string &fn_name,
913  const Target &target = get_target_from_environment());
914  void compile_to_object(const std::string &filename, const std::vector<Argument> &,
915  const Target &target = get_target_from_environment());
916  // @}
917 
918  /** Emit a header file with the given filename for this
919  * function. The header will define a function with the type
920  * signature given by the second argument, and a name given by the
921  * third. The name defaults to the same name as this halide
922  * function. You don't actually have to have defined this function
923  * yet to call this. You probably don't want to use this directly;
924  * call compile_to_static_library or compile_to_file instead. */
925  void compile_to_header(const std::string &filename, const std::vector<Argument> &, const std::string &fn_name = "",
926  const Target &target = get_target_from_environment());
927 
928  /** Statically compile this function to text assembly equivalent
929  * to the object file generated by compile_to_object. This is
930  * useful for checking what Halide is producing without having to
931  * disassemble anything, or if you need to feed the assembly into
932  * some custom toolchain to produce an object file (e.g. iOS) */
933  //@{
934  void compile_to_assembly(const std::string &filename, const std::vector<Argument> &, const std::string &fn_name,
935  const Target &target = get_target_from_environment());
936  void compile_to_assembly(const std::string &filename, const std::vector<Argument> &,
937  const Target &target = get_target_from_environment());
938  // @}
939 
940  /** Statically compile this function to C source code. This is
941  * useful for providing fallback code paths that will compile on
942  * many platforms. Vectorization will fail, and parallelization
943  * will produce serial code. */
944  void compile_to_c(const std::string &filename,
945  const std::vector<Argument> &,
946  const std::string &fn_name = "",
947  const Target &target = get_target_from_environment());
948 
949  /** Write out an internal representation of lowered code. Useful
950  * for analyzing and debugging scheduling. Can emit html or plain
951  * text. */
952  void compile_to_lowered_stmt(const std::string &filename,
953  const std::vector<Argument> &args,
954  StmtOutputFormat fmt = Text,
955  const Target &target = get_target_from_environment());
956 
957  /** Write out a conceptual representation of lowered code, before any parallel loop
958  * get factored out into separate functions, or GPU loops are offloaded to kernel code.r
959  * Useful for analyzing and debugging scheduling. Can emit html or plain text. */
960  void compile_to_conceptual_stmt(const std::string &filename,
961  const std::vector<Argument> &args,
962  StmtOutputFormat fmt = Text,
963  const Target &target = get_target_from_environment());
964 
965  /** Write out the loop nests specified by the schedule for this
966  * Function. Helpful for understanding what a schedule is
967  * doing. */
969 
970  /** Compile to object file and header pair, with the given
971  * arguments. The name defaults to the same name as this halide
972  * function.
973  */
974  void compile_to_file(const std::string &filename_prefix, const std::vector<Argument> &args,
975  const std::string &fn_name = "",
976  const Target &target = get_target_from_environment());
977 
978  /** Compile to static-library file and header pair, with the given
979  * arguments. The name defaults to the same name as this halide
980  * function.
981  */
982  void compile_to_static_library(const std::string &filename_prefix, const std::vector<Argument> &args,
983  const std::string &fn_name = "",
984  const Target &target = get_target_from_environment());
985 
986  /** Compile to static-library file and header pair once for each target;
987  * each resulting function will be considered (in order) via halide_can_use_target_features()
988  * at runtime, with the first appropriate match being selected for subsequent use.
989  * This is typically useful for specializations that may vary unpredictably by machine
990  * (e.g., SSE4.1/AVX/AVX2 on x86 desktop machines).
991  * All targets must have identical arch-os-bits.
992  */
993  void compile_to_multitarget_static_library(const std::string &filename_prefix,
994  const std::vector<Argument> &args,
995  const std::vector<Target> &targets);
996 
997  /** Like compile_to_multitarget_static_library(), except that the object files
998  * are all output as object files (rather than bundled into a static library).
999  *
1000  * `suffixes` is an optional list of strings to use for as the suffix for each object
1001  * file. If nonempty, it must be the same length as `targets`. (If empty, Target::to_string()
1002  * will be used for each suffix.)
1003  *
1004  * Note that if `targets.size()` > 1, the wrapper code (to select the subtarget)
1005  * will be generated with the filename `${filename_prefix}_wrapper.o`
1006  *
1007  * Note that if `targets.size()` > 1 and `no_runtime` is not specified, the runtime
1008  * will be generated with the filename `${filename_prefix}_runtime.o`
1009  */
1010  void compile_to_multitarget_object_files(const std::string &filename_prefix,
1011  const std::vector<Argument> &args,
1012  const std::vector<Target> &targets,
1013  const std::vector<std::string> &suffixes);
1014 
1015  /** Store an internal representation of lowered code as a self
1016  * contained Module suitable for further compilation. */
1017  Module compile_to_module(const std::vector<Argument> &args, const std::string &fn_name = "",
1018  const Target &target = get_target_from_environment());
1019 
1020  /** Compile and generate multiple target files with single call.
1021  * Deduces target files based on filenames specified in
1022  * output_files map.
1023  */
1024  void compile_to(const std::map<OutputFileType, std::string> &output_files,
1025  const std::vector<Argument> &args,
1026  const std::string &fn_name,
1027  const Target &target = get_target_from_environment());
1028 
1029  /** Eagerly jit compile the function to machine code. This
1030  * normally happens on the first call to realize. If you're
1031  * running your halide pipeline inside time-sensitive code and
1032  * wish to avoid including the time taken to compile a pipeline,
1033  * then you can call this ahead of time. Default is to use the Target
1034  * returned from Halide::get_jit_target_from_environment()
1035  */
1037 
1038  /** Get a struct containing the currently set custom functions
1039  * used by JIT. This can be mutated. Changes will take effect the
1040  * next time this Func is realized. */
1042 
1043  /** Eagerly jit compile the function to machine code and return a callable
1044  * struct that behaves like a function pointer. The calling convention
1045  * will exactly match that of an AOT-compiled version of this Func
1046  * with the same Argument list.
1047  */
1048  Callable compile_to_callable(const std::vector<Argument> &args,
1049  const Target &target = get_jit_target_from_environment());
1050 
1051  /** Add a custom pass to be used during lowering. It is run after
1052  * all other lowering passes. Can be used to verify properties of
1053  * the lowered Stmt, instrument it with extra code, or otherwise
1054  * modify it. The Func takes ownership of the pass, and will call
1055  * delete on it when the Func goes out of scope. So don't pass a
1056  * stack object, or share pass instances between multiple
1057  * Funcs. */
1058  template<typename T>
1060  // Template instantiate a custom deleter for this type, then
1061  // wrap in a lambda. The custom deleter lives in user code, so
1062  // that deletion is on the same heap as construction (I hate Windows).
1063  add_custom_lowering_pass(pass, [pass]() { delete_lowering_pass<T>(pass); });
1064  }
1065 
1066  /** Add a custom pass to be used during lowering, with the
1067  * function that will be called to delete it also passed in. Set
1068  * it to nullptr if you wish to retain ownership of the object. */
1069  void add_custom_lowering_pass(Internal::IRMutator *pass, std::function<void()> deleter);
1070 
1071  /** Remove all previously-set custom lowering passes */
1073 
1074  /** Get the custom lowering passes. */
1075  const std::vector<CustomLoweringPass> &custom_lowering_passes();
1076 
1077  /** When this function is compiled, include code that dumps its
1078  * values to a file after it is realized, for the purpose of
1079  * debugging.
1080  *
1081  * If filename ends in ".tif" or ".tiff" (case insensitive) the file
1082  * is in TIFF format and can be read by standard tools. Oherwise, the
1083  * file format is as follows:
1084  *
1085  * All data is in the byte-order of the target platform. First, a
1086  * 20 byte-header containing four 32-bit ints, giving the extents
1087  * of the first four dimensions. Dimensions beyond four are
1088  * folded into the fourth. Then, a fifth 32-bit int giving the
1089  * data type of the function. The typecodes are given by: float =
1090  * 0, double = 1, uint8_t = 2, int8_t = 3, uint16_t = 4, int16_t =
1091  * 5, uint32_t = 6, int32_t = 7, uint64_t = 8, int64_t = 9. The
1092  * data follows the header, as a densely packed array of the given
1093  * size and the given type. If given the extension .tmp, this file
1094  * format can be natively read by the program ImageStack. */
1095  void debug_to_file(const std::string &filename);
1096 
1097  /** The name of this function, either given during construction,
1098  * or automatically generated. */
1099  const std::string &name() const;
1100 
1101  /** Get the pure arguments. */
1102  std::vector<Var> args() const;
1103 
1104  /** The right-hand-side value of the pure definition of this
1105  * function. Causes an error if there's no pure definition, or if
1106  * the function is defined to return multiple values. */
1107  Expr value() const;
1108 
1109  /** The values returned by this function. An error if the function
1110  * has not been been defined. Returns a Tuple with one element for
1111  * functions defined to return a single value. */
1112  Tuple values() const;
1113 
1114  /** Does this function have at least a pure definition. */
1115  bool defined() const;
1116 
1117  /** Get the left-hand-side of the update definition. An empty
1118  * vector if there's no update definition. If there are
1119  * multiple update definitions for this function, use the
1120  * argument to select which one you want. */
1121  const std::vector<Expr> &update_args(int idx = 0) const;
1122 
1123  /** Get the right-hand-side of an update definition. An error if
1124  * there's no update definition. If there are multiple
1125  * update definitions for this function, use the argument to
1126  * select which one you want. */
1127  Expr update_value(int idx = 0) const;
1128 
1129  /** Get the right-hand-side of an update definition for
1130  * functions that returns multiple values. An error if there's no
1131  * update definition. Returns a Tuple with one element for
1132  * functions that return a single value. */
1133  Tuple update_values(int idx = 0) const;
1134 
1135  /** Get the RVars of the reduction domain for an update definition, if there is
1136  * one. */
1137  std::vector<RVar> rvars(int idx = 0) const;
1138 
1139  /** Does this function have at least one update definition? */
1141 
1142  /** How many update definitions does this function have? */
1144 
1145  /** Is this function an external stage? That is, was it defined
1146  * using define_extern? */
1147  bool is_extern() const;
1148 
1149  /** Add an extern definition for this Func. This lets you define a
1150  * Func that represents an external pipeline stage. You can, for
1151  * example, use it to wrap a call to an extern library such as
1152  * fftw. */
1153  // @{
1154  void define_extern(const std::string &function_name,
1155  const std::vector<ExternFuncArgument> &params, Type t,
1156  int dimensionality,
1158  DeviceAPI device_api = DeviceAPI::Host) {
1159  define_extern(function_name, params, t,
1160  Internal::make_argument_list(dimensionality), mangling,
1161  device_api);
1162  }
1163 
1164  void define_extern(const std::string &function_name,
1165  const std::vector<ExternFuncArgument> &params,
1166  const std::vector<Type> &types, int dimensionality,
1167  NameMangling mangling) {
1168  define_extern(function_name, params, types,
1169  Internal::make_argument_list(dimensionality), mangling);
1170  }
1171 
1172  void define_extern(const std::string &function_name,
1173  const std::vector<ExternFuncArgument> &params,
1174  const std::vector<Type> &types, int dimensionality,
1176  DeviceAPI device_api = DeviceAPI::Host) {
1177  define_extern(function_name, params, types,
1178  Internal::make_argument_list(dimensionality), mangling,
1179  device_api);
1180  }
1181 
1182  void define_extern(const std::string &function_name,
1183  const std::vector<ExternFuncArgument> &params, Type t,
1184  const std::vector<Var> &arguments,
1186  DeviceAPI device_api = DeviceAPI::Host) {
1187  define_extern(function_name, params, std::vector<Type>{t}, arguments,
1188  mangling, device_api);
1189  }
1190 
1191  void define_extern(const std::string &function_name,
1192  const std::vector<ExternFuncArgument> &params,
1193  const std::vector<Type> &types,
1194  const std::vector<Var> &arguments,
1196  DeviceAPI device_api = DeviceAPI::Host);
1197  // @}
1198 
1199  /** Get the type(s) of the outputs of this Func.
1200  *
1201  * It is not legal to call type() unless the Func has non-Tuple elements.
1202  *
1203  * If the Func isn't yet defined, and was not specified with required types,
1204  * a runtime error will occur.
1205  *
1206  * If the Func isn't yet defined, but *was* specified with required types,
1207  * the requirements will be returned. */
1208  // @{
1209  const Type &type() const;
1210  const std::vector<Type> &types() const;
1211  // @}
1212 
1213  /** Get the number of outputs of this Func. Corresponds to the
1214  * size of the Tuple this Func was defined to return.
1215  * If the Func isn't yet defined, but was specified with required types,
1216  * the number of outputs specified in the requirements will be returned. */
1217  int outputs() const;
1218 
1219  /** Get the name of the extern function called for an extern
1220  * definition. */
1221  const std::string &extern_function_name() const;
1222 
1223  /** The dimensionality (number of arguments) of this function.
1224  * If the Func isn't yet defined, but was specified with required dimensionality,
1225  * the dimensionality specified in the requirements will be returned. */
1226  int dimensions() const;
1227 
1228  /** Construct either the left-hand-side of a definition, or a call
1229  * to a functions that happens to only contain vars as
1230  * arguments. If the function has already been defined, and fewer
1231  * arguments are given than the function has dimensions, then
1232  * enough implicit vars are added to the end of the argument list
1233  * to make up the difference (see \ref Var::implicit) */
1234  // @{
1235  FuncRef operator()(std::vector<Var>) const;
1236 
1237  template<typename... Args>
1239  operator()(Args &&...args) const {
1240  std::vector<Var> collected_args{std::forward<Args>(args)...};
1241  return this->operator()(collected_args);
1242  }
1243  // @}
1244 
1245  /** Either calls to the function, or the left-hand-side of
1246  * an update definition (see \ref RDom). If the function has
1247  * already been defined, and fewer arguments are given than the
1248  * function has dimensions, then enough implicit vars are added to
1249  * the end of the argument list to make up the difference. (see
1250  * \ref Var::implicit)*/
1251  // @{
1252  FuncRef operator()(std::vector<Expr>) const;
1253 
1254  template<typename... Args>
1256  operator()(const Expr &x, Args &&...args) const {
1257  std::vector<Expr> collected_args{x, std::forward<Args>(args)...};
1258  return (*this)(collected_args);
1259  }
1260  // @}
1261 
1262  /** Creates and returns a new identity Func that wraps this Func. During
1263  * compilation, Halide replaces all calls to this Func done by 'f'
1264  * with calls to the wrapper. If this Func is already wrapped for
1265  * use in 'f', will return the existing wrapper.
1266  *
1267  * For example, g.in(f) would rewrite a pipeline like this:
1268  \code
1269  g(x, y) = ...
1270  f(x, y) = ... g(x, y) ...
1271  \endcode
1272  * into a pipeline like this:
1273  \code
1274  g(x, y) = ...
1275  g_wrap(x, y) = g(x, y)
1276  f(x, y) = ... g_wrap(x, y)
1277  \endcode
1278  *
1279  * This has a variety of uses. You can use it to schedule this
1280  * Func differently in the different places it is used:
1281  \code
1282  g(x, y) = ...
1283  f1(x, y) = ... g(x, y) ...
1284  f2(x, y) = ... g(x, y) ...
1285  g.in(f1).compute_at(f1, y).vectorize(x, 8);
1286  g.in(f2).compute_at(f2, x).unroll(x);
1287  \endcode
1288  *
1289  * You can also use it to stage loads from this Func via some
1290  * intermediate buffer (perhaps on the stack as in
1291  * test/performance/block_transpose.cpp, or in shared GPU memory
1292  * as in test/performance/wrap.cpp). In this we compute the
1293  * wrapper at tiles of the consuming Funcs like so:
1294  \code
1295  g.compute_root()...
1296  g.in(f).compute_at(f, tiles)...
1297  \endcode
1298  *
1299  * Func::in() can also be used to compute pieces of a Func into a
1300  * smaller scratch buffer (perhaps on the GPU) and then copy them
1301  * into a larger output buffer one tile at a time. See
1302  * apps/interpolate/interpolate.cpp for an example of this. In
1303  * this case we compute the Func at tiles of its own wrapper:
1304  \code
1305  f.in(g).compute_root().gpu_tile(...)...
1306  f.compute_at(f.in(g), tiles)...
1307  \endcode
1308  *
1309  * A similar use of Func::in() wrapping Funcs with multiple update
1310  * stages in a pure wrapper. The following code:
1311  \code
1312  f(x, y) = x + y;
1313  f(x, y) += 5;
1314  g(x, y) = f(x, y);
1315  f.compute_root();
1316  \endcode
1317  *
1318  * Is equivalent to:
1319  \code
1320  for y:
1321  for x:
1322  f(x, y) = x + y;
1323  for y:
1324  for x:
1325  f(x, y) += 5
1326  for y:
1327  for x:
1328  g(x, y) = f(x, y)
1329  \endcode
1330  * using Func::in(), we can write:
1331  \code
1332  f(x, y) = x + y;
1333  f(x, y) += 5;
1334  g(x, y) = f(x, y);
1335  f.in(g).compute_root();
1336  \endcode
1337  * which instead produces:
1338  \code
1339  for y:
1340  for x:
1341  f(x, y) = x + y;
1342  f(x, y) += 5
1343  f_wrap(x, y) = f(x, y)
1344  for y:
1345  for x:
1346  g(x, y) = f_wrap(x, y)
1347  \endcode
1348  */
1349  Func in(const Func &f);
1350 
1351  /** Create and return an identity wrapper shared by all the Funcs in
1352  * 'fs'. If any of the Funcs in 'fs' already have a custom wrapper,
1353  * this will throw an error. */
1354  Func in(const std::vector<Func> &fs);
1355 
1356  /** Create and return a global identity wrapper, which wraps all calls to
1357  * this Func by any other Func. If a global wrapper already exists,
1358  * returns it. The global identity wrapper is only used by callers for
1359  * which no custom wrapper has been specified.
1360  */
1362 
1363  /** Similar to \ref Func::in; however, instead of replacing the call to
1364  * this Func with an identity Func that refers to it, this replaces the
1365  * call with a clone of this Func.
1366  *
1367  * For example, f.clone_in(g) would rewrite a pipeline like this:
1368  \code
1369  f(x, y) = x + y;
1370  g(x, y) = f(x, y) + 2;
1371  h(x, y) = f(x, y) - 3;
1372  \endcode
1373  * into a pipeline like this:
1374  \code
1375  f(x, y) = x + y;
1376  f_clone(x, y) = x + y;
1377  g(x, y) = f_clone(x, y) + 2;
1378  h(x, y) = f(x, y) - 3;
1379  \endcode
1380  *
1381  */
1382  //@{
1383  Func clone_in(const Func &f);
1384  Func clone_in(const std::vector<Func> &fs);
1385  //@}
1386 
1387  /** Declare that this function should be implemented by a call to
1388  * halide_buffer_copy with the given target device API. Asserts
1389  * that the Func has a pure definition which is a simple call to a
1390  * single input, and no update definitions. The wrapper Funcs
1391  * returned by in() are suitable candidates. Consumes all pure
1392  * variables, and rewrites the Func to have an extern definition
1393  * that calls halide_buffer_copy. */
1395 
1396  /** Declare that this function should be implemented by a call to
1397  * halide_buffer_copy with a NULL target device API. Equivalent to
1398  * copy_to_device(DeviceAPI::Host). Asserts that the Func has a
1399  * pure definition which is a simple call to a single input, and
1400  * no update definitions. The wrapper Funcs returned by in() are
1401  * suitable candidates. Consumes all pure variables, and rewrites
1402  * the Func to have an extern definition that calls
1403  * halide_buffer_copy.
1404  *
1405  * Note that if the source Func is already valid in host memory,
1406  * this compiles to code that does the minimum number of calls to
1407  * memcpy.
1408  */
1410 
1411  /** Split a dimension into inner and outer subdimensions with the
1412  * given names, where the inner dimension iterates from 0 to
1413  * factor-1. The inner and outer subdimensions can then be dealt
1414  * with using the other scheduling calls. It's ok to reuse the old
1415  * variable name as either the inner or outer variable. The final
1416  * argument specifies how the tail should be handled if the split
1417  * factor does not provably divide the extent. */
1418  Func &split(const VarOrRVar &old, const VarOrRVar &outer, const VarOrRVar &inner, const Expr &factor, TailStrategy tail = TailStrategy::Auto);
1419 
1420  /** Join two dimensions into a single fused dimension. The fused dimension
1421  * covers the product of the extents of the inner and outer dimensions
1422  * given. The loop type (e.g. parallel, vectorized) of the resulting fused
1423  * dimension is inherited from the first argument. */
1424  Func &fuse(const VarOrRVar &inner, const VarOrRVar &outer, const VarOrRVar &fused);
1425 
1426  /** Mark a dimension to be traversed serially. This is the default. */
1427  Func &serial(const VarOrRVar &var);
1428 
1429  /** Mark a dimension to be traversed in parallel */
1430  Func &parallel(const VarOrRVar &var);
1431 
1432  /** Split a dimension by the given task_size, and the parallelize the
1433  * outer dimension. This creates parallel tasks that have size
1434  * task_size. After this call, var refers to the outer dimension of
1435  * the split. The inner dimension has a new anonymous name. If you
1436  * wish to mutate it, or schedule with respect to it, do the split
1437  * manually. */
1438  Func &parallel(const VarOrRVar &var, const Expr &task_size, TailStrategy tail = TailStrategy::Auto);
1439 
1440  /** Mark a dimension to be computed all-at-once as a single
1441  * vector. The dimension should have constant extent -
1442  * e.g. because it is the inner dimension following a split by a
1443  * constant factor. For most uses of vectorize you want the two
1444  * argument form. The variable to be vectorized should be the
1445  * innermost one. */
1446  Func &vectorize(const VarOrRVar &var);
1447 
1448  /** Mark a dimension to be completely unrolled. The dimension
1449  * should have constant extent - e.g. because it is the inner
1450  * dimension following a split by a constant factor. For most uses
1451  * of unroll you want the two-argument form. */
1452  Func &unroll(const VarOrRVar &var);
1453 
1454  /** Split a dimension by the given factor, then vectorize the
1455  * inner dimension. This is how you vectorize a loop of unknown
1456  * size. The variable to be vectorized should be the innermost
1457  * one. After this call, var refers to the outer dimension of the
1458  * split. 'factor' must be an integer. */
1459  Func &vectorize(const VarOrRVar &var, const Expr &factor, TailStrategy tail = TailStrategy::Auto);
1460 
1461  /** Split a dimension by the given factor, then unroll the inner
1462  * dimension. This is how you unroll a loop of unknown size by
1463  * some constant factor. After this call, var refers to the outer
1464  * dimension of the split. 'factor' must be an integer. */
1465  Func &unroll(const VarOrRVar &var, const Expr &factor, TailStrategy tail = TailStrategy::Auto);
1466 
1467  /** Set the loop partition policy. Loop partitioning can be useful to
1468  * optimize boundary conditions (such as clamp_edge). Loop partitioning
1469  * splits a for loop into three for loops: a prologue, a steady-state,
1470  * and an epilogue.
1471  * The default policy is Auto. */
1472  Func &partition(const VarOrRVar &var, Partition partition_policy);
1473 
1474  /** Set the loop partition policy to Never for a vector of Vars and
1475  * RVars. */
1476  Func &never_partition(const std::vector<VarOrRVar> &vars);
1477 
1478  /** Set the loop partition policy to Never for some number of Vars and RVars. */
1479  template<typename... Args>
1481  never_partition(const VarOrRVar &x, Args &&...args) {
1482  std::vector<VarOrRVar> collected_args{x, std::forward<Args>(args)...};
1483  return never_partition(collected_args);
1484  }
1485 
1486  /** Set the loop partition policy to Never for all Vars and RVar of the
1487  * initial definition of the Func. It must be called separately on any
1488  * update definitions. */
1490 
1491  /** Set the loop partition policy to Always for a vector of Vars and
1492  * RVars. */
1493  Func &always_partition(const std::vector<VarOrRVar> &vars);
1494 
1495  /** Set the loop partition policy to Always for some number of Vars and RVars. */
1496  template<typename... Args>
1498  always_partition(const VarOrRVar &x, Args &&...args) {
1499  std::vector<VarOrRVar> collected_args{x, std::forward<Args>(args)...};
1500  return always_partition(collected_args);
1501  }
1502 
1503  /** Set the loop partition policy to Always for all Vars and RVar of the
1504  * initial definition of the Func. It must be called separately on any
1505  * update definitions. */
1507 
1508  /** Statically declare that the range over which a function should
1509  * be evaluated is given by the second and third arguments. This
1510  * can let Halide perform some optimizations. E.g. if you know
1511  * there are going to be 4 color channels, you can completely
1512  * vectorize the color channel dimension without the overhead of
1513  * splitting it up. If bounds inference decides that it requires
1514  * more of this function than the bounds you have stated, a
1515  * runtime error will occur when you try to run your pipeline. */
1516  Func &bound(const Var &var, Expr min, Expr extent);
1517 
1518  /** Statically declare the range over which the function will be
1519  * evaluated in the general case. This provides a basis for the auto
1520  * scheduler to make trade-offs and scheduling decisions. The auto
1521  * generated schedules might break when the sizes of the dimensions are
1522  * very different from the estimates specified. These estimates are used
1523  * only by the auto scheduler if the function is a pipeline output. */
1524  Func &set_estimate(const Var &var, const Expr &min, const Expr &extent);
1525 
1526  /** Set (min, extent) estimates for all dimensions in the Func
1527  * at once; this is equivalent to calling `set_estimate(args()[n], min, extent)`
1528  * repeatedly, but slightly terser. The size of the estimates vector
1529  * must match the dimensionality of the Func. */
1530  Func &set_estimates(const Region &estimates);
1531 
1532  /** Expand the region computed so that the min coordinates is
1533  * congruent to 'remainder' modulo 'modulus', and the extent is a
1534  * multiple of 'modulus'. For example, f.align_bounds(x, 2) forces
1535  * the min and extent realized to be even, and calling
1536  * f.align_bounds(x, 2, 1) forces the min to be odd and the extent
1537  * to be even. The region computed always contains the region that
1538  * would have been computed without this directive, so no
1539  * assertions are injected.
1540  */
1541  Func &align_bounds(const Var &var, Expr modulus, Expr remainder = 0);
1542 
1543  /** Expand the region computed so that the extent is a
1544  * multiple of 'modulus'. For example, f.align_extent(x, 2) forces
1545  * the extent realized to be even. The region computed always contains the
1546  * region that would have been computed without this directive, so no
1547  * assertions are injected. (This is essentially equivalent to align_bounds(),
1548  * but always leaving the min untouched.)
1549  */
1550  Func &align_extent(const Var &var, Expr modulus);
1551 
1552  /** Bound the extent of a Func's realization, but not its
1553  * min. This means the dimension can be unrolled or vectorized
1554  * even when its min is not fixed (for example because it is
1555  * compute_at tiles of another Func). This can also be useful for
1556  * forcing a function's allocation to be a fixed size, which often
1557  * means it can go on the stack. */
1558  Func &bound_extent(const Var &var, Expr extent);
1559 
1560  /** Split two dimensions at once by the given factors, and then
1561  * reorder the resulting dimensions to be xi, yi, xo, yo from
1562  * innermost outwards. This gives a tiled traversal. */
1563  Func &tile(const VarOrRVar &x, const VarOrRVar &y,
1564  const VarOrRVar &xo, const VarOrRVar &yo,
1565  const VarOrRVar &xi, const VarOrRVar &yi,
1566  const Expr &xfactor, const Expr &yfactor,
1568 
1569  /** A shorter form of tile, which reuses the old variable names as
1570  * the new outer dimensions */
1571  Func &tile(const VarOrRVar &x, const VarOrRVar &y,
1572  const VarOrRVar &xi, const VarOrRVar &yi,
1573  const Expr &xfactor, const Expr &yfactor,
1575 
1576  /** A more general form of tile, which defines tiles of any dimensionality. */
1577  Func &tile(const std::vector<VarOrRVar> &previous,
1578  const std::vector<VarOrRVar> &outers,
1579  const std::vector<VarOrRVar> &inners,
1580  const std::vector<Expr> &factors,
1581  const std::vector<TailStrategy> &tails);
1582 
1583  /** The generalized tile, with a single tail strategy to apply to all vars. */
1584  Func &tile(const std::vector<VarOrRVar> &previous,
1585  const std::vector<VarOrRVar> &outers,
1586  const std::vector<VarOrRVar> &inners,
1587  const std::vector<Expr> &factors,
1589 
1590  /** Generalized tiling, reusing the previous names as the outer names. */
1591  Func &tile(const std::vector<VarOrRVar> &previous,
1592  const std::vector<VarOrRVar> &inners,
1593  const std::vector<Expr> &factors,
1595 
1596  /** Reorder variables to have the given nesting order, from
1597  * innermost out */
1598  Func &reorder(const std::vector<VarOrRVar> &vars);
1599 
1600  template<typename... Args>
1602  reorder(const VarOrRVar &x, const VarOrRVar &y, Args &&...args) {
1603  std::vector<VarOrRVar> collected_args{x, y, std::forward<Args>(args)...};
1604  return reorder(collected_args);
1605  }
1606 
1607  /** Rename a dimension. Equivalent to split with a inner size of one. */
1608  Func &rename(const VarOrRVar &old_name, const VarOrRVar &new_name);
1609 
1610  /** Specify that race conditions are permitted for this Func,
1611  * which enables parallelizing over RVars even when Halide cannot
1612  * prove that it is safe to do so. Use this with great caution,
1613  * and only if you can prove to yourself that this is safe, as it
1614  * may result in a non-deterministic routine that returns
1615  * different values at different times or on different machines. */
1617 
1618  /** Issue atomic updates for this Func. This allows parallelization
1619  * on associative RVars. The function throws a compile error when
1620  * Halide fails to prove associativity. Use override_associativity_test
1621  * to disable the associativity test if you believe the function is
1622  * associative or the order of reduction variable execution does not
1623  * matter.
1624  * Halide compiles this into hardware atomic operations whenever possible,
1625  * and falls back to a mutex lock per storage element if it is impossible
1626  * to atomically update.
1627  * There are three possible outcomes of the compiled code:
1628  * atomic add, compare-and-swap loop, and mutex lock.
1629  * For example:
1630  *
1631  * hist(x) = 0;
1632  * hist(im(r)) += 1;
1633  * hist.compute_root();
1634  * hist.update().atomic().parallel();
1635  *
1636  * will be compiled to atomic add operations.
1637  *
1638  * hist(x) = 0;
1639  * hist(im(r)) = min(hist(im(r)) + 1, 100);
1640  * hist.compute_root();
1641  * hist.update().atomic().parallel();
1642  *
1643  * will be compiled to compare-and-swap loops.
1644  *
1645  * arg_max() = {0, im(0)};
1646  * Expr old_index = arg_max()[0];
1647  * Expr old_max = arg_max()[1];
1648  * Expr new_index = select(old_max < im(r), r, old_index);
1649  * Expr new_max = max(im(r), old_max);
1650  * arg_max() = {new_index, new_max};
1651  * arg_max.compute_root();
1652  * arg_max.update().atomic().parallel();
1653  *
1654  * will be compiled to updates guarded by a mutex lock,
1655  * since it is impossible to atomically update two different locations.
1656  *
1657  * Currently the atomic operation is supported by x86, CUDA, and OpenCL backends.
1658  * Compiling to other backends results in a compile error.
1659  * If an operation is compiled into a mutex lock, and is vectorized or is
1660  * compiled to CUDA or OpenCL, it also results in a compile error,
1661  * since per-element mutex lock on vectorized operation leads to a
1662  * deadlock.
1663  * Vectorization of predicated RVars (through rdom.where()) on CPU
1664  * is also unsupported yet (see https://github.com/halide/Halide/issues/4298).
1665  * 8-bit and 16-bit atomics on GPU are also not supported. */
1666  Func &atomic(bool override_associativity_test = false);
1667 
1668  /** Specialize a Func. This creates a special-case version of the
1669  * Func where the given condition is true. The most effective
1670  * conditions are those of the form param == value, and boolean
1671  * Params. Consider a simple example:
1672  \code
1673  f(x) = x + select(cond, 0, 1);
1674  f.compute_root();
1675  \endcode
1676  * This is equivalent to:
1677  \code
1678  for (int x = 0; x < width; x++) {
1679  f[x] = x + (cond ? 0 : 1);
1680  }
1681  \endcode
1682  * Adding the scheduling directive:
1683  \code
1684  f.specialize(cond)
1685  \endcode
1686  * makes it equivalent to:
1687  \code
1688  if (cond) {
1689  for (int x = 0; x < width; x++) {
1690  f[x] = x;
1691  }
1692  } else {
1693  for (int x = 0; x < width; x++) {
1694  f[x] = x + 1;
1695  }
1696  }
1697  \endcode
1698  * Note that the inner loops have been simplified. In the first
1699  * path Halide knows that cond is true, and in the second path
1700  * Halide knows that it is false.
1701  *
1702  * The specialized version gets its own schedule, which inherits
1703  * every directive made about the parent Func's schedule so far
1704  * except for its specializations. This method returns a handle to
1705  * the new schedule. If you wish to retrieve the specialized
1706  * sub-schedule again later, you can call this method with the
1707  * same condition. Consider the following example of scheduling
1708  * the specialized version:
1709  *
1710  \code
1711  f(x) = x;
1712  f.compute_root();
1713  f.specialize(width > 1).unroll(x, 2);
1714  \endcode
1715  * Assuming for simplicity that width is even, this is equivalent to:
1716  \code
1717  if (width > 1) {
1718  for (int x = 0; x < width/2; x++) {
1719  f[2*x] = 2*x;
1720  f[2*x + 1] = 2*x + 1;
1721  }
1722  } else {
1723  for (int x = 0; x < width/2; x++) {
1724  f[x] = x;
1725  }
1726  }
1727  \endcode
1728  * For this case, it may be better to schedule the un-specialized
1729  * case instead:
1730  \code
1731  f(x) = x;
1732  f.compute_root();
1733  f.specialize(width == 1); // Creates a copy of the schedule so far.
1734  f.unroll(x, 2); // Only applies to the unspecialized case.
1735  \endcode
1736  * This is equivalent to:
1737  \code
1738  if (width == 1) {
1739  f[0] = 0;
1740  } else {
1741  for (int x = 0; x < width/2; x++) {
1742  f[2*x] = 2*x;
1743  f[2*x + 1] = 2*x + 1;
1744  }
1745  }
1746  \endcode
1747  * This can be a good way to write a pipeline that splits,
1748  * vectorizes, or tiles, but can still handle small inputs.
1749  *
1750  * If a Func has several specializations, the first matching one
1751  * will be used, so the order in which you define specializations
1752  * is significant. For example:
1753  *
1754  \code
1755  f(x) = x + select(cond1, a, b) - select(cond2, c, d);
1756  f.specialize(cond1);
1757  f.specialize(cond2);
1758  \endcode
1759  * is equivalent to:
1760  \code
1761  if (cond1) {
1762  for (int x = 0; x < width; x++) {
1763  f[x] = x + a - (cond2 ? c : d);
1764  }
1765  } else if (cond2) {
1766  for (int x = 0; x < width; x++) {
1767  f[x] = x + b - c;
1768  }
1769  } else {
1770  for (int x = 0; x < width; x++) {
1771  f[x] = x + b - d;
1772  }
1773  }
1774  \endcode
1775  *
1776  * Specializations may in turn be specialized, which creates a
1777  * nested if statement in the generated code.
1778  *
1779  \code
1780  f(x) = x + select(cond1, a, b) - select(cond2, c, d);
1781  f.specialize(cond1).specialize(cond2);
1782  \endcode
1783  * This is equivalent to:
1784  \code
1785  if (cond1) {
1786  if (cond2) {
1787  for (int x = 0; x < width; x++) {
1788  f[x] = x + a - c;
1789  }
1790  } else {
1791  for (int x = 0; x < width; x++) {
1792  f[x] = x + a - d;
1793  }
1794  }
1795  } else {
1796  for (int x = 0; x < width; x++) {
1797  f[x] = x + b - (cond2 ? c : d);
1798  }
1799  }
1800  \endcode
1801  * To create a 4-way if statement that simplifies away all of the
1802  * ternary operators above, you could say:
1803  \code
1804  f.specialize(cond1).specialize(cond2);
1805  f.specialize(cond2);
1806  \endcode
1807  * or
1808  \code
1809  f.specialize(cond1 && cond2);
1810  f.specialize(cond1);
1811  f.specialize(cond2);
1812  \endcode
1813  *
1814  * Any prior Func which is compute_at some variable of this Func
1815  * gets separately included in all paths of the generated if
1816  * statement. The Var in the compute_at call to must exist in all
1817  * paths, but it may have been generated via a different path of
1818  * splits, fuses, and renames. This can be used somewhat
1819  * creatively. Consider the following code:
1820  \code
1821  g(x, y) = 8*x;
1822  f(x, y) = g(x, y) + 1;
1823  f.compute_root().specialize(cond);
1824  Var g_loop;
1825  f.specialize(cond).rename(y, g_loop);
1826  f.rename(x, g_loop);
1827  g.compute_at(f, g_loop);
1828  \endcode
1829  * When cond is true, this is equivalent to g.compute_at(f,y).
1830  * When it is false, this is equivalent to g.compute_at(f,x).
1831  */
1832  Stage specialize(const Expr &condition);
1833 
1834  /** Add a specialization to a Func that always terminates execution
1835  * with a call to halide_error(). By itself, this is of limited use,
1836  * but can be useful to terminate chains of specialize() calls where
1837  * no "default" case is expected (thus avoiding unnecessary code generation).
1838  *
1839  * For instance, say we want to optimize a pipeline to process images
1840  * in planar and interleaved format; we might typically do something like:
1841  \code
1842  ImageParam im(UInt(8), 3);
1843  Func f = do_something_with(im);
1844  f.specialize(im.dim(0).stride() == 1).vectorize(x, 8); // planar
1845  f.specialize(im.dim(2).stride() == 1).reorder(c, x, y).vectorize(c); // interleaved
1846  \endcode
1847  * This code will vectorize along rows for the planar case, and across pixel
1848  * components for the interleaved case... but there is an implicit "else"
1849  * for the unhandled cases, which generates unoptimized code. If we never
1850  * anticipate passing any other sort of images to this, we code streamline
1851  * our code by adding specialize_fail():
1852  \code
1853  ImageParam im(UInt(8), 3);
1854  Func f = do_something(im);
1855  f.specialize(im.dim(0).stride() == 1).vectorize(x, 8); // planar
1856  f.specialize(im.dim(2).stride() == 1).reorder(c, x, y).vectorize(c); // interleaved
1857  f.specialize_fail("Unhandled image format");
1858  \endcode
1859  * Conceptually, this produces codes like:
1860  \code
1861  if (im.dim(0).stride() == 1) {
1862  do_something_planar();
1863  } else if (im.dim(2).stride() == 1) {
1864  do_something_interleaved();
1865  } else {
1866  halide_error("Unhandled image format");
1867  }
1868  \endcode
1869  *
1870  * Note that calling specialize_fail() terminates the specialization chain
1871  * for a given Func; you cannot create new specializations for the Func
1872  * afterwards (though you can retrieve handles to previous specializations).
1873  */
1874  void specialize_fail(const std::string &message);
1875 
1876  /** Tell Halide that the following dimensions correspond to GPU
1877  * thread indices. This is useful if you compute a producer
1878  * function within the block indices of a consumer function, and
1879  * want to control how that function's dimensions map to GPU
1880  * threads. If the selected target is not an appropriate GPU, this
1881  * just marks those dimensions as parallel. */
1882  // @{
1883  Func &gpu_threads(const VarOrRVar &thread_x, DeviceAPI device_api = DeviceAPI::Default_GPU);
1884  Func &gpu_threads(const VarOrRVar &thread_x, const VarOrRVar &thread_y, DeviceAPI device_api = DeviceAPI::Default_GPU);
1885  Func &gpu_threads(const VarOrRVar &thread_x, const VarOrRVar &thread_y, const VarOrRVar &thread_z, DeviceAPI device_api = DeviceAPI::Default_GPU);
1886  // @}
1887 
1888  /** The given dimension corresponds to the lanes in a GPU
1889  * warp. GPU warp lanes are distinguished from GPU threads by the
1890  * fact that all warp lanes run together in lockstep, which
1891  * permits lightweight communication of data from one lane to
1892  * another. */
1893  Func &gpu_lanes(const VarOrRVar &thread_x, DeviceAPI device_api = DeviceAPI::Default_GPU);
1894 
1895  /** Tell Halide to run this stage using a single gpu thread and
1896  * block. This is not an efficient use of your GPU, but it can be
1897  * useful to avoid copy-back for intermediate update stages that
1898  * touch a very small part of your Func. */
1900 
1901  /** Tell Halide that the following dimensions correspond to GPU
1902  * block indices. This is useful for scheduling stages that will
1903  * run serially within each GPU block. If the selected target is
1904  * not ptx, this just marks those dimensions as parallel. */
1905  // @{
1907  Func &gpu_blocks(const VarOrRVar &block_x, const VarOrRVar &block_y, DeviceAPI device_api = DeviceAPI::Default_GPU);
1908  Func &gpu_blocks(const VarOrRVar &block_x, const VarOrRVar &block_y, const VarOrRVar &block_z, DeviceAPI device_api = DeviceAPI::Default_GPU);
1909  // @}
1910 
1911  /** Tell Halide that the following dimensions correspond to GPU
1912  * block indices and thread indices. If the selected target is not
1913  * ptx, these just mark the given dimensions as parallel. The
1914  * dimensions are consumed by this call, so do all other
1915  * unrolling, reordering, etc first. */
1916  // @{
1917  Func &gpu(const VarOrRVar &block_x, const VarOrRVar &thread_x, DeviceAPI device_api = DeviceAPI::Default_GPU);
1918  Func &gpu(const VarOrRVar &block_x, const VarOrRVar &block_y,
1919  const VarOrRVar &thread_x, const VarOrRVar &thread_y, DeviceAPI device_api = DeviceAPI::Default_GPU);
1920  Func &gpu(const VarOrRVar &block_x, const VarOrRVar &block_y, const VarOrRVar &block_z,
1921  const VarOrRVar &thread_x, const VarOrRVar &thread_y, const VarOrRVar &thread_z, DeviceAPI device_api = DeviceAPI::Default_GPU);
1922  // @}
1923 
1924  /** Short-hand for tiling a domain and mapping the tile indices
1925  * to GPU block indices and the coordinates within each tile to
1926  * GPU thread indices. Consumes the variables given, so do all
1927  * other scheduling first. */
1928  // @{
1929  Func &gpu_tile(const VarOrRVar &x, const VarOrRVar &bx, const VarOrRVar &tx, const Expr &x_size,
1931  DeviceAPI device_api = DeviceAPI::Default_GPU);
1932 
1933  Func &gpu_tile(const VarOrRVar &x, const VarOrRVar &tx, const Expr &x_size,
1935  DeviceAPI device_api = DeviceAPI::Default_GPU);
1936  Func &gpu_tile(const VarOrRVar &x, const VarOrRVar &y,
1937  const VarOrRVar &bx, const VarOrRVar &by,
1938  const VarOrRVar &tx, const VarOrRVar &ty,
1939  const Expr &x_size, const Expr &y_size,
1941  DeviceAPI device_api = DeviceAPI::Default_GPU);
1942 
1943  Func &gpu_tile(const VarOrRVar &x, const VarOrRVar &y,
1944  const VarOrRVar &tx, const VarOrRVar &ty,
1945  const Expr &x_size, const Expr &y_size,
1947  DeviceAPI device_api = DeviceAPI::Default_GPU);
1948 
1949  Func &gpu_tile(const VarOrRVar &x, const VarOrRVar &y, const VarOrRVar &z,
1950  const VarOrRVar &bx, const VarOrRVar &by, const VarOrRVar &bz,
1951  const VarOrRVar &tx, const VarOrRVar &ty, const VarOrRVar &tz,
1952  const Expr &x_size, const Expr &y_size, const Expr &z_size,
1954  DeviceAPI device_api = DeviceAPI::Default_GPU);
1955  Func &gpu_tile(const VarOrRVar &x, const VarOrRVar &y, const VarOrRVar &z,
1956  const VarOrRVar &tx, const VarOrRVar &ty, const VarOrRVar &tz,
1957  const Expr &x_size, const Expr &y_size, const Expr &z_size,
1959  DeviceAPI device_api = DeviceAPI::Default_GPU);
1960  // @}
1961 
1962  /** Schedule for execution on Hexagon. When a loop is marked with
1963  * Hexagon, that loop is executed on a Hexagon DSP. */
1965 
1966  /** Prefetch data written to or read from a Func or an ImageParam by a
1967  * subsequent loop iteration, at an optionally specified iteration offset. You may specify
1968  * specification of different vars for the location of the prefetch() instruction
1969  * vs. the location that is being prefetched:
1970  *
1971  * - the first var specified, 'at', indicates the loop in which the prefetch will be placed
1972  * - the second var specified, 'from', determines the var used to find the bounds to prefetch
1973  * (in conjunction with 'offset')
1974  *
1975  * If 'at' and 'from' are distinct vars, then 'from' must be at a nesting level outside 'at.'
1976  * Note that the value for 'offset' applies only to 'from', not 'at'.
1977  *
1978  * The final argument specifies how prefetch of region outside bounds
1979  * should be handled.
1980  *
1981  * For example, consider this pipeline:
1982  \code
1983  Func f, g;
1984  Var x, y, z;
1985  f(x, y) = x + y;
1986  g(x, y) = 2 * f(x, y);
1987  h(x, y) = 3 * f(x, y);
1988  \endcode
1989  *
1990  * The following schedule:
1991  \code
1992  f.compute_root();
1993  g.prefetch(f, x, x, 2, PrefetchBoundStrategy::NonFaulting);
1994  h.prefetch(f, x, y, 2, PrefetchBoundStrategy::NonFaulting);
1995  \endcode
1996  *
1997  * will inject prefetch call at the innermost loop of 'g' and 'h' and generate
1998  * the following loop nest:
1999  \code
2000  for y = ...
2001  for x = ...
2002  f(x, y) = x + y
2003  for y = ..
2004  for x = ...
2005  prefetch(&f[x + 2, y], 1, 16);
2006  g(x, y) = 2 * f(x, y)
2007  for y = ..
2008  for x = ...
2009  prefetch(&f[x, y + 2], 1, 16);
2010  h(x, y) = 3 * f(x, y)
2011  \endcode
2012  *
2013  * Note that the 'from' nesting level need not be adjacent to 'at':
2014  \code
2015  Func f, g;
2016  Var x, y, z, w;
2017  f(x, y, z, w) = x + y + z + w;
2018  g(x, y, z, w) = 2 * f(x, y, z, w);
2019  \endcode
2020  *
2021  * The following schedule:
2022  \code
2023  f.compute_root();
2024  g.prefetch(f, y, w, 2, PrefetchBoundStrategy::NonFaulting);
2025  \endcode
2026  *
2027  * will produce code that prefetches a tile of data:
2028  \code
2029  for w = ...
2030  for z = ...
2031  for y = ...
2032  for x = ...
2033  f(x, y, z, w) = x + y + z + w
2034  for w = ...
2035  for z = ...
2036  for y = ...
2037  for x0 = ...
2038  prefetch(&f[x0, y, z, w + 2], 1, 16);
2039  for x = ...
2040  g(x, y, z, w) = 2 * f(x, y, z, w)
2041  \endcode
2042  *
2043  * Note that calling prefetch() with the same var for both 'at' and 'from'
2044  * is equivalent to calling prefetch() with that var.
2045  */
2046  // @{
2047  Func &prefetch(const Func &f, const VarOrRVar &at, const VarOrRVar &from, Expr offset = 1,
2049  Func &prefetch(const Parameter &param, const VarOrRVar &at, const VarOrRVar &from, Expr offset = 1,
2051  template<typename T>
2052  Func &prefetch(const T &image, const VarOrRVar &at, const VarOrRVar &from, Expr offset = 1,
2054  return prefetch(image.parameter(), at, from, std::move(offset), strategy);
2055  }
2056  // @}
2057 
2058  /** Specify how the storage for the function is laid out. These
2059  * calls let you specify the nesting order of the dimensions. For
2060  * example, foo.reorder_storage(y, x) tells Halide to use
2061  * column-major storage for any realizations of foo, without
2062  * changing how you refer to foo in the code. You may want to do
2063  * this if you intend to vectorize across y. When representing
2064  * color images, foo.reorder_storage(c, x, y) specifies packed
2065  * storage (red, green, and blue values adjacent in memory), and
2066  * foo.reorder_storage(x, y, c) specifies planar storage (entire
2067  * red, green, and blue images one after the other in memory).
2068  *
2069  * If you leave out some dimensions, those remain in the same
2070  * positions in the nesting order while the specified variables
2071  * are reordered around them. */
2072  // @{
2073  Func &reorder_storage(const std::vector<Var> &dims);
2074 
2075  Func &reorder_storage(const Var &x, const Var &y);
2076  template<typename... Args>
2078  reorder_storage(const Var &x, const Var &y, Args &&...args) {
2079  std::vector<Var> collected_args{x, y, std::forward<Args>(args)...};
2080  return reorder_storage(collected_args);
2081  }
2082  // @}
2083 
2084  /** Pad the storage extent of a particular dimension of
2085  * realizations of this function up to be a multiple of the
2086  * specified alignment. This guarantees that the strides for the
2087  * dimensions stored outside of dim will be multiples of the
2088  * specified alignment, where the strides and alignment are
2089  * measured in numbers of elements.
2090  *
2091  * For example, to guarantee that a function foo(x, y, c)
2092  * representing an image has scanlines starting on offsets
2093  * aligned to multiples of 16, use foo.align_storage(x, 16). */
2094  Func &align_storage(const Var &dim, const Expr &alignment);
2095 
2096  /** Store realizations of this function in a circular buffer of a
2097  * given extent. This is more efficient when the extent of the
2098  * circular buffer is a power of 2. If the fold factor is too
2099  * small, or the dimension is not accessed monotonically, the
2100  * pipeline will generate an error at runtime.
2101  *
2102  * The fold_forward option indicates that the new values of the
2103  * producer are accessed by the consumer in a monotonically
2104  * increasing order. Folding storage of producers is also
2105  * supported if the new values are accessed in a monotonically
2106  * decreasing order by setting fold_forward to false.
2107  *
2108  * For example, consider the pipeline:
2109  \code
2110  Func f, g;
2111  Var x, y;
2112  g(x, y) = x*y;
2113  f(x, y) = g(x, y) + g(x, y+1);
2114  \endcode
2115  *
2116  * If we schedule f like so:
2117  *
2118  \code
2119  g.compute_at(f, y).store_root().fold_storage(y, 2);
2120  \endcode
2121  *
2122  * Then g will be computed at each row of f and stored in a buffer
2123  * with an extent in y of 2, alternately storing each computed row
2124  * of g in row y=0 or y=1.
2125  */
2126  Func &fold_storage(const Var &dim, const Expr &extent, bool fold_forward = true);
2127 
2128  /** Compute this function as needed for each unique value of the
2129  * given var for the given calling function f.
2130  *
2131  * For example, consider the simple pipeline:
2132  \code
2133  Func f, g;
2134  Var x, y;
2135  g(x, y) = x*y;
2136  f(x, y) = g(x, y) + g(x, y+1) + g(x+1, y) + g(x+1, y+1);
2137  \endcode
2138  *
2139  * If we schedule f like so:
2140  *
2141  \code
2142  g.compute_at(f, x);
2143  \endcode
2144  *
2145  * Then the C code equivalent to this pipeline will look like this
2146  *
2147  \code
2148 
2149  int f[height][width];
2150  for (int y = 0; y < height; y++) {
2151  for (int x = 0; x < width; x++) {
2152  int g[2][2];
2153  g[0][0] = x*y;
2154  g[0][1] = (x+1)*y;
2155  g[1][0] = x*(y+1);
2156  g[1][1] = (x+1)*(y+1);
2157  f[y][x] = g[0][0] + g[1][0] + g[0][1] + g[1][1];
2158  }
2159  }
2160 
2161  \endcode
2162  *
2163  * The allocation and computation of g is within f's loop over x,
2164  * and enough of g is computed to satisfy all that f will need for
2165  * that iteration. This has excellent locality - values of g are
2166  * used as soon as they are computed, but it does redundant
2167  * work. Each value of g ends up getting computed four times. If
2168  * we instead schedule f like so:
2169  *
2170  \code
2171  g.compute_at(f, y);
2172  \endcode
2173  *
2174  * The equivalent C code is:
2175  *
2176  \code
2177  int f[height][width];
2178  for (int y = 0; y < height; y++) {
2179  int g[2][width+1];
2180  for (int x = 0; x < width; x++) {
2181  g[0][x] = x*y;
2182  g[1][x] = x*(y+1);
2183  }
2184  for (int x = 0; x < width; x++) {
2185  f[y][x] = g[0][x] + g[1][x] + g[0][x+1] + g[1][x+1];
2186  }
2187  }
2188  \endcode
2189  *
2190  * The allocation and computation of g is within f's loop over y,
2191  * and enough of g is computed to satisfy all that f will need for
2192  * that iteration. This does less redundant work (each point in g
2193  * ends up being evaluated twice), but the locality is not quite
2194  * as good, and we have to allocate more temporary memory to store
2195  * g.
2196  */
2197  Func &compute_at(const Func &f, const Var &var);
2198 
2199  /** Schedule a function to be computed within the iteration over
2200  * some dimension of an update domain. Produces equivalent code
2201  * to the version of compute_at that takes a Var. */
2202  Func &compute_at(const Func &f, const RVar &var);
2203 
2204  /** Schedule a function to be computed within the iteration over
2205  * a given LoopLevel. */
2206  Func &compute_at(LoopLevel loop_level);
2207 
2208  /** Schedule the iteration over the initial definition of this function
2209  * to be fused with another stage 's' from outermost loop to a
2210  * given LoopLevel. */
2211  // @{
2212  Func &compute_with(const Stage &s, const VarOrRVar &var, const std::vector<std::pair<VarOrRVar, LoopAlignStrategy>> &align);
2214  Func &compute_with(LoopLevel loop_level, const std::vector<std::pair<VarOrRVar, LoopAlignStrategy>> &align);
2216 
2217  /** Compute all of this function once ahead of time. Reusing
2218  * the example in \ref Func::compute_at :
2219  *
2220  \code
2221  Func f, g;
2222  Var x, y;
2223  g(x, y) = x*y;
2224  f(x, y) = g(x, y) + g(x, y+1) + g(x+1, y) + g(x+1, y+1);
2225 
2226  g.compute_root();
2227  \endcode
2228  *
2229  * is equivalent to
2230  *
2231  \code
2232  int f[height][width];
2233  int g[height+1][width+1];
2234  for (int y = 0; y < height+1; y++) {
2235  for (int x = 0; x < width+1; x++) {
2236  g[y][x] = x*y;
2237  }
2238  }
2239  for (int y = 0; y < height; y++) {
2240  for (int x = 0; x < width; x++) {
2241  f[y][x] = g[y][x] + g[y+1][x] + g[y][x+1] + g[y+1][x+1];
2242  }
2243  }
2244  \endcode
2245  *
2246  * g is computed once ahead of time, and enough is computed to
2247  * satisfy all uses of it. This does no redundant work (each point
2248  * in g is evaluated once), but has poor locality (values of g are
2249  * probably not still in cache when they are used by f), and
2250  * allocates lots of temporary memory to store g.
2251  */
2253 
2254  /** Use the halide_memoization_cache_... interface to store a
2255  * computed version of this function across invocations of the
2256  * Func.
2257  *
2258  * If an eviction_key is provided, it must be constructed with
2259  * Expr of integer or handle type. The key Expr will be promoted
2260  * to a uint64_t and can be used with halide_memoization_cache_evict
2261  * to remove memoized entries using this eviction key from the
2262  * cache. Memoized computations that do not provide an eviction
2263  * key will never be evicted by this mechanism.
2264  */
2265  Func &memoize(const EvictionKey &eviction_key = EvictionKey());
2266 
2267  /** Produce this Func asynchronously in a separate
2268  * thread. Consumers will be run by the task system when the
2269  * production is complete. If this Func's store level is different
2270  * to its compute level, consumers will be run concurrently,
2271  * blocking as necessary to prevent reading ahead of what the
2272  * producer has computed. If storage is folded, then the producer
2273  * will additionally not be permitted to run too far ahead of the
2274  * consumer, to avoid clobbering data that has not yet been
2275  * used.
2276  *
2277  * Take special care when combining this with custom thread pool
2278  * implementations, as avoiding deadlock with producer-consumer
2279  * parallelism requires a much more sophisticated parallel runtime
2280  * than with data parallelism alone. It is strongly recommended
2281  * you just use Halide's default thread pool, which guarantees no
2282  * deadlock and a bound on the number of threads launched.
2283  */
2285 
2286  /** Expands the storage of the function by an extra dimension
2287  * to enable ring buffering. For this to be useful the storage
2288  * of the function has to be hoisted to an upper loop level using
2289  * \ref Func::hoist_storage. The index for the new ring buffer dimension
2290  * is calculated implicitly based on a linear combination of the all of
2291  * the loop variables between hoist_storage and compute_at/store_at
2292  * loop levels. Scheduling a function with ring_buffer increases the
2293  * amount of memory required for this function by an *extent* times.
2294  * ring_buffer is especially useful in combination with \ref Func::async,
2295  * but can be used without it.
2296  *
2297  * The extent is expected to be a positive integer.
2298  */
2300 
2301  /** Bound the extent of a Func's storage, but not extent of its
2302  * compute. This can be useful for forcing a function's allocation
2303  * to be a fixed size, which often means it can go on the stack.
2304  * If bounds inference decides that it requires more storage for
2305  * this function than the allocation size you have stated, a runtime
2306  * error will occur when you try to run the pipeline. */
2307  Func &bound_storage(const Var &dim, const Expr &bound);
2308 
2309  /** Allocate storage for this function within f's loop over
2310  * var. Scheduling storage is optional, and can be used to
2311  * separate the loop level at which storage occurs from the loop
2312  * level at which computation occurs to trade off between locality
2313  * and redundant work. This can open the door for two types of
2314  * optimization.
2315  *
2316  * Consider again the pipeline from \ref Func::compute_at :
2317  \code
2318  Func f, g;
2319  Var x, y;
2320  g(x, y) = x*y;
2321  f(x, y) = g(x, y) + g(x+1, y) + g(x, y+1) + g(x+1, y+1);
2322  \endcode
2323  *
2324  * If we schedule it like so:
2325  *
2326  \code
2327  g.compute_at(f, x).store_at(f, y);
2328  \endcode
2329  *
2330  * Then the computation of g takes place within the loop over x,
2331  * but the storage takes place within the loop over y:
2332  *
2333  \code
2334  int f[height][width];
2335  for (int y = 0; y < height; y++) {
2336  int g[2][width+1];
2337  for (int x = 0; x < width; x++) {
2338  g[0][x] = x*y;
2339  g[0][x+1] = (x+1)*y;
2340  g[1][x] = x*(y+1);
2341  g[1][x+1] = (x+1)*(y+1);
2342  f[y][x] = g[0][x] + g[1][x] + g[0][x+1] + g[1][x+1];
2343  }
2344  }
2345  \endcode
2346  *
2347  * Provided the for loop over x is serial, halide then
2348  * automatically performs the following sliding window
2349  * optimization:
2350  *
2351  \code
2352  int f[height][width];
2353  for (int y = 0; y < height; y++) {
2354  int g[2][width+1];
2355  for (int x = 0; x < width; x++) {
2356  if (x == 0) {
2357  g[0][x] = x*y;
2358  g[1][x] = x*(y+1);
2359  }
2360  g[0][x+1] = (x+1)*y;
2361  g[1][x+1] = (x+1)*(y+1);
2362  f[y][x] = g[0][x] + g[1][x] + g[0][x+1] + g[1][x+1];
2363  }
2364  }
2365  \endcode
2366  *
2367  * Two of the assignments to g only need to be done when x is
2368  * zero. The rest of the time, those sites have already been
2369  * filled in by a previous iteration. This version has the
2370  * locality of compute_at(f, x), but allocates more memory and
2371  * does much less redundant work.
2372  *
2373  * Halide then further optimizes this pipeline like so:
2374  *
2375  \code
2376  int f[height][width];
2377  for (int y = 0; y < height; y++) {
2378  int g[2][2];
2379  for (int x = 0; x < width; x++) {
2380  if (x == 0) {
2381  g[0][0] = x*y;
2382  g[1][0] = x*(y+1);
2383  }
2384  g[0][(x+1)%2] = (x+1)*y;
2385  g[1][(x+1)%2] = (x+1)*(y+1);
2386  f[y][x] = g[0][x%2] + g[1][x%2] + g[0][(x+1)%2] + g[1][(x+1)%2];
2387  }
2388  }
2389  \endcode
2390  *
2391  * Halide has detected that it's possible to use a circular buffer
2392  * to represent g, and has reduced all accesses to g modulo 2 in
2393  * the x dimension. This optimization only triggers if the for
2394  * loop over x is serial, and if halide can statically determine
2395  * some power of two large enough to cover the range needed. For
2396  * powers of two, the modulo operator compiles to more efficient
2397  * bit-masking. This optimization reduces memory usage, and also
2398  * improves locality by reusing recently-accessed memory instead
2399  * of pulling new memory into cache.
2400  *
2401  */
2402  Func &store_at(const Func &f, const Var &var);
2403 
2404  /** Equivalent to the version of store_at that takes a Var, but
2405  * schedules storage within the loop over a dimension of a
2406  * reduction domain */
2407  Func &store_at(const Func &f, const RVar &var);
2408 
2409  /** Equivalent to the version of store_at that takes a Var, but
2410  * schedules storage at a given LoopLevel. */
2411  Func &store_at(LoopLevel loop_level);
2412 
2413  /** Equivalent to \ref Func::store_at, but schedules storage
2414  * outside the outermost loop. */
2416 
2417  /** Hoist storage for this function within f's loop over
2418  * var. This is different from \ref Func::store_at, because hoist_storage
2419  * simply moves an actual allocation to a given loop level and
2420  * doesn't trigger any of the optimizations such as sliding window.
2421  * Hoisting storage is optional and can be used as an optimization
2422  * to avoid unnecessary allocations by moving it out from an inner
2423  * loop.
2424  *
2425  * Consider again the pipeline from \ref Func::compute_at :
2426  \code
2427  Func f, g;
2428  Var x, y;
2429  g(x, y) = x*y;
2430  f(x, y) = g(x, y) + g(x, y+1) + g(x+1, y) + g(x+1, y+1);
2431  \endcode
2432  *
2433  * If we schedule f like so:
2434  *
2435  \code
2436  g.compute_at(f, x);
2437  \endcode
2438  *
2439  * Then the C code equivalent to this pipeline will look like this
2440  *
2441  \code
2442 
2443  int f[height][width];
2444  for (int y = 0; y < height; y++) {
2445  for (int x = 0; x < width; x++) {
2446  int g[2][2];
2447  g[0][0] = x*y;
2448  g[0][1] = (x+1)*y;
2449  g[1][0] = x*(y+1);
2450  g[1][1] = (x+1)*(y+1);
2451  f[y][x] = g[0][0] + g[1][0] + g[0][1] + g[1][1];
2452  }
2453  }
2454 
2455  \endcode
2456  *
2457  * Note the allocation for g inside of the loop over variable x which
2458  * can happen for each iteration of the inner loop (in total height * width times).
2459  * In some cases allocation can be expensive, so it might be better to do it once
2460  * and reuse allocated memory across all iterations of the loop.
2461  *
2462  * This can be done by scheduling g like so:
2463  *
2464  \code
2465  g.compute_at(f, x).hoist_storage(f, Var::outermost());
2466  \endcode
2467  *
2468  * Then the C code equivalent to this pipeline will look like this
2469  *
2470  \code
2471 
2472  int f[height][width];
2473  int g[2][2];
2474  for (int y = 0; y < height; y++) {
2475  for (int x = 0; x < width; x++) {
2476  g[0][0] = x*y;
2477  g[0][1] = (x+1)*y;
2478  g[1][0] = x*(y+1);
2479  g[1][1] = (x+1)*(y+1);
2480  f[y][x] = g[0][0] + g[1][0] + g[0][1] + g[1][1];
2481  }
2482  }
2483 
2484  \endcode
2485  *
2486  * hoist_storage can be used together with \ref Func::store_at and
2487  * \ref Func::fold_storage (for example, to hoist the storage allocated
2488  * after sliding window optimization).
2489  *
2490  */
2491  Func &hoist_storage(const Func &f, const Var &var);
2492 
2493  /** Equivalent to the version of hoist_storage that takes a Var, but
2494  * schedules storage within the loop over a dimension of a
2495  * reduction domain */
2496  Func &hoist_storage(const Func &f, const RVar &var);
2497 
2498  /** Equivalent to the version of hoist_storage that takes a Var, but
2499  * schedules storage at a given LoopLevel. */
2501 
2502  /** Equivalent to \ref Func::hoist_storage_root, but schedules storage
2503  * outside the outermost loop. */
2505 
2506  /** Aggressively inline all uses of this function. This is the
2507  * default schedule, so you're unlikely to need to call this. For
2508  * a Func with an update definition, that means it gets computed
2509  * as close to the innermost loop as possible.
2510  *
2511  * Consider once more the pipeline from \ref Func::compute_at :
2512  *
2513  \code
2514  Func f, g;
2515  Var x, y;
2516  g(x, y) = x*y;
2517  f(x, y) = g(x, y) + g(x+1, y) + g(x, y+1) + g(x+1, y+1);
2518  \endcode
2519  *
2520  * Leaving g as inline, this compiles to code equivalent to the following C:
2521  *
2522  \code
2523  int f[height][width];
2524  for (int y = 0; y < height; y++) {
2525  for (int x = 0; x < width; x++) {
2526  f[y][x] = x*y + x*(y+1) + (x+1)*y + (x+1)*(y+1);
2527  }
2528  }
2529  \endcode
2530  */
2532 
2533  /** Get a handle on an update step for the purposes of scheduling
2534  * it. */
2535  Stage update(int idx = 0);
2536 
2537  /** Set the type of memory this Func should be stored in. Controls
2538  * whether allocations go on the stack or the heap on the CPU, and
2539  * in global vs shared vs local on the GPU. See the documentation
2540  * on MemoryType for more detail. */
2541  Func &store_in(MemoryType memory_type);
2542 
2543  /** Trace all loads from this Func by emitting calls to
2544  * halide_trace. If the Func is inlined, this has no
2545  * effect. */
2547 
2548  /** Trace all stores to the buffer backing this Func by emitting
2549  * calls to halide_trace. If the Func is inlined, this call
2550  * has no effect. */
2552 
2553  /** Trace all realizations of this Func by emitting calls to
2554  * halide_trace. */
2556 
2557  /** Add a string of arbitrary text that will be passed thru to trace
2558  * inspection code if the Func is realized in trace mode. (Funcs that are
2559  * inlined won't have their tags emitted.) Ignored entirely if
2560  * tracing is not enabled for the Func (or globally).
2561  */
2562  Func &add_trace_tag(const std::string &trace_tag);
2563 
2564  /** Marks this function as a function that should not be profiled
2565  * when using the target feature Profile or ProfileByTimer.
2566  * This is useful when this function is does too little work at once
2567  * such that the overhead of setting the profiling token might
2568  * become significant, or that the measured time is not representative
2569  * due to modern processors (instruction level parallelism, out-of-order
2570  * execution). */
2572 
2573  /** Get a handle on the internal halide function that this Func
2574  * represents. Useful if you want to do introspection on Halide
2575  * functions */
2576  Internal::Function function() const {
2577  return func;
2578  }
2579 
2580  /** You can cast a Func to its pure stage for the purposes of
2581  * scheduling it. */
2582  operator Stage() const;
2583 
2584  /** Get a handle on the output buffer for this Func. Only relevant
2585  * if this is the output Func in a pipeline. Useful for making
2586  * static promises about strides, mins, and extents. */
2587  // @{
2589  std::vector<OutputImageParam> output_buffers() const;
2590  // @}
2591 
2592  /** Use a Func as an argument to an external stage. */
2593  operator ExternFuncArgument() const;
2594 
2595  /** Infer the arguments to the Func, sorted into a canonical order:
2596  * all buffers (sorted alphabetically by name), followed by all non-buffers
2597  * (sorted alphabetically by name).
2598  This lets you write things like:
2599  \code
2600  func.compile_to_assembly("/dev/stdout", func.infer_arguments());
2601  \endcode
2602  */
2603  std::vector<Argument> infer_arguments() const;
2604 
2605  /** Return the current StageSchedule associated with this initial
2606  * Stage of this Func. For introspection only: to modify schedule,
2607  * use the Func interface. */
2609  return Stage(*this).get_schedule();
2610  }
2611 };
2612 
2613 namespace Internal {
2614 
2615 template<typename Last>
2616 inline void check_types(const Tuple &t, int idx) {
2617  using T = typename std::remove_pointer<typename std::remove_reference<Last>::type>::type;
2618  user_assert(t[idx].type() == type_of<T>())
2619  << "Can't evaluate expression "
2620  << t[idx] << " of type " << t[idx].type()
2621  << " as a scalar of type " << type_of<T>() << "\n";
2622 }
2623 
2624 template<typename First, typename Second, typename... Rest>
2625 inline void check_types(const Tuple &t, int idx) {
2626  check_types<First>(t, idx);
2627  check_types<Second, Rest...>(t, idx + 1);
2628 }
2629 
2630 template<typename Last>
2631 inline void assign_results(Realization &r, int idx, Last last) {
2632  using T = typename std::remove_pointer<typename std::remove_reference<Last>::type>::type;
2633  *last = Buffer<T>(r[idx])();
2634 }
2635 
2636 template<typename First, typename Second, typename... Rest>
2637 inline void assign_results(Realization &r, int idx, First first, Second second, Rest &&...rest) {
2638  assign_results<First>(r, idx, first);
2639  assign_results<Second, Rest...>(r, idx + 1, second, rest...);
2640 }
2641 
2642 } // namespace Internal
2643 
2644 /** JIT-Compile and run enough code to evaluate a Halide
2645  * expression. This can be thought of as a scalar version of
2646  * \ref Func::realize */
2647 template<typename T>
2649  user_assert(e.type() == type_of<T>())
2650  << "Can't evaluate expression "
2651  << e << " of type " << e.type()
2652  << " as a scalar of type " << type_of<T>() << "\n";
2653  Func f;
2654  f() = e;
2655  Buffer<T, 0> im = f.realize(ctx);
2656  return im();
2657 }
2658 
2659 /** evaluate with a default user context */
2660 template<typename T>
2662  return evaluate<T>(nullptr, e);
2663 }
2664 
2665 /** JIT-compile and run enough code to evaluate a Halide Tuple. */
2666 template<typename First, typename... Rest>
2667 HALIDE_NO_USER_CODE_INLINE void evaluate(JITUserContext *ctx, Tuple t, First first, Rest &&...rest) {
2668  Internal::check_types<First, Rest...>(t, 0);
2669 
2670  Func f;
2671  f() = t;
2672  Realization r = f.realize(ctx);
2673  Internal::assign_results(r, 0, first, rest...);
2674 }
2675 
2676 /** JIT-compile and run enough code to evaluate a Halide Tuple. */
2677 template<typename First, typename... Rest>
2678 HALIDE_NO_USER_CODE_INLINE void evaluate(Tuple t, First first, Rest &&...rest) {
2679  evaluate<First, Rest...>(nullptr, std::move(t), std::forward<First>(first), std::forward<Rest...>(rest...));
2680 }
2681 
2682 namespace Internal {
2683 
2684 inline void schedule_scalar(Func f) {
2686  if (t.has_gpu_feature()) {
2687  f.gpu_single_thread();
2688  }
2689  if (t.has_feature(Target::HVX)) {
2690  f.hexagon();
2691  }
2692 }
2693 
2694 } // namespace Internal
2695 
2696 /** JIT-Compile and run enough code to evaluate a Halide
2697  * expression. This can be thought of as a scalar version of
2698  * \ref Func::realize. Can use GPU if jit target from environment
2699  * specifies one.
2700  */
2701 template<typename T>
2703  user_assert(e.type() == type_of<T>())
2704  << "Can't evaluate expression "
2705  << e << " of type " << e.type()
2706  << " as a scalar of type " << type_of<T>() << "\n";
2707  Func f;
2708  f() = e;
2710  Buffer<T, 0> im = f.realize();
2711  return im();
2712 }
2713 
2714 /** JIT-compile and run enough code to evaluate a Halide Tuple. Can
2715  * use GPU if jit target from environment specifies one. */
2716 // @{
2717 template<typename First, typename... Rest>
2718 HALIDE_NO_USER_CODE_INLINE void evaluate_may_gpu(Tuple t, First first, Rest &&...rest) {
2719  Internal::check_types<First, Rest...>(t, 0);
2720 
2721  Func f;
2722  f() = t;
2724  Realization r = f.realize();
2725  Internal::assign_results(r, 0, first, rest...);
2726 }
2727 // @}
2728 
2729 } // namespace Halide
2730 
2731 #endif
Defines a type used for expressing the type signature of a generated halide pipeline.
#define internal_assert(c)
Definition: Errors.h:19
Base classes for Halide expressions (Halide::Expr) and statements (Halide::Internal::Stmt)
Defines the struct representing lifetime and dependencies of a JIT compiled halide pipeline.
Defines Module, an IR container that fully describes a Halide program.
Classes for declaring scalar parameters to halide pipelines.
Defines the front-end class representing an entire Halide imaging pipeline.
Defines the front-end syntax for reduction domains and reduction variables.
Defines the structure that describes a Halide target.
Defines Tuple - the front-end handle on small arrays of expressions.
#define HALIDE_NO_USER_CODE_INLINE
Definition: Util.h:47
Defines the Var - the front-end variable.
A Halide::Buffer is a named shared reference to a Halide::Runtime::Buffer.
Definition: Buffer.h:122
Helper class for identifying purpose of an Expr passed to memoize.
Definition: Func.h:685
EvictionKey(const Expr &expr=Expr())
Definition: Func.h:691
A halide function.
Definition: Func.h:700
void print_loop_nest()
Write out the loop nests specified by the schedule for this Function.
Func & unroll(const VarOrRVar &var)
Mark a dimension to be completely unrolled.
bool is_extern() const
Is this function an external stage? That is, was it defined using define_extern?
FuncRef operator()(std::vector< Expr >) const
Either calls to the function, or the left-hand-side of an update definition (see RDom).
Func & hexagon(const VarOrRVar &x=Var::outermost())
Schedule for execution on Hexagon.
Func(const std::string &name)
Declare a new undefined function with the given name.
void compile_to_multitarget_object_files(const std::string &filename_prefix, const std::vector< Argument > &args, const std::vector< Target > &targets, const std::vector< std::string > &suffixes)
Like compile_to_multitarget_static_library(), except that the object files are all output as object f...
Func & align_extent(const Var &var, Expr modulus)
Expand the region computed so that the extent is a multiple of 'modulus'.
HALIDE_NO_USER_CODE_INLINE std::enable_if< Internal::all_are_convertible< Var, Args... >::value, FuncRef >::type operator()(Args &&...args) const
Definition: Func.h:1239
Func & always_partition_all()
Set the loop partition policy to Always for all Vars and RVar of the initial definition of the Func.
Func & hoist_storage_root()
Equivalent to Func::hoist_storage_root, but schedules storage outside the outermost loop.
Func & tile(const VarOrRVar &x, const VarOrRVar &y, const VarOrRVar &xo, const VarOrRVar &yo, const VarOrRVar &xi, const VarOrRVar &yi, const Expr &xfactor, const Expr &yfactor, TailStrategy tail=TailStrategy::Auto)
Split two dimensions at once by the given factors, and then reorder the resulting dimensions to be xi...
void specialize_fail(const std::string &message)
Add a specialization to a Func that always terminates execution with a call to halide_error().
Func & memoize(const EvictionKey &eviction_key=EvictionKey())
Use the halide_memoization_cache_...
void compile_to_assembly(const std::string &filename, const std::vector< Argument > &, const std::string &fn_name, const Target &target=get_target_from_environment())
Statically compile this function to text assembly equivalent to the object file generated by compile_...
Func & gpu_tile(const VarOrRVar &x, const VarOrRVar &y, const VarOrRVar &z, const VarOrRVar &tx, const VarOrRVar &ty, const VarOrRVar &tz, const Expr &x_size, const Expr &y_size, const Expr &z_size, TailStrategy tail=TailStrategy::Auto, DeviceAPI device_api=DeviceAPI::Default_GPU)
Func & allow_race_conditions()
Specify that race conditions are permitted for this Func, which enables parallelizing over RVars even...
bool has_update_definition() const
Does this function have at least one update definition?
void compile_jit(const Target &target=get_jit_target_from_environment())
Eagerly jit compile the function to machine code.
Func & bound_storage(const Var &dim, const Expr &bound)
Bound the extent of a Func's storage, but not extent of its compute.
Func()
Declare a new undefined function with an automatically-generated unique name.
Func & async()
Produce this Func asynchronously in a separate thread.
void compile_to_bitcode(const std::string &filename, const std::vector< Argument > &, const Target &target=get_target_from_environment())
Func & hoist_storage(const Func &f, const Var &var)
Hoist storage for this function within f's loop over var.
void infer_input_bounds(Pipeline::RealizationArg outputs, const Target &target=get_jit_target_from_environment())
HALIDE_NO_USER_CODE_INLINE std::enable_if< Internal::all_are_convertible< VarOrRVar, Args... >::value, Func & >::type always_partition(const VarOrRVar &x, Args &&...args)
Set the loop partition policy to Always for some number of Vars and RVars.
Definition: Func.h:1498
Func & gpu(const VarOrRVar &block_x, const VarOrRVar &block_y, const VarOrRVar &block_z, const VarOrRVar &thread_x, const VarOrRVar &thread_y, const VarOrRVar &thread_z, DeviceAPI device_api=DeviceAPI::Default_GPU)
Func & gpu_threads(const VarOrRVar &thread_x, const VarOrRVar &thread_y, const VarOrRVar &thread_z, DeviceAPI device_api=DeviceAPI::Default_GPU)
Func & compute_root()
Compute all of this function once ahead of time.
Func & tile(const std::vector< VarOrRVar > &previous, const std::vector< VarOrRVar > &inners, const std::vector< Expr > &factors, TailStrategy tail=TailStrategy::Auto)
Generalized tiling, reusing the previous names as the outer names.
Func & gpu(const VarOrRVar &block_x, const VarOrRVar &thread_x, DeviceAPI device_api=DeviceAPI::Default_GPU)
Tell Halide that the following dimensions correspond to GPU block indices and thread indices.
Func & compute_with(const Stage &s, const VarOrRVar &var, const std::vector< std::pair< VarOrRVar, LoopAlignStrategy >> &align)
Schedule the iteration over the initial definition of this function to be fused with another stage 's...
void compile_to_lowered_stmt(const std::string &filename, const std::vector< Argument > &args, StmtOutputFormat fmt=Text, const Target &target=get_target_from_environment())
Write out an internal representation of lowered code.
void compile_to_c(const std::string &filename, const std::vector< Argument > &, const std::string &fn_name="", const Target &target=get_target_from_environment())
Statically compile this function to C source code.
Func & fuse(const VarOrRVar &inner, const VarOrRVar &outer, const VarOrRVar &fused)
Join two dimensions into a single fused dimension.
Func & fold_storage(const Var &dim, const Expr &extent, bool fold_forward=true)
Store realizations of this function in a circular buffer of a given extent.
Func & store_at(LoopLevel loop_level)
Equivalent to the version of store_at that takes a Var, but schedules storage at a given LoopLevel.
Stage update(int idx=0)
Get a handle on an update step for the purposes of scheduling it.
Func & reorder_storage(const Var &x, const Var &y)
HALIDE_NO_USER_CODE_INLINE std::enable_if< Internal::all_are_convertible< Expr, Args... >::value, FuncRef >::type operator()(const Expr &x, Args &&...args) const
Definition: Func.h:1256
Func(const Type &required_type, int required_dims, const std::string &name)
Declare a new undefined function with the given name.
bool defined() const
Does this function have at least a pure definition.
Func(const std::vector< Type > &required_types, int required_dims, const std::string &name)
Declare a new undefined function with the given name.
Func & compute_at(LoopLevel loop_level)
Schedule a function to be computed within the iteration over a given LoopLevel.
const Internal::StageSchedule & get_schedule() const
Return the current StageSchedule associated with this initial Stage of this Func.
Definition: Func.h:2608
Func & gpu_blocks(const VarOrRVar &block_x, DeviceAPI device_api=DeviceAPI::Default_GPU)
Tell Halide that the following dimensions correspond to GPU block indices.
Func & store_at(const Func &f, const Var &var)
Allocate storage for this function within f's loop over var.
Func copy_to_host()
Declare that this function should be implemented by a call to halide_buffer_copy with a NULL target d...
Func & split(const VarOrRVar &old, const VarOrRVar &outer, const VarOrRVar &inner, const Expr &factor, TailStrategy tail=TailStrategy::Auto)
Split a dimension into inner and outer subdimensions with the given names, where the inner dimension ...
void infer_input_bounds(JITUserContext *context, Pipeline::RealizationArg outputs, const Target &target=get_jit_target_from_environment())
Func & compute_with(LoopLevel loop_level, const std::vector< std::pair< VarOrRVar, LoopAlignStrategy >> &align)
std::vector< Argument > infer_arguments() const
Infer the arguments to the Func, sorted into a canonical order: all buffers (sorted alphabetically by...
void compile_to_header(const std::string &filename, const std::vector< Argument > &, const std::string &fn_name="", const Target &target=get_target_from_environment())
Emit a header file with the given filename for this function.
const Type & type() const
Get the type(s) of the outputs of this Func.
Func & hoist_storage(LoopLevel loop_level)
Equivalent to the version of hoist_storage that takes a Var, but schedules storage at a given LoopLev...
Func & prefetch(const T &image, const VarOrRVar &at, const VarOrRVar &from, Expr offset=1, PrefetchBoundStrategy strategy=PrefetchBoundStrategy::GuardWithIf)
Definition: Func.h:2052
std::vector< Var > args() const
Get the pure arguments.
Func(const Expr &e)
Declare a new function with an automatically-generated unique name, and define it to return the given...
Func & add_trace_tag(const std::string &trace_tag)
Add a string of arbitrary text that will be passed thru to trace inspection code if the Func is reali...
int dimensions() const
The dimensionality (number of arguments) of this function.
HALIDE_NO_USER_CODE_INLINE std::enable_if< Internal::all_are_convertible< Var, Args... >::value, Func & >::type reorder_storage(const Var &x, const Var &y, Args &&...args)
Definition: Func.h:2078
Func & align_bounds(const Var &var, Expr modulus, Expr remainder=0)
Expand the region computed so that the min coordinates is congruent to 'remainder' modulo 'modulus',...
void compile_to_conceptual_stmt(const std::string &filename, const std::vector< Argument > &args, StmtOutputFormat fmt=Text, const Target &target=get_target_from_environment())
Write out a conceptual representation of lowered code, before any parallel loop get factored out into...
Func & compute_with(LoopLevel loop_level, LoopAlignStrategy align=LoopAlignStrategy::Auto)
HALIDE_NO_USER_CODE_INLINE std::enable_if< Internal::all_are_convertible< VarOrRVar, Args... >::value, Func & >::type reorder(const VarOrRVar &x, const VarOrRVar &y, Args &&...args)
Definition: Func.h:1602
Func & store_root()
Equivalent to Func::store_at, but schedules storage outside the outermost loop.
Realization realize(JITUserContext *context, std::vector< int32_t > sizes={}, const Target &target=Target())
Same as above, but takes a custom user-provided context to be passed to runtime functions.
int outputs() const
Get the number of outputs of this Func.
Func & never_partition_all()
Set the loop partition policy to Never for all Vars and RVar of the initial definition of the Func.
Tuple update_values(int idx=0) const
Get the right-hand-side of an update definition for functions that returns multiple values.
void compile_to_bitcode(const std::string &filename, const std::vector< Argument > &, const std::string &fn_name, const Target &target=get_target_from_environment())
Statically compile this function to llvm bitcode, with the given filename (which should probably end ...
int num_update_definitions() const
How many update definitions does this function have?
Func & rename(const VarOrRVar &old_name, const VarOrRVar &new_name)
Rename a dimension.
Func & vectorize(const VarOrRVar &var)
Mark a dimension to be computed all-at-once as a single vector.
Func & tile(const std::vector< VarOrRVar > &previous, const std::vector< VarOrRVar > &outers, const std::vector< VarOrRVar > &inners, const std::vector< Expr > &factors, const std::vector< TailStrategy > &tails)
A more general form of tile, which defines tiles of any dimensionality.
Func & bound_extent(const Var &var, Expr extent)
Bound the extent of a Func's realization, but not its min.
Func & trace_stores()
Trace all stores to the buffer backing this Func by emitting calls to halide_trace.
Func & set_estimates(const Region &estimates)
Set (min, extent) estimates for all dimensions in the Func at once; this is equivalent to calling set...
Stage specialize(const Expr &condition)
Specialize a Func.
Callable compile_to_callable(const std::vector< Argument > &args, const Target &target=get_jit_target_from_environment())
Eagerly jit compile the function to machine code and return a callable struct that behaves like a fun...
Func & compute_at(const Func &f, const Var &var)
Compute this function as needed for each unique value of the given var for the given calling function...
Func & tile(const std::vector< VarOrRVar > &previous, const std::vector< VarOrRVar > &outers, const std::vector< VarOrRVar > &inners, const std::vector< Expr > &factors, TailStrategy tail=TailStrategy::Auto)
The generalized tile, with a single tail strategy to apply to all vars.
Func & reorder_storage(const std::vector< Var > &dims)
Specify how the storage for the function is laid out.
Func & compute_at(const Func &f, const RVar &var)
Schedule a function to be computed within the iteration over some dimension of an update domain.
Func & gpu_tile(const VarOrRVar &x, const VarOrRVar &bx, const VarOrRVar &tx, const Expr &x_size, TailStrategy tail=TailStrategy::Auto, DeviceAPI device_api=DeviceAPI::Default_GPU)
Short-hand for tiling a domain and mapping the tile indices to GPU block indices and the coordinates ...
const std::vector< Expr > & update_args(int idx=0) const
Get the left-hand-side of the update definition.
Func & gpu_blocks(const VarOrRVar &block_x, const VarOrRVar &block_y, DeviceAPI device_api=DeviceAPI::Default_GPU)
Func & store_at(const Func &f, const RVar &var)
Equivalent to the version of store_at that takes a Var, but schedules storage within the loop over a ...
HALIDE_NO_USER_CODE_INLINE Func(Buffer< T, Dims > &im)
Construct a new Func to wrap a Buffer.
Definition: Func.h:759
void define_extern(const std::string &function_name, const std::vector< ExternFuncArgument > &params, const std::vector< Type > &types, const std::vector< Var > &arguments, NameMangling mangling=NameMangling::Default, DeviceAPI device_api=DeviceAPI::Host)
Func & parallel(const VarOrRVar &var, const Expr &task_size, TailStrategy tail=TailStrategy::Auto)
Split a dimension by the given task_size, and the parallelize the outer dimension.
JITHandlers & jit_handlers()
Get a struct containing the currently set custom functions used by JIT.
Expr value() const
The right-hand-side value of the pure definition of this function.
Func & tile(const VarOrRVar &x, const VarOrRVar &y, const VarOrRVar &xi, const VarOrRVar &yi, const Expr &xfactor, const Expr &yfactor, TailStrategy tail=TailStrategy::Auto)
A shorter form of tile, which reuses the old variable names as the new outer dimensions.
Func & hoist_storage(const Func &f, const RVar &var)
Equivalent to the version of hoist_storage that takes a Var, but schedules storage within the loop ov...
void infer_input_bounds(const std::vector< int32_t > &sizes, const Target &target=get_jit_target_from_environment())
For a given size of output, or a given output buffer, determine the bounds required of all unbound Im...
Func & gpu_tile(const VarOrRVar &x, const VarOrRVar &tx, const Expr &x_size, TailStrategy tail=TailStrategy::Auto, DeviceAPI device_api=DeviceAPI::Default_GPU)
Func clone_in(const std::vector< Func > &fs)
Module compile_to_module(const std::vector< Argument > &args, const std::string &fn_name="", const Target &target=get_target_from_environment())
Store an internal representation of lowered code as a self contained Module suitable for further comp...
void define_extern(const std::string &function_name, const std::vector< ExternFuncArgument > &params, const std::vector< Type > &types, int dimensionality, NameMangling mangling=NameMangling::Default, DeviceAPI device_api=DeviceAPI::Host)
Definition: Func.h:1172
void realize(Pipeline::RealizationArg outputs, const Target &target=Target())
Evaluate this function into an existing allocated buffer or buffers.
Func in()
Create and return a global identity wrapper, which wraps all calls to this Func by any other Func.
Func & vectorize(const VarOrRVar &var, const Expr &factor, TailStrategy tail=TailStrategy::Auto)
Split a dimension by the given factor, then vectorize the inner dimension.
OutputImageParam output_buffer() const
Get a handle on the output buffer for this Func.
Expr update_value(int idx=0) const
Get the right-hand-side of an update definition.
Func & bound(const Var &var, Expr min, Expr extent)
Statically declare that the range over which a function should be evaluated is given by the second an...
void compile_to(const std::map< OutputFileType, std::string > &output_files, const std::vector< Argument > &args, const std::string &fn_name, const Target &target=get_target_from_environment())
Compile and generate multiple target files with single call.
Func & partition(const VarOrRVar &var, Partition partition_policy)
Set the loop partition policy.
void compile_to_llvm_assembly(const std::string &filename, const std::vector< Argument > &, const Target &target=get_target_from_environment())
Func & gpu_threads(const VarOrRVar &thread_x, const VarOrRVar &thread_y, DeviceAPI device_api=DeviceAPI::Default_GPU)
void add_custom_lowering_pass(T *pass)
Add a custom pass to be used during lowering.
Definition: Func.h:1059
Func in(const std::vector< Func > &fs)
Create and return an identity wrapper shared by all the Funcs in 'fs'.
Func & gpu(const VarOrRVar &block_x, const VarOrRVar &block_y, const VarOrRVar &thread_x, const VarOrRVar &thread_y, DeviceAPI device_api=DeviceAPI::Default_GPU)
Realization realize(std::vector< int32_t > sizes={}, const Target &target=Target())
Evaluate this function over some rectangular domain and return the resulting buffer or buffers.
void realize(JITUserContext *context, Pipeline::RealizationArg outputs, const Target &target=Target())
Same as above, but takes a custom user-provided context to be passed to runtime functions.
Func & parallel(const VarOrRVar &var)
Mark a dimension to be traversed in parallel.
Func & serial(const VarOrRVar &var)
Mark a dimension to be traversed serially.
Func & never_partition(const std::vector< VarOrRVar > &vars)
Set the loop partition policy to Never for a vector of Vars and RVars.
Func & prefetch(const Func &f, const VarOrRVar &at, const VarOrRVar &from, Expr offset=1, PrefetchBoundStrategy strategy=PrefetchBoundStrategy::GuardWithIf)
Prefetch data written to or read from a Func or an ImageParam by a subsequent loop iteration,...
const std::string & name() const
The name of this function, either given during construction, or automatically generated.
void define_extern(const std::string &function_name, const std::vector< ExternFuncArgument > &params, Type t, int dimensionality, NameMangling mangling=NameMangling::Default, DeviceAPI device_api=DeviceAPI::Host)
Add an extern definition for this Func.
Definition: Func.h:1154
Func & align_storage(const Var &dim, const Expr &alignment)
Pad the storage extent of a particular dimension of realizations of this function up to be a multiple...
void compile_to_file(const std::string &filename_prefix, const std::vector< Argument > &args, const std::string &fn_name="", const Target &target=get_target_from_environment())
Compile to object file and header pair, with the given arguments.
Func & gpu_threads(const VarOrRVar &thread_x, DeviceAPI device_api=DeviceAPI::Default_GPU)
Tell Halide that the following dimensions correspond to GPU thread indices.
void add_custom_lowering_pass(Internal::IRMutator *pass, std::function< void()> deleter)
Add a custom pass to be used during lowering, with the function that will be called to delete it also...
void clear_custom_lowering_passes()
Remove all previously-set custom lowering passes.
Func & prefetch(const Parameter &param, const VarOrRVar &at, const VarOrRVar &from, Expr offset=1, PrefetchBoundStrategy strategy=PrefetchBoundStrategy::GuardWithIf)
void compile_to_llvm_assembly(const std::string &filename, const std::vector< Argument > &, const std::string &fn_name, const Target &target=get_target_from_environment())
Statically compile this function to llvm assembly, with the given filename (which should probably end...
void compile_to_multitarget_static_library(const std::string &filename_prefix, const std::vector< Argument > &args, const std::vector< Target > &targets)
Compile to static-library file and header pair once for each target; each resulting function will be ...
Func & gpu_lanes(const VarOrRVar &thread_x, DeviceAPI device_api=DeviceAPI::Default_GPU)
The given dimension corresponds to the lanes in a GPU warp.
std::vector< OutputImageParam > output_buffers() const
HALIDE_NO_USER_CODE_INLINE std::enable_if< Internal::all_are_convertible< VarOrRVar, Args... >::value, Func & >::type never_partition(const VarOrRVar &x, Args &&...args)
Set the loop partition policy to Never for some number of Vars and RVars.
Definition: Func.h:1481
Func & gpu_tile(const VarOrRVar &x, const VarOrRVar &y, const VarOrRVar &z, const VarOrRVar &bx, const VarOrRVar &by, const VarOrRVar &bz, const VarOrRVar &tx, const VarOrRVar &ty, const VarOrRVar &tz, const Expr &x_size, const Expr &y_size, const Expr &z_size, TailStrategy tail=TailStrategy::Auto, DeviceAPI device_api=DeviceAPI::Default_GPU)
Func & store_in(MemoryType memory_type)
Set the type of memory this Func should be stored in.
void compile_to_assembly(const std::string &filename, const std::vector< Argument > &, const Target &target=get_target_from_environment())
Func clone_in(const Func &f)
Similar to Func::in; however, instead of replacing the call to this Func with an identity Func that r...
std::vector< RVar > rvars(int idx=0) const
Get the RVars of the reduction domain for an update definition, if there is one.
Func & gpu_single_thread(DeviceAPI device_api=DeviceAPI::Default_GPU)
Tell Halide to run this stage using a single gpu thread and block.
Func(Internal::Function f)
Construct a new Func to wrap an existing, already-define Function object.
const std::vector< Type > & types() const
void compile_to_object(const std::string &filename, const std::vector< Argument > &, const std::string &fn_name, const Target &target=get_target_from_environment())
Statically compile this function to an object file, with the given filename (which should probably en...
const std::string & extern_function_name() const
Get the name of the extern function called for an extern definition.
Func & ring_buffer(Expr extent)
Expands the storage of the function by an extra dimension to enable ring buffering.
Func & compute_with(const Stage &s, const VarOrRVar &var, LoopAlignStrategy align=LoopAlignStrategy::Auto)
Func & trace_realizations()
Trace all realizations of this Func by emitting calls to halide_trace.
Tuple values() const
The values returned by this function.
Func & gpu_tile(const VarOrRVar &x, const VarOrRVar &y, const VarOrRVar &bx, const VarOrRVar &by, const VarOrRVar &tx, const VarOrRVar &ty, const Expr &x_size, const Expr &y_size, TailStrategy tail=TailStrategy::Auto, DeviceAPI device_api=DeviceAPI::Default_GPU)
Func & compute_inline()
Aggressively inline all uses of this function.
Func copy_to_device(DeviceAPI d=DeviceAPI::Default_GPU)
Declare that this function should be implemented by a call to halide_buffer_copy with the given targe...
Func & gpu_tile(const VarOrRVar &x, const VarOrRVar &y, const VarOrRVar &tx, const VarOrRVar &ty, const Expr &x_size, const Expr &y_size, TailStrategy tail=TailStrategy::Auto, DeviceAPI device_api=DeviceAPI::Default_GPU)
void compile_to_object(const std::string &filename, const std::vector< Argument > &, const Target &target=get_target_from_environment())
void define_extern(const std::string &function_name, const std::vector< ExternFuncArgument > &params, Type t, const std::vector< Var > &arguments, NameMangling mangling=NameMangling::Default, DeviceAPI device_api=DeviceAPI::Host)
Definition: Func.h:1182
Func & reorder(const std::vector< VarOrRVar > &vars)
Reorder variables to have the given nesting order, from innermost out.
Func & atomic(bool override_associativity_test=false)
Issue atomic updates for this Func.
const std::vector< CustomLoweringPass > & custom_lowering_passes()
Get the custom lowering passes.
void infer_input_bounds(JITUserContext *context, const std::vector< int32_t > &sizes, const Target &target=get_jit_target_from_environment())
Versions of infer_input_bounds that take a custom user context to pass to runtime functions.
void debug_to_file(const std::string &filename)
When this function is compiled, include code that dumps its values to a file after it is realized,...
Func & gpu_blocks(const VarOrRVar &block_x, const VarOrRVar &block_y, const VarOrRVar &block_z, DeviceAPI device_api=DeviceAPI::Default_GPU)
Func & no_profiling()
Marks this function as a function that should not be profiled when using the target feature Profile o...
Func in(const Func &f)
Creates and returns a new identity Func that wraps this Func.
void compile_to_static_library(const std::string &filename_prefix, const std::vector< Argument > &args, const std::string &fn_name="", const Target &target=get_target_from_environment())
Compile to static-library file and header pair, with the given arguments.
Func & set_estimate(const Var &var, const Expr &min, const Expr &extent)
Statically declare the range over which the function will be evaluated in the general case.
Func & unroll(const VarOrRVar &var, const Expr &factor, TailStrategy tail=TailStrategy::Auto)
Split a dimension by the given factor, then unroll the inner dimension.
Func & trace_loads()
Trace all loads from this Func by emitting calls to halide_trace.
FuncRef operator()(std::vector< Var >) const
Construct either the left-hand-side of a definition, or a call to a functions that happens to only co...
void define_extern(const std::string &function_name, const std::vector< ExternFuncArgument > &params, const std::vector< Type > &types, int dimensionality, NameMangling mangling)
Definition: Func.h:1164
Func & always_partition(const std::vector< VarOrRVar > &vars)
Set the loop partition policy to Always for a vector of Vars and RVars.
A fragment of front-end syntax of the form f(x, y, z), where x, y, z are Vars or Exprs.
Definition: Func.h:491
Stage operator*=(const FuncRef &)
FuncTupleElementRef operator[](int) const
When a FuncRef refers to a function that provides multiple outputs, you can access each output as an ...
Stage operator-=(const FuncRef &)
size_t size() const
How many outputs does the function this refers to produce.
Internal::Function function() const
What function is this calling?
Definition: Func.h:588
Stage operator+=(Expr)
Define a stage that adds the given expression to this Func.
Stage operator-=(Expr)
Define a stage that adds the negative of the given expression to this Func.
Stage operator*=(Expr)
Define a stage that multiplies this Func by the given expression.
Stage operator-=(const Tuple &)
Stage operator/=(Expr)
Define a stage that divides this Func by the given expression.
Stage operator+=(const FuncRef &)
Stage operator=(const Expr &)
Use this as the left-hand-side of a definition or an update definition (see RDom).
Stage operator=(const FuncRef &)
FuncRef(Internal::Function, const std::vector< Var > &, int placeholder_pos=-1, int count=0)
Stage operator+=(const Tuple &)
FuncRef(const Internal::Function &, const std::vector< Expr > &, int placeholder_pos=-1, int count=0)
Stage operator/=(const FuncRef &)
Stage operator*=(const Tuple &)
Stage operator/=(const Tuple &)
Stage operator=(const Tuple &)
Use this as the left-hand-side of a definition or an update definition for a Func with multiple outpu...
A fragment of front-end syntax of the form f(x, y, z)[index], where x, y, z are Vars or Exprs.
Definition: Func.h:610
int index() const
Return index to the function outputs.
Definition: Func.h:674
Stage operator+=(const Expr &e)
Define a stage that adds the given expression to Tuple component 'idx' of this Func.
Stage operator*=(const Expr &e)
Define a stage that multiplies Tuple component 'idx' of this Func by the given expression.
Stage operator/=(const Expr &e)
Define a stage that divides Tuple component 'idx' of this Func by the given expression.
Stage operator=(const Expr &e)
Use this as the left-hand-side of an update definition of Tuple component 'idx' of a Func (see RDom).
Stage operator=(const FuncRef &e)
Stage operator-=(const Expr &e)
Define a stage that adds the negative of the given expression to Tuple component 'idx' of this Func.
FuncTupleElementRef(const FuncRef &ref, const std::vector< Expr > &args, int idx)
An Image parameter to a halide pipeline.
Definition: ImageParam.h:23
A Function definition which can either represent a init or an update definition.
Definition: Definition.h:38
const StageSchedule & schedule() const
Get the default (no-specialization) stage-specific schedule associated with this definition.
const std::vector< Expr > & args() const
Get the default (no-specialization) arguments (left-hand-side) of the definition.
bool defined() const
Definition objects are nullable.
A reference-counted handle to Halide's internal representation of a function.
Definition: Function.h:39
A base class for passes over the IR which modify it (e.g.
Definition: IRMutator.h:26
A schedule for a single stage of a Halide pipeline.
Definition: Schedule.h:685
A reference to a site in a Halide statement at the top of the body of a particular for loop.
Definition: Schedule.h:203
A halide module.
Definition: Module.h:142
A handle on the output buffer of a pipeline.
A reference-counted handle to a parameter to a halide pipeline.
Definition: Parameter.h:40
A class representing a Halide pipeline.
Definition: Pipeline.h:107
A multi-dimensional domain over which to iterate.
Definition: RDom.h:193
A reduction variable represents a single dimension of a reduction domain (RDom).
Definition: RDom.h:29
const std::string & name() const
The name of this reduction variable.
A Realization is a vector of references to existing Buffer objects.
Definition: Realization.h:19
A single definition of a Func.
Definition: Func.h:69
Stage & prefetch(const T &image, const VarOrRVar &at, const VarOrRVar &from, Expr offset=1, PrefetchBoundStrategy strategy=PrefetchBoundStrategy::GuardWithIf)
Definition: Func.h:468
Stage & always_partition_all()
std::string name() const
Return the name of this stage, e.g.
Stage & rename(const VarOrRVar &old_name, const VarOrRVar &new_name)
Stage & gpu_threads(const VarOrRVar &thread_x, const VarOrRVar &thread_y, DeviceAPI device_api=DeviceAPI::Default_GPU)
Stage & gpu(const VarOrRVar &block_x, const VarOrRVar &block_y, const VarOrRVar &block_z, const VarOrRVar &thread_x, const VarOrRVar &thread_y, const VarOrRVar &thread_z, DeviceAPI device_api=DeviceAPI::Default_GPU)
Stage & gpu(const VarOrRVar &block_x, const VarOrRVar &block_y, const VarOrRVar &thread_x, const VarOrRVar &thread_y, DeviceAPI device_api=DeviceAPI::Default_GPU)
Stage & prefetch(const Func &f, const VarOrRVar &at, const VarOrRVar &from, Expr offset=1, PrefetchBoundStrategy strategy=PrefetchBoundStrategy::GuardWithIf)
Stage & tile(const std::vector< VarOrRVar > &previous, const std::vector< VarOrRVar > &outers, const std::vector< VarOrRVar > &inners, const std::vector< Expr > &factors, TailStrategy tail=TailStrategy::Auto)
Stage & gpu_tile(const VarOrRVar &x, const VarOrRVar &tx, const Expr &x_size, TailStrategy tail=TailStrategy::Auto, DeviceAPI device_api=DeviceAPI::Default_GPU)
HALIDE_NO_USER_CODE_INLINE std::enable_if< Internal::all_are_convertible< VarOrRVar, Args... >::value, Stage & >::type reorder(const VarOrRVar &x, const VarOrRVar &y, Args &&...args)
Definition: Func.h:383
Stage & gpu(const VarOrRVar &block_x, const VarOrRVar &thread_x, DeviceAPI device_api=DeviceAPI::Default_GPU)
Stage & gpu_tile(const VarOrRVar &x, const VarOrRVar &y, const VarOrRVar &bx, const VarOrRVar &by, const VarOrRVar &tx, const VarOrRVar &ty, const Expr &x_size, const Expr &y_size, TailStrategy tail=TailStrategy::Auto, DeviceAPI device_api=DeviceAPI::Default_GPU)
Stage & hexagon(const VarOrRVar &x=Var::outermost())
Func rfactor(const RVar &r, const Var &v)
Stage & always_partition(const std::vector< VarOrRVar > &vars)
Stage & compute_with(const Stage &s, const VarOrRVar &var, LoopAlignStrategy align=LoopAlignStrategy::Auto)
Stage & vectorize(const VarOrRVar &var)
Stage & gpu_single_thread(DeviceAPI device_api=DeviceAPI::Default_GPU)
Stage & compute_with(LoopLevel loop_level, LoopAlignStrategy align=LoopAlignStrategy::Auto)
Stage & unroll(const VarOrRVar &var)
Stage & parallel(const VarOrRVar &var)
Stage & allow_race_conditions()
Stage & serial(const VarOrRVar &var)
Stage & gpu_tile(const VarOrRVar &x, const VarOrRVar &y, const VarOrRVar &z, const VarOrRVar &bx, const VarOrRVar &by, const VarOrRVar &bz, const VarOrRVar &tx, const VarOrRVar &ty, const VarOrRVar &tz, const Expr &x_size, const Expr &y_size, const Expr &z_size, TailStrategy tail=TailStrategy::Auto, DeviceAPI device_api=DeviceAPI::Default_GPU)
Stage & gpu_tile(const VarOrRVar &x, const VarOrRVar &bx, const VarOrRVar &tx, const Expr &x_size, TailStrategy tail=TailStrategy::Auto, DeviceAPI device_api=DeviceAPI::Default_GPU)
Stage & tile(const std::vector< VarOrRVar > &previous, const std::vector< VarOrRVar > &outers, const std::vector< VarOrRVar > &inners, const std::vector< Expr > &factors, const std::vector< TailStrategy > &tails)
Stage & prefetch(const Parameter &param, const VarOrRVar &at, const VarOrRVar &from, Expr offset=1, PrefetchBoundStrategy strategy=PrefetchBoundStrategy::GuardWithIf)
HALIDE_NO_USER_CODE_INLINE std::enable_if< Internal::all_are_convertible< VarOrRVar, Args... >::value, Stage & >::type never_partition(const VarOrRVar &x, Args &&...args)
Definition: Func.h:390
Stage specialize(const Expr &condition)
Stage & compute_with(LoopLevel loop_level, const std::vector< std::pair< VarOrRVar, LoopAlignStrategy >> &align)
Schedule the iteration over this stage to be fused with another stage 's' from outermost loop to a gi...
Stage & tile(const VarOrRVar &x, const VarOrRVar &y, const VarOrRVar &xo, const VarOrRVar &yo, const VarOrRVar &xi, const VarOrRVar &yi, const Expr &xfactor, const Expr &yfactor, TailStrategy tail=TailStrategy::Auto)
Stage & split(const VarOrRVar &old, const VarOrRVar &outer, const VarOrRVar &inner, const Expr &factor, TailStrategy tail=TailStrategy::Auto)
Scheduling calls that control how the domain of this stage is traversed.
Stage & fuse(const VarOrRVar &inner, const VarOrRVar &outer, const VarOrRVar &fused)
Stage(Internal::Function f, Internal::Definition d, size_t stage_index)
Definition: Func.h:93
HALIDE_NO_USER_CODE_INLINE std::enable_if< Internal::all_are_convertible< VarOrRVar, Args... >::value, Stage & >::type always_partition(const VarOrRVar &x, Args &&...args)
Definition: Func.h:397
Stage & vectorize(const VarOrRVar &var, const Expr &factor, TailStrategy tail=TailStrategy::Auto)
Func rfactor(std::vector< std::pair< RVar, Var >> preserved)
Calling rfactor() on an associative update definition a Func will split the update into an intermedia...
Stage & parallel(const VarOrRVar &var, const Expr &task_size, TailStrategy tail=TailStrategy::Auto)
Stage & gpu_blocks(const VarOrRVar &block_x, const VarOrRVar &block_y, const VarOrRVar &block_z, DeviceAPI device_api=DeviceAPI::Default_GPU)
const Internal::StageSchedule & get_schedule() const
Return the current StageSchedule associated with this Stage.
Definition: Func.h:106
Stage & gpu_tile(const VarOrRVar &x, const VarOrRVar &y, const VarOrRVar &z, const VarOrRVar &tx, const VarOrRVar &ty, const VarOrRVar &tz, const Expr &x_size, const Expr &y_size, const Expr &z_size, TailStrategy tail=TailStrategy::Auto, DeviceAPI device_api=DeviceAPI::Default_GPU)
Stage & reorder(const std::vector< VarOrRVar > &vars)
Stage & gpu_blocks(const VarOrRVar &block_x, DeviceAPI device_api=DeviceAPI::Default_GPU)
Stage & gpu_blocks(const VarOrRVar &block_x, const VarOrRVar &block_y, DeviceAPI device_api=DeviceAPI::Default_GPU)
Stage & tile(const std::vector< VarOrRVar > &previous, const std::vector< VarOrRVar > &inners, const std::vector< Expr > &factors, TailStrategy tail=TailStrategy::Auto)
void specialize_fail(const std::string &message)
Stage & gpu_threads(const VarOrRVar &thread_x, const VarOrRVar &thread_y, const VarOrRVar &thread_z, DeviceAPI device_api=DeviceAPI::Default_GPU)
Stage & tile(const VarOrRVar &x, const VarOrRVar &y, const VarOrRVar &xi, const VarOrRVar &yi, const Expr &xfactor, const Expr &yfactor, TailStrategy tail=TailStrategy::Auto)
Stage & compute_with(const Stage &s, const VarOrRVar &var, const std::vector< std::pair< VarOrRVar, LoopAlignStrategy >> &align)
Stage & unroll(const VarOrRVar &var, const Expr &factor, TailStrategy tail=TailStrategy::Auto)
Stage & never_partition_all()
Stage & atomic(bool override_associativity_test=false)
Stage & gpu_threads(const VarOrRVar &thread_x, DeviceAPI device_api=DeviceAPI::Default_GPU)
Stage & gpu_lanes(const VarOrRVar &thread_x, DeviceAPI device_api=DeviceAPI::Default_GPU)
Stage & gpu_tile(const VarOrRVar &x, const VarOrRVar &y, const VarOrRVar &tx, const VarOrRVar &ty, const Expr &x_size, const Expr &y_size, TailStrategy tail=TailStrategy::Auto, DeviceAPI device_api=DeviceAPI::Default_GPU)
std::string dump_argument_list() const
Return a string describing the current var list taking into account all the splits,...
Stage & never_partition(const std::vector< VarOrRVar > &vars)
void unscheduled()
Assert that this stage has intentionally been given no schedule, and suppress the warning about unsch...
Stage & partition(const VarOrRVar &var, Partition partition_policy)
Create a small array of Exprs for defining and calling functions with multiple outputs.
Definition: Tuple.h:18
A Halide variable, to be used when defining functions.
Definition: Var.h:19
const std::string & name() const
Get the name of a Var.
static Var outermost()
A Var that represents the location outside the outermost loop.
Definition: Var.h:163
void schedule_scalar(Func f)
Definition: Func.h:2684
void assign_results(Realization &r, int idx, Last last)
Definition: Func.h:2631
void check_types(const Tuple &t, int idx)
Definition: Func.h:2616
ForType
An enum describing a type of loop traversal.
Definition: Expr.h:406
std::vector< Var > make_argument_list(int dimensionality)
Make a list of unique arguments for definitions with unnamed arguments.
This file defines the class FunctionDAG, which is our representation of a Halide pipeline,...
@ Internal
Not visible externally, similar to 'static' linkage in C.
PrefetchBoundStrategy
Different ways to handle accesses outside the original extents in a prefetch.
@ GuardWithIf
Guard the prefetch with if-guards that ignores the prefetch if any of the prefetched region ever goes...
HALIDE_NO_USER_CODE_INLINE T evaluate_may_gpu(const Expr &e)
JIT-Compile and run enough code to evaluate a Halide expression.
Definition: Func.h:2702
TailStrategy
Different ways to handle a tail case in a split when the factor does not provably divide the extent.
Definition: Schedule.h:33
@ Auto
For pure definitions use ShiftInwards.
LoopAlignStrategy
Different ways to handle the case when the start/end of the loops of stages computed with (fused) are...
Definition: Schedule.h:137
@ Auto
By default, LoopAlignStrategy is set to NoAlign.
Expr min(const FuncRef &a, const FuncRef &b)
Explicit overloads of min and max for FuncRef.
Definition: Func.h:597
NameMangling
An enum to specify calling convention for extern stages.
Definition: Function.h:26
@ Default
Match whatever is specified in the Target.
Target get_jit_target_from_environment()
Return the target that Halide will use for jit-compilation.
DeviceAPI
An enum describing a type of device API.
Definition: DeviceAPI.h:15
@ Host
Used to denote for loops that run on the same device as the containing code.
Target get_target_from_environment()
Return the target that Halide will use.
StmtOutputFormat
Used to determine if the output printed to file should be as a normal string or as an HTML file which...
Definition: Pipeline.h:72
@ Text
Definition: Pipeline.h:73
Stage ScheduleHandle
Definition: Func.h:482
std::vector< Range > Region
A multi-dimensional box.
Definition: Expr.h:350
Expr max(const FuncRef &a, const FuncRef &b)
Definition: Func.h:600
MemoryType
An enum describing different address spaces to be used with Func::store_in.
Definition: Expr.h:353
Partition
Different ways to handle loops with a potentially optimizable boundary conditions.
HALIDE_NO_USER_CODE_INLINE T evaluate(JITUserContext *ctx, const Expr &e)
JIT-Compile and run enough code to evaluate a Halide expression.
Definition: Func.h:2648
A fragment of Halide syntax.
Definition: Expr.h:258
HALIDE_ALWAYS_INLINE Type type() const
Get the type of this expression node.
Definition: Expr.h:327
An argument to an extern-defined Func.
A set of custom overrides of runtime functions.
Definition: JITModule.h:35
A context to be passed to Pipeline::realize.
Definition: JITModule.h:136
A struct representing a target machine and os to generate code for.
Definition: Target.h:19
bool has_gpu_feature() const
Is a fully feature GPU compute runtime enabled? I.e.
bool has_feature(Feature f) const
Types in the halide type system.
Definition: Type.h:283
A class that can represent Vars or RVars.
Definition: Func.h:29
bool is_rvar
Definition: Func.h:57
VarOrRVar(const Var &v)
Definition: Func.h:33
VarOrRVar(const RVar &r)
Definition: Func.h:36
const std::string & name() const
Definition: Func.h:47
VarOrRVar(const std::string &n, bool r)
Definition: Func.h:30
VarOrRVar(const ImplicitVar< N > &u)
Definition: Func.h:43
VarOrRVar(const RDom &r)
Definition: Func.h:39
#define user_assert(c)
Definition: test.h:10