Halide 20.0.0
Halide compiler and libraries
Loading...
Searching...
No Matches
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
23namespace Halide {
24
25class OutputImageParam;
26
27/** A class that can represent Vars or RVars. Used for reorder calls
28 * which can accept a mix of either. */
29struct 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
60class ImageParam;
61
62namespace Internal {
63struct AssociativeOp;
64class Function;
65struct Split;
66struct StorageDim;
67} // namespace Internal
68
69/** A single definition of a Func. May be a pure or update definition. */
70class Stage {
71 /** Reference to the Function this stage (or definition) belongs to. */
72 Internal::Function function;
73 Internal::Definition definition;
74 /** Indicate which stage the definition belongs to (0 for initial
75 * definition, 1 for first update, etc.). */
76 size_t stage_index;
77 /** Pure Vars of the Function (from the init definition). */
78 std::vector<Var> dim_vars;
79
80 void set_dim_type(const VarOrRVar &var, Internal::ForType t);
81 void set_dim_device_api(const VarOrRVar &var, DeviceAPI device_api);
82 void split(const std::string &old, const std::string &outer, const std::string &inner,
83 const Expr &factor, bool exact, TailStrategy tail);
84 void remove(const std::string &var);
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 std::pair<std::vector<Internal::Split>, std::vector<Internal::Split>>
93 rfactor_validate_args(const std::vector<std::pair<RVar, Var>> &preserved, const Internal::AssociativeOp &prover_result);
94
95public:
97 : function(std::move(f)), definition(std::move(d)), stage_index(stage_index) {
98 internal_assert(definition.defined());
99
100 dim_vars.reserve(function.args().size());
101 for (const auto &arg : function.args()) {
102 dim_vars.emplace_back(arg);
103 }
104 internal_assert(definition.args().size() == dim_vars.size());
105 }
106
107 /** Return the current StageSchedule associated with this Stage. For
108 * introspection only: to modify schedule, use the Func interface. */
110 return definition.schedule();
111 }
112
113 /** Return a string describing the current var list taking into
114 * account all the splits, reorders, and tiles. */
115 std::string dump_argument_list() const;
116
117 /** Return the name of this stage, e.g. "f.update(2)" */
118 std::string name() const;
119
120 /** Calling rfactor() on an associative update definition a Func will split
121 * the update into an intermediate which computes the partial results and
122 * replaces the current update definition with a new definition which merges
123 * the partial results. If called on a init/pure definition, this will
124 * throw an error. rfactor() will automatically infer the associative reduction
125 * operator and identity of the operator. If it can't prove the operation
126 * is associative or if it cannot find an identity for that operator, this
127 * will throw an error. In addition, commutativity of the operator is required
128 * if rfactor() is called on the inner dimension but excluding the outer
129 * dimensions.
130 *
131 * rfactor() takes as input 'preserved', which is a list of <RVar, Var> pairs.
132 * The rvars not listed in 'preserved' are removed from the original Func and
133 * are lifted to the intermediate Func. The remaining rvars (the ones in
134 * 'preserved') are made pure in the intermediate Func. The intermediate Func's
135 * update definition inherits all scheduling directives (e.g. split,fuse, etc.)
136 * applied to the original Func's update definition. The loop order of the
137 * intermediate Func's update definition is the same as the original, although
138 * the RVars in 'preserved' are replaced by the new pure Vars. The loop order of the
139 * intermediate Func's init definition from innermost to outermost is the args'
140 * order of the original Func's init definition followed by the new pure Vars.
141 *
142 * The intermediate Func also inherits storage order from the original Func
143 * with the new pure Vars added to the outermost.
144 *
145 * For example, f.update(0).rfactor({{r.y, u}}) would rewrite a pipeline like this:
146 \code
147 f(x, y) = 0;
148 f(x, y) += g(r.x, r.y);
149 \endcode
150 * into a pipeline like this:
151 \code
152 f_intm(x, y, u) = 0;
153 f_intm(x, y, u) += g(r.x, u);
154
155 f(x, y) = 0;
156 f(x, y) += f_intm(x, y, r.y);
157 \endcode
158 *
159 * This has a variety of uses. You can use it to split computation of an associative reduction:
160 \code
161 f(x, y) = 10;
162 RDom r(0, 96);
163 f(x, y) = max(f(x, y), g(x, y, r.x));
164 f.update(0).split(r.x, rxo, rxi, 8).reorder(y, x).parallel(x);
165 f.update(0).rfactor({{rxo, u}}).compute_root().parallel(u).update(0).parallel(u);
166 \endcode
167 *
168 *, which is equivalent to:
169 \code
170 parallel for u = 0 to 11:
171 for y:
172 for x:
173 f_intm(x, y, u) = -inf
174 parallel for x:
175 for y:
176 parallel for u = 0 to 11:
177 for rxi = 0 to 7:
178 f_intm(x, y, u) = max(f_intm(x, y, u), g(8*u + rxi))
179 for y:
180 for x:
181 f(x, y) = 10
182 parallel for x:
183 for y:
184 for rxo = 0 to 11:
185 f(x, y) = max(f(x, y), f_intm(x, y, rxo))
186 \endcode
187 *
188 */
189 // @{
190 Func rfactor(const std::vector<std::pair<RVar, Var>> &preserved);
191 Func rfactor(const RVar &r, const Var &v);
192 // @}
193
194 /** Schedule the iteration over this stage to be fused with another
195 * stage 's' from outermost loop to a given LoopLevel. 'this' stage will
196 * be computed AFTER 's' in the innermost fused dimension. There should not
197 * be any dependencies between those two fused stages. If either of the
198 * stages being fused is a stage of an extern Func, this will throw an error.
199 *
200 * Note that the two stages that are fused together should have the same
201 * exact schedule from the outermost to the innermost fused dimension, and
202 * the stage we are calling compute_with on should not have specializations,
203 * e.g. f2.compute_with(f1, x) is allowed only if f2 has no specializations.
204 *
205 * Also, if a producer is desired to be computed at the fused loop level,
206 * the function passed to the compute_at() needs to be the "parent". Consider
207 * the following code:
208 \code
209 input(x, y) = x + y;
210 f(x, y) = input(x, y);
211 f(x, y) += 5;
212 g(x, y) = x - y;
213 g(x, y) += 10;
214 f.compute_with(g, y);
215 f.update().compute_with(g.update(), y);
216 \endcode
217 *
218 * To compute 'input' at the fused loop level at dimension y, we specify
219 * input.compute_at(g, y) instead of input.compute_at(f, y) since 'g' is
220 * the "parent" for this fused loop (i.e. 'g' is computed first before 'f'
221 * is computed). On the other hand, to compute 'input' at the innermost
222 * dimension of 'f', we specify input.compute_at(f, x) instead of
223 * input.compute_at(g, x) since the x dimension of 'f' is not fused
224 * (only the y dimension is).
225 *
226 * Given the constraints, this has a variety of uses. Consider the
227 * following code:
228 \code
229 f(x, y) = x + y;
230 g(x, y) = x - y;
231 h(x, y) = f(x, y) + g(x, y);
232 f.compute_root();
233 g.compute_root();
234 f.split(x, xo, xi, 8);
235 g.split(x, xo, xi, 8);
236 g.compute_with(f, xo);
237 \endcode
238 *
239 * This is equivalent to:
240 \code
241 for y:
242 for xo:
243 for xi:
244 f(8*xo + xi) = (8*xo + xi) + y
245 for xi:
246 g(8*xo + xi) = (8*xo + xi) - y
247 for y:
248 for x:
249 h(x, y) = f(x, y) + g(x, y)
250 \endcode
251 *
252 * The size of the dimensions of the stages computed_with do not have
253 * to match. Consider the following code where 'g' is half the size of 'f':
254 \code
255 Image<int> f_im(size, size), g_im(size/2, size/2);
256 input(x, y) = x + y;
257 f(x, y) = input(x, y);
258 g(x, y) = input(2*x, 2*y);
259 g.compute_with(f, y);
260 input.compute_at(f, y);
261 Pipeline({f, g}).realize({f_im, g_im});
262 \endcode
263 *
264 * This is equivalent to:
265 \code
266 for y = 0 to size-1:
267 for x = 0 to size-1:
268 input(x, y) = x + y;
269 for x = 0 to size-1:
270 f(x, y) = input(x, y)
271 for x = 0 to size/2-1:
272 if (y < size/2-1):
273 g(x, y) = input(2*x, 2*y)
274 \endcode
275 *
276 * 'align' specifies how the loop iteration of each dimension of the
277 * two stages being fused should be aligned in the fused loop nests
278 * (see LoopAlignStrategy for options). Consider the following loop nests:
279 \code
280 for z = f_min_z to f_max_z:
281 for y = f_min_y to f_max_y:
282 for x = f_min_x to f_max_x:
283 f(x, y, z) = x + y + z
284 for z = g_min_z to g_max_z:
285 for y = g_min_y to g_max_y:
286 for x = g_min_x to g_max_x:
287 g(x, y, z) = x - y - z
288 \endcode
289 *
290 * If no alignment strategy is specified, the following loop nest will be
291 * generated:
292 \code
293 for z = min(f_min_z, g_min_z) to max(f_max_z, g_max_z):
294 for y = min(f_min_y, g_min_y) to max(f_max_y, g_max_y):
295 for x = f_min_x to f_max_x:
296 if (f_min_z <= z <= f_max_z):
297 if (f_min_y <= y <= f_max_y):
298 f(x, y, z) = x + y + z
299 for x = g_min_x to g_max_x:
300 if (g_min_z <= z <= g_max_z):
301 if (g_min_y <= y <= g_max_y):
302 g(x, y, z) = x - y - z
303 \endcode
304 *
305 * Instead, these alignment strategies:
306 \code
307 g.compute_with(f, y, {{z, LoopAlignStrategy::AlignStart}, {y, LoopAlignStrategy::AlignEnd}});
308 \endcode
309 * will produce the following loop nest:
310 \code
311 f_loop_min_z = f_min_z
312 f_loop_max_z = max(f_max_z, (f_min_z - g_min_z) + g_max_z)
313 for z = f_min_z to f_loop_max_z:
314 f_loop_min_y = min(f_min_y, (f_max_y - g_max_y) + g_min_y)
315 f_loop_max_y = f_max_y
316 for y = f_loop_min_y to f_loop_max_y:
317 for x = f_min_x to f_max_x:
318 if (f_loop_min_z <= z <= f_loop_max_z):
319 if (f_loop_min_y <= y <= f_loop_max_y):
320 f(x, y, z) = x + y + z
321 for x = g_min_x to g_max_x:
322 g_shift_z = g_min_z - f_loop_min_z
323 g_shift_y = g_max_y - f_loop_max_y
324 if (g_min_z <= (z + g_shift_z) <= g_max_z):
325 if (g_min_y <= (y + g_shift_y) <= g_max_y):
326 g(x, y + g_shift_y, z + g_shift_z) = x - (y + g_shift_y) - (z + g_shift_z)
327 \endcode
328 *
329 * LoopAlignStrategy::AlignStart on dimension z will shift the loop iteration
330 * of 'g' at dimension z so that its starting value matches that of 'f'.
331 * Likewise, LoopAlignStrategy::AlignEnd on dimension y will shift the loop
332 * iteration of 'g' at dimension y so that its end value matches that of 'f'.
333 */
334 // @{
335 Stage &compute_with(LoopLevel loop_level, const std::vector<std::pair<VarOrRVar, LoopAlignStrategy>> &align);
337 Stage &compute_with(const Stage &s, const VarOrRVar &var, const std::vector<std::pair<VarOrRVar, LoopAlignStrategy>> &align);
339 // @}
340
341 /** Scheduling calls that control how the domain of this stage is
342 * traversed. See the documentation for Func for the meanings. */
343 // @{
344
345 Stage &split(const VarOrRVar &old, const VarOrRVar &outer, const VarOrRVar &inner, const Expr &factor, TailStrategy tail = TailStrategy::Auto);
346 Stage &fuse(const VarOrRVar &inner, const VarOrRVar &outer, const VarOrRVar &fused);
347 Stage &serial(const VarOrRVar &var);
350 Stage &unroll(const VarOrRVar &var);
352 Stage &vectorize(const VarOrRVar &var, const Expr &factor, TailStrategy tail = TailStrategy::Auto);
353 Stage &unroll(const VarOrRVar &var, const Expr &factor, TailStrategy tail = TailStrategy::Auto);
354 Stage &partition(const VarOrRVar &var, Partition partition_policy);
356 Stage &never_partition(const std::vector<VarOrRVar> &vars);
358 Stage &always_partition(const std::vector<VarOrRVar> &vars);
359
360 Stage &tile(const VarOrRVar &x, const VarOrRVar &y,
361 const VarOrRVar &xo, const VarOrRVar &yo,
362 const VarOrRVar &xi, const VarOrRVar &yi, const Expr &xfactor, const Expr &yfactor,
364 Stage &tile(const VarOrRVar &x, const VarOrRVar &y,
365 const VarOrRVar &xi, const VarOrRVar &yi,
366 const Expr &xfactor, const Expr &yfactor,
368 Stage &tile(const std::vector<VarOrRVar> &previous,
369 const std::vector<VarOrRVar> &outers,
370 const std::vector<VarOrRVar> &inners,
371 const std::vector<Expr> &factors,
372 const std::vector<TailStrategy> &tails);
373 Stage &tile(const std::vector<VarOrRVar> &previous,
374 const std::vector<VarOrRVar> &outers,
375 const std::vector<VarOrRVar> &inners,
376 const std::vector<Expr> &factors,
378 Stage &tile(const std::vector<VarOrRVar> &previous,
379 const std::vector<VarOrRVar> &inners,
380 const std::vector<Expr> &factors,
382 Stage &reorder(const std::vector<VarOrRVar> &vars);
383
384 template<typename... Args>
385 HALIDE_NO_USER_CODE_INLINE typename std::enable_if<Internal::all_are_convertible<VarOrRVar, Args...>::value, Stage &>::type
386 reorder(const VarOrRVar &x, const VarOrRVar &y, Args &&...args) {
387 std::vector<VarOrRVar> collected_args{x, y, std::forward<Args>(args)...};
388 return reorder(collected_args);
389 }
390
391 template<typename... Args>
392 HALIDE_NO_USER_CODE_INLINE typename std::enable_if<Internal::all_are_convertible<VarOrRVar, Args...>::value, Stage &>::type
393 never_partition(const VarOrRVar &x, Args &&...args) {
394 std::vector<VarOrRVar> collected_args{x, std::forward<Args>(args)...};
396 }
397
398 template<typename... Args>
399 HALIDE_NO_USER_CODE_INLINE typename std::enable_if<Internal::all_are_convertible<VarOrRVar, Args...>::value, Stage &>::type
400 always_partition(const VarOrRVar &x, Args &&...args) {
401 std::vector<VarOrRVar> collected_args{x, std::forward<Args>(args)...};
403 }
404
406 Stage specialize(const Expr &condition);
407 void specialize_fail(const std::string &message);
408
412
414
416
420
423 const VarOrRVar &thread_x, const VarOrRVar &thread_y,
424 DeviceAPI device_api = DeviceAPI::Default_GPU);
426 const VarOrRVar &thread_x, const VarOrRVar &thread_y, const VarOrRVar &thread_z,
427 DeviceAPI device_api = DeviceAPI::Default_GPU);
428
429 Stage &gpu_tile(const VarOrRVar &x, const VarOrRVar &bx, const VarOrRVar &tx, const Expr &x_size,
431 DeviceAPI device_api = DeviceAPI::Default_GPU);
432
433 Stage &gpu_tile(const VarOrRVar &x, const VarOrRVar &tx, const Expr &x_size,
435 DeviceAPI device_api = DeviceAPI::Default_GPU);
436 Stage &gpu_tile(const VarOrRVar &x, const VarOrRVar &y,
437 const VarOrRVar &bx, const VarOrRVar &by,
438 const VarOrRVar &tx, const VarOrRVar &ty,
439 const Expr &x_size, const Expr &y_size,
441 DeviceAPI device_api = DeviceAPI::Default_GPU);
442
443 Stage &gpu_tile(const VarOrRVar &x, const VarOrRVar &y,
444 const VarOrRVar &tx, const VarOrRVar &ty,
445 const Expr &x_size, const Expr &y_size,
447 DeviceAPI device_api = DeviceAPI::Default_GPU);
448
449 Stage &gpu_tile(const VarOrRVar &x, const VarOrRVar &y, const VarOrRVar &z,
450 const VarOrRVar &bx, const VarOrRVar &by, const VarOrRVar &bz,
451 const VarOrRVar &tx, const VarOrRVar &ty, const VarOrRVar &tz,
452 const Expr &x_size, const Expr &y_size, const Expr &z_size,
454 DeviceAPI device_api = DeviceAPI::Default_GPU);
455 Stage &gpu_tile(const VarOrRVar &x, const VarOrRVar &y, const VarOrRVar &z,
456 const VarOrRVar &tx, const VarOrRVar &ty, const VarOrRVar &tz,
457 const Expr &x_size, const Expr &y_size, const Expr &z_size,
459 DeviceAPI device_api = DeviceAPI::Default_GPU);
460
463
465
466 Stage &prefetch(const Func &f, const VarOrRVar &at, const VarOrRVar &from, Expr offset = 1,
468 Stage &prefetch(const Parameter &param, const VarOrRVar &at, const VarOrRVar &from, Expr offset = 1,
470 template<typename T>
471 Stage &prefetch(const T &image, const VarOrRVar &at, const VarOrRVar &from, Expr offset = 1,
473 return prefetch(image.parameter(), at, from, std::move(offset), strategy);
474 }
475 // @}
476
477 /** Assert that this stage has intentionally been given no schedule, and
478 * suppress the warning about unscheduled update definitions that would
479 * otherwise fire. This counts as a schedule, so calling this twice on the
480 * same Stage will fail the assertion. */
482};
483
484// For backwards compatibility, keep the ScheduleHandle name.
486
488
489/** A fragment of front-end syntax of the form f(x, y, z), where x, y,
490 * z are Vars or Exprs. If could be the left hand side of a definition or
491 * an update definition, or it could be a call to a function. We don't know
492 * until we see how this object gets used.
493 */
494class FuncRef {
496 int implicit_placeholder_pos;
497 int implicit_count;
498 std::vector<Expr> args;
499 std::vector<Expr> args_with_implicit_vars(const std::vector<Expr> &e) const;
500
501 /** Helper for function update by Tuple. If the function does not
502 * already have a pure definition, init_val will be used as RHS of
503 * each tuple element in the initial function definition. */
504 template<typename BinaryOp>
505 Stage func_ref_update(const Tuple &e, int init_val);
506
507 /** Helper for function update by Expr. If the function does not
508 * already have a pure definition, init_val will be used as RHS in
509 * the initial function definition. */
510 template<typename BinaryOp>
511 Stage func_ref_update(const Expr &e, int init_val);
512
513public:
514 FuncRef(const Internal::Function &, const std::vector<Expr> &,
515 int placeholder_pos = -1, int count = 0);
516 FuncRef(Internal::Function, const std::vector<Var> &,
517 int placeholder_pos = -1, int count = 0);
518
519 /** Use this as the left-hand-side of a definition or an update definition
520 * (see \ref RDom).
521 */
523
524 /** Use this as the left-hand-side of a definition or an update definition
525 * for a Func with multiple outputs. */
527
528 /** Define a stage that adds the given expression to this Func. If the
529 * expression refers to some RDom, this performs a sum reduction of the
530 * expression over the domain. If the function does not already have a
531 * pure definition, this sets it to zero.
532 */
533 // @{
537 // @}
538
539 /** Define a stage that adds the negative of the given expression to this
540 * Func. If the expression refers to some RDom, this performs a sum reduction
541 * of the negative of the expression over the domain. If the function does
542 * not already have a pure definition, this sets it to zero.
543 */
544 // @{
548 // @}
549
550 /** Define a stage that multiplies this Func by the given expression. If the
551 * expression refers to some RDom, this performs a product reduction of the
552 * expression over the domain. If the function does not already have a pure
553 * definition, this sets it to 1.
554 */
555 // @{
559 // @}
560
561 /** Define a stage that divides this Func by the given expression.
562 * If the expression refers to some RDom, this performs a product
563 * reduction of the inverse of the expression over the domain. If the
564 * function does not already have a pure definition, this sets it to 1.
565 */
566 // @{
570 // @}
571
572 /* Override the usual assignment operator, so that
573 * f(x, y) = g(x, y) defines f.
574 */
576
577 /** Use this as a call to the function, and not the left-hand-side
578 * of a definition. Only works for single-output Funcs. */
579 operator Expr() const;
580
581 /** When a FuncRef refers to a function that provides multiple
582 * outputs, you can access each output as an Expr using
583 * operator[].
584 */
586
587 /** How many outputs does the function this refers to produce. */
588 size_t size() const;
589
590 /** What function is this calling? */
592 return func;
593 }
594};
595
596/** Explicit overloads of min and max for FuncRef. These exist to
597 * disambiguate calls to min on FuncRefs when a user has pulled both
598 * Halide::min and std::min into their namespace. */
599// @{
600inline Expr min(const FuncRef &a, const FuncRef &b) {
601 return min(Expr(a), Expr(b));
602}
603inline Expr max(const FuncRef &a, const FuncRef &b) {
604 return max(Expr(a), Expr(b));
605}
606// @}
607
608/** A fragment of front-end syntax of the form f(x, y, z)[index], where x, y,
609 * z are Vars or Exprs. If could be the left hand side of an update
610 * definition, or it could be a call to a function. We don't know
611 * until we see how this object gets used.
612 */
614 FuncRef func_ref;
615 std::vector<Expr> args; // args to the function
616 int idx; // Index to function outputs
617
618 /** Helper function that generates a Tuple where element at 'idx' is set
619 * to 'e' and the rests are undef. */
620 Tuple values_with_undefs(const Expr &e) const;
621
622public:
623 FuncTupleElementRef(const FuncRef &ref, const std::vector<Expr> &args, int idx);
624
625 /** Use this as the left-hand-side of an update definition of Tuple
626 * component 'idx' of a Func (see \ref RDom). The function must
627 * already have an initial definition.
628 */
630
631 /** Define a stage that adds the given expression to Tuple component 'idx'
632 * of this Func. The other Tuple components are unchanged. If the expression
633 * refers to some RDom, this performs a sum reduction of the expression over
634 * the domain. The function must already have an initial definition.
635 */
637
638 /** Define a stage that adds the negative of the given expression to Tuple
639 * component 'idx' of this Func. The other Tuple components are unchanged.
640 * If the expression refers to some RDom, this performs a sum reduction of
641 * the negative of the expression over the domain. The function must already
642 * have an initial definition.
643 */
645
646 /** Define a stage that multiplies Tuple component 'idx' of this Func by
647 * the given expression. The other Tuple components are unchanged. If the
648 * expression refers to some RDom, this performs a product reduction of
649 * the expression over the domain. The function must already have an
650 * initial definition.
651 */
653
654 /** Define a stage that divides Tuple component 'idx' of this Func by
655 * the given expression. The other Tuple components are unchanged.
656 * If the expression refers to some RDom, this performs a product
657 * reduction of the inverse of the expression over the domain. The function
658 * must already have an initial definition.
659 */
661
662 /* Override the usual assignment operator, so that
663 * f(x, y)[index] = g(x, y) defines f.
664 */
666
667 /** Use this as a call to Tuple component 'idx' of a Func, and not the
668 * left-hand-side of a definition. */
669 operator Expr() const;
670
671 /** What function is this calling? */
673 return func_ref.function();
674 }
675
676 /** Return index to the function outputs. */
677 int index() const {
678 return idx;
679 }
680};
681
682namespace Internal {
683class IRMutator;
684} // namespace Internal
685
686/** Helper class for identifying purpose of an Expr passed to memoize.
687 */
689protected:
691 friend class Func;
692
693public:
694 explicit EvictionKey(const Expr &expr = Expr())
695 : key(expr) {
696 }
697};
698
699/** A halide function. This class represents one stage in a Halide
700 * pipeline, and is the unit by which we schedule things. By default
701 * they are aggressively inlined, so you are encouraged to make lots
702 * of little functions, rather than storing things in Exprs. */
703class Func {
704
705 /** A handle on the internal halide function that this
706 * represents */
708
709 /** When you make a reference to this function with fewer
710 * arguments than it has dimensions, the argument list is bulked
711 * up with 'implicit' vars with canonical names. This lets you
712 * pass around partially applied Halide functions. */
713 // @{
714 std::pair<int, int> add_implicit_vars(std::vector<Var> &) const;
715 std::pair<int, int> add_implicit_vars(std::vector<Expr> &) const;
716 // @}
717
718 /** The imaging pipeline that outputs this Func alone. */
719 Pipeline pipeline_;
720
721 /** Get the imaging pipeline that outputs this Func alone,
722 * creating it (and freezing the Func) if necessary. */
723 Pipeline pipeline();
724
725 // Helper function for recursive reordering support
726 Func &reorder_storage(const std::vector<Var> &dims, size_t start);
727
728 void invalidate_cache();
729
730public:
731 /** Declare a new undefined function with the given name */
732 explicit Func(const std::string &name);
733
734 /** Declare a new undefined function with the given name.
735 * The function will be constrained to represent Exprs of required_type.
736 * If required_dims is not AnyDims, the function will be constrained to exactly
737 * that many dimensions. */
738 explicit Func(const Type &required_type, int required_dims, const std::string &name);
739
740 /** Declare a new undefined function with the given name.
741 * If required_types is not empty, the function will be constrained to represent
742 * Tuples of the same arity and types. (If required_types is empty, there is no constraint.)
743 * If required_dims is not AnyDims, the function will be constrained to exactly
744 * that many dimensions. */
745 explicit Func(const std::vector<Type> &required_types, int required_dims, const std::string &name);
746
747 /** Declare a new undefined function with an
748 * automatically-generated unique name */
750
751 /** Declare a new function with an automatically-generated unique
752 * name, and define it to return the given expression (which may
753 * not contain free variables). */
754 explicit Func(const Expr &e);
755
756 /** Construct a new Func to wrap an existing, already-define
757 * Function object. */
759
760 /** Construct a new Func to wrap a Buffer. */
761 template<typename T, int Dims>
763 : Func() {
764 (*this)(_) = im(_);
765 }
766
767 /** Evaluate this function over some rectangular domain and return
768 * the resulting buffer or buffers. Performs compilation if the
769 * Func has not previously been realized and compile_jit has not
770 * been called. If the final stage of the pipeline is on the GPU,
771 * data is copied back to the host before being returned. The
772 * returned Realization should probably be instantly converted to
773 * a Buffer class of the appropriate type. That is, do this:
774 *
775 \code
776 f(x) = sin(x);
777 Buffer<float> im = f.realize(...);
778 \endcode
779 *
780 * If your Func has multiple values, because you defined it using
781 * a Tuple, then casting the result of a realize call to a buffer
782 * or image will produce a run-time error. Instead you should do the
783 * following:
784 *
785 \code
786 f(x) = Tuple(x, sin(x));
787 Realization r = f.realize(...);
788 Buffer<int> im0 = r[0];
789 Buffer<float> im1 = r[1];
790 \endcode
791 *
792 * In Halide formal arguments of a computation are specified using
793 * Param<T> and ImageParam objects in the expressions defining the
794 * computation. Note that this method is not thread-safe, in that
795 * Param<T> and ImageParam are globals shared by all threads; to call
796 * jitted code in a thread-safe manner, use compile_to_callable() instead.
797 *
798 \code
799 Param<int32> p(42);
800 ImageParam img(Int(32), 1);
801 f(x) = img(x) + p;
802
803 Buffer<int32_t) arg_img(10, 10);
804 <fill in arg_img...>
805
806 Target t = get_jit_target_from_environment();
807 Buffer<int32_t> result = f.realize({10, 10}, t);
808 \endcode
809 *
810 * Alternatively, an initializer list can be used
811 * directly in the realize call to pass this information:
812 *
813 \code
814 Param<int32> p(42);
815 ImageParam img(Int(32), 1);
816 f(x) = img(x) + p;
817
818 Buffer<int32_t) arg_img(10, 10);
819 <fill in arg_img...>
820
821 Target t = get_jit_target_from_environment();
822 Buffer<int32_t> result = f.realize({10, 10}, t, { { p, 17 }, { img, arg_img } });
823 \endcode
824 *
825 * If the Func cannot be realized into a buffer of the given size
826 * due to scheduling constraints on scattering update definitions,
827 * it will be realized into a larger buffer of the minimum size
828 * possible, and a cropped view at the requested size will be
829 * returned. It is thus not safe to assume the returned buffers
830 * are contiguous in memory. This behavior can be disabled with
831 * the NoBoundsQuery target flag, in which case an error about
832 * writing out of bounds on the output buffer will trigger
833 * instead.
834 *
835 */
836 Realization realize(std::vector<int32_t> sizes = {}, const Target &target = Target());
837
838 /** Same as above, but takes a custom user-provided context to be
839 * passed to runtime functions. This can be used to pass state to
840 * runtime overrides in a thread-safe manner. A nullptr context is
841 * legal, and is equivalent to calling the variant of realize
842 * that does not take a context. */
844 std::vector<int32_t> sizes = {},
845 const Target &target = Target());
846
847 /** Evaluate this function into an existing allocated buffer or
848 * buffers. If the buffer is also one of the arguments to the
849 * function, strange things may happen, as the pipeline isn't
850 * necessarily safe to run in-place. If you pass multiple buffers,
851 * they must have matching sizes. This form of realize does *not*
852 * automatically copy data back from the GPU. */
854
855 /** Same as above, but takes a custom user-provided context to be
856 * passed to runtime functions. This can be used to pass state to
857 * runtime overrides in a thread-safe manner. A nullptr context is
858 * legal, and is equivalent to calling the variant of realize
859 * that does not take a context. */
860 void realize(JITUserContext *context,
862 const Target &target = Target());
863
864 /** For a given size of output, or a given output buffer,
865 * determine the bounds required of all unbound ImageParams
866 * referenced. Communicates the result by allocating new buffers
867 * of the appropriate size and binding them to the unbound
868 * ImageParams.
869 */
870 // @{
871 void infer_input_bounds(const std::vector<int32_t> &sizes,
872 const Target &target = get_jit_target_from_environment());
874 const Target &target = get_jit_target_from_environment());
875 // @}
876
877 /** Versions of infer_input_bounds that take a custom user context
878 * to pass to runtime functions. */
879 // @{
881 const std::vector<int32_t> &sizes,
882 const Target &target = get_jit_target_from_environment());
885 const Target &target = get_jit_target_from_environment());
886 // @}
887 /** Statically compile this function to llvm bitcode, with the
888 * given filename (which should probably end in .bc), type
889 * signature, and C function name (which defaults to the same name
890 * as this halide function */
891 //@{
892 void compile_to_bitcode(const std::string &filename, const std::vector<Argument> &, const std::string &fn_name,
893 const Target &target = get_target_from_environment());
894 void compile_to_bitcode(const std::string &filename, const std::vector<Argument> &,
895 const Target &target = get_target_from_environment());
896 // @}
897
898 /** Statically compile this function to llvm assembly, with the
899 * given filename (which should probably end in .ll), type
900 * signature, and C function name (which defaults to the same name
901 * as this halide function */
902 //@{
903 void compile_to_llvm_assembly(const std::string &filename, const std::vector<Argument> &, const std::string &fn_name,
904 const Target &target = get_target_from_environment());
905 void compile_to_llvm_assembly(const std::string &filename, const std::vector<Argument> &,
906 const Target &target = get_target_from_environment());
907 // @}
908
909 /** Statically compile this function to an object file, with the
910 * given filename (which should probably end in .o or .obj), type
911 * signature, and C function name (which defaults to the same name
912 * as this halide function. You probably don't want to use this
913 * directly; call compile_to_static_library or compile_to_file instead. */
914 //@{
915 void compile_to_object(const std::string &filename, const std::vector<Argument> &, const std::string &fn_name,
916 const Target &target = get_target_from_environment());
917 void compile_to_object(const std::string &filename, const std::vector<Argument> &,
918 const Target &target = get_target_from_environment());
919 // @}
920
921 /** Emit a header file with the given filename for this
922 * function. The header will define a function with the type
923 * signature given by the second argument, and a name given by the
924 * third. The name defaults to the same name as this halide
925 * function. You don't actually have to have defined this function
926 * yet to call this. You probably don't want to use this directly;
927 * call compile_to_static_library or compile_to_file instead. */
928 void compile_to_header(const std::string &filename, const std::vector<Argument> &, const std::string &fn_name = "",
929 const Target &target = get_target_from_environment());
930
931 /** Statically compile this function to text assembly equivalent
932 * to the object file generated by compile_to_object. This is
933 * useful for checking what Halide is producing without having to
934 * disassemble anything, or if you need to feed the assembly into
935 * some custom toolchain to produce an object file (e.g. iOS) */
936 //@{
937 void compile_to_assembly(const std::string &filename, const std::vector<Argument> &, const std::string &fn_name,
938 const Target &target = get_target_from_environment());
939 void compile_to_assembly(const std::string &filename, const std::vector<Argument> &,
940 const Target &target = get_target_from_environment());
941 // @}
942
943 /** Statically compile this function to C source code. This is
944 * useful for providing fallback code paths that will compile on
945 * many platforms. Vectorization will fail, and parallelization
946 * will produce serial code. */
947 void compile_to_c(const std::string &filename,
948 const std::vector<Argument> &,
949 const std::string &fn_name = "",
950 const Target &target = get_target_from_environment());
951
952 /** Write out an internal representation of lowered code. Useful
953 * for analyzing and debugging scheduling. Can emit html or plain
954 * text. */
955 void compile_to_lowered_stmt(const std::string &filename,
956 const std::vector<Argument> &args,
958 const Target &target = get_target_from_environment());
959
960 /** Write out a conceptual representation of lowered code, before any parallel loop
961 * get factored out into separate functions, or GPU loops are offloaded to kernel code.r
962 * Useful for analyzing and debugging scheduling. Can emit html or plain text. */
963 void compile_to_conceptual_stmt(const std::string &filename,
964 const std::vector<Argument> &args,
966 const Target &target = get_target_from_environment());
967
968 /** Write out the loop nests specified by the schedule for this
969 * Function. Helpful for understanding what a schedule is
970 * doing. */
972
973 /** Compile to object file and header pair, with the given
974 * arguments. The name defaults to the same name as this halide
975 * function.
976 */
977 void compile_to_file(const std::string &filename_prefix, const std::vector<Argument> &args,
978 const std::string &fn_name = "",
979 const Target &target = get_target_from_environment());
980
981 /** Compile to static-library file and header pair, with the given
982 * arguments. The name defaults to the same name as this halide
983 * function.
984 */
985 void compile_to_static_library(const std::string &filename_prefix, const std::vector<Argument> &args,
986 const std::string &fn_name = "",
987 const Target &target = get_target_from_environment());
988
989 /** Compile to static-library file and header pair once for each target;
990 * each resulting function will be considered (in order) via halide_can_use_target_features()
991 * at runtime, with the first appropriate match being selected for subsequent use.
992 * This is typically useful for specializations that may vary unpredictably by machine
993 * (e.g., SSE4.1/AVX/AVX2 on x86 desktop machines).
994 * All targets must have identical arch-os-bits.
995 */
997 const std::vector<Argument> &args,
998 const std::vector<Target> &targets);
999
1000 /** Like compile_to_multitarget_static_library(), except that the object files
1001 * are all output as object files (rather than bundled into a static library).
1002 *
1003 * `suffixes` is an optional list of strings to use for as the suffix for each object
1004 * file. If nonempty, it must be the same length as `targets`. (If empty, Target::to_string()
1005 * will be used for each suffix.)
1006 *
1007 * Note that if `targets.size()` > 1, the wrapper code (to select the subtarget)
1008 * will be generated with the filename `${filename_prefix}_wrapper.o`
1009 *
1010 * Note that if `targets.size()` > 1 and `no_runtime` is not specified, the runtime
1011 * will be generated with the filename `${filename_prefix}_runtime.o`
1012 */
1014 const std::vector<Argument> &args,
1015 const std::vector<Target> &targets,
1016 const std::vector<std::string> &suffixes);
1017
1018 /** Store an internal representation of lowered code as a self
1019 * contained Module suitable for further compilation. */
1020 Module compile_to_module(const std::vector<Argument> &args, const std::string &fn_name = "",
1021 const Target &target = get_target_from_environment());
1022
1023 /** Compile and generate multiple target files with single call.
1024 * Deduces target files based on filenames specified in
1025 * output_files map.
1026 */
1027 void compile_to(const std::map<OutputFileType, std::string> &output_files,
1028 const std::vector<Argument> &args,
1029 const std::string &fn_name,
1030 const Target &target = get_target_from_environment());
1031
1032 /** Eagerly jit compile the function to machine code. This
1033 * normally happens on the first call to realize. If you're
1034 * running your halide pipeline inside time-sensitive code and
1035 * wish to avoid including the time taken to compile a pipeline,
1036 * then you can call this ahead of time. Default is to use the Target
1037 * returned from Halide::get_jit_target_from_environment()
1038 */
1040
1041 /** Get a struct containing the currently set custom functions
1042 * used by JIT. This can be mutated. Changes will take effect the
1043 * next time this Func is realized. */
1045
1046 /** Eagerly jit compile the function to machine code and return a callable
1047 * struct that behaves like a function pointer. The calling convention
1048 * will exactly match that of an AOT-compiled version of this Func
1049 * with the same Argument list.
1050 */
1051 Callable compile_to_callable(const std::vector<Argument> &args,
1052 const Target &target = get_jit_target_from_environment());
1053
1054 /** Add a custom pass to be used during lowering. It is run after
1055 * all other lowering passes. Can be used to verify properties of
1056 * the lowered Stmt, instrument it with extra code, or otherwise
1057 * modify it. The Func takes ownership of the pass, and will call
1058 * delete on it when the Func goes out of scope. So don't pass a
1059 * stack object, or share pass instances between multiple
1060 * Funcs. */
1061 template<typename T>
1063 // Template instantiate a custom deleter for this type, then
1064 // wrap in a lambda. The custom deleter lives in user code, so
1065 // that deletion is on the same heap as construction (I hate Windows).
1066 add_custom_lowering_pass(pass, [pass]() { delete_lowering_pass<T>(pass); });
1067 }
1068
1069 /** Add a custom pass to be used during lowering, with the
1070 * function that will be called to delete it also passed in. Set
1071 * it to nullptr if you wish to retain ownership of the object. */
1072 void add_custom_lowering_pass(Internal::IRMutator *pass, std::function<void()> deleter);
1073
1074 /** Remove all previously-set custom lowering passes */
1076
1077 /** Get the custom lowering passes. */
1078 const std::vector<CustomLoweringPass> &custom_lowering_passes();
1079
1080 /** When this function is compiled, include code that dumps its
1081 * values to a file after it is realized, for the purpose of
1082 * debugging.
1083 *
1084 * If filename ends in ".tif" or ".tiff" (case insensitive) the file
1085 * is in TIFF format and can be read by standard tools. Oherwise, the
1086 * file format is as follows:
1087 *
1088 * All data is in the byte-order of the target platform. First, a
1089 * 20 byte-header containing four 32-bit ints, giving the extents
1090 * of the first four dimensions. Dimensions beyond four are
1091 * folded into the fourth. Then, a fifth 32-bit int giving the
1092 * data type of the function. The typecodes are given by: float =
1093 * 0, double = 1, uint8_t = 2, int8_t = 3, uint16_t = 4, int16_t =
1094 * 5, uint32_t = 6, int32_t = 7, uint64_t = 8, int64_t = 9. The
1095 * data follows the header, as a densely packed array of the given
1096 * size and the given type. If given the extension .tmp, this file
1097 * format can be natively read by the program ImageStack. */
1098 void debug_to_file(const std::string &filename);
1099
1100 /** The name of this function, either given during construction,
1101 * or automatically generated. */
1102 const std::string &name() const;
1103
1104 /** Get the pure arguments. */
1105 std::vector<Var> args() const;
1106
1107 /** The right-hand-side value of the pure definition of this
1108 * function. Causes an error if there's no pure definition, or if
1109 * the function is defined to return multiple values. */
1110 Expr value() const;
1111
1112 /** The values returned by this function. An error if the function
1113 * has not been been defined. Returns a Tuple with one element for
1114 * functions defined to return a single value. */
1115 Tuple values() const;
1116
1117 /** Does this function have at least a pure definition. */
1118 bool defined() const;
1119
1120 /** Get the left-hand-side of the update definition. An empty
1121 * vector if there's no update definition. If there are
1122 * multiple update definitions for this function, use the
1123 * argument to select which one you want. */
1124 const std::vector<Expr> &update_args(int idx = 0) const;
1125
1126 /** Get the right-hand-side of an update definition. An error if
1127 * there's no update definition. If there are multiple
1128 * update definitions for this function, use the argument to
1129 * select which one you want. */
1130 Expr update_value(int idx = 0) const;
1131
1132 /** Get the right-hand-side of an update definition for
1133 * functions that returns multiple values. An error if there's no
1134 * update definition. Returns a Tuple with one element for
1135 * functions that return a single value. */
1136 Tuple update_values(int idx = 0) const;
1137
1138 /** Get the RVars of the reduction domain for an update definition, if there is
1139 * one. */
1140 std::vector<RVar> rvars(int idx = 0) const;
1141
1142 /** Does this function have at least one update definition? */
1144
1145 /** How many update definitions does this function have? */
1147
1148 /** Is this function an external stage? That is, was it defined
1149 * using define_extern? */
1150 bool is_extern() const;
1151
1152 /** Add an extern definition for this Func. This lets you define a
1153 * Func that represents an external pipeline stage. You can, for
1154 * example, use it to wrap a call to an extern library such as
1155 * fftw. */
1156 // @{
1157 void define_extern(const std::string &function_name,
1158 const std::vector<ExternFuncArgument> &params, Type t,
1159 int dimensionality,
1161 DeviceAPI device_api = DeviceAPI::Host) {
1162 define_extern(function_name, params, t,
1164 device_api);
1165 }
1166
1167 void define_extern(const std::string &function_name,
1168 const std::vector<ExternFuncArgument> &params,
1169 const std::vector<Type> &types, int dimensionality,
1171 define_extern(function_name, params, types,
1173 }
1174
1175 void define_extern(const std::string &function_name,
1176 const std::vector<ExternFuncArgument> &params,
1177 const std::vector<Type> &types, int dimensionality,
1179 DeviceAPI device_api = DeviceAPI::Host) {
1180 define_extern(function_name, params, types,
1182 device_api);
1183 }
1184
1185 void define_extern(const std::string &function_name,
1186 const std::vector<ExternFuncArgument> &params, Type t,
1187 const std::vector<Var> &arguments,
1189 DeviceAPI device_api = DeviceAPI::Host) {
1190 define_extern(function_name, params, std::vector<Type>{t}, arguments,
1191 mangling, device_api);
1192 }
1193
1194 void define_extern(const std::string &function_name,
1195 const std::vector<ExternFuncArgument> &params,
1196 const std::vector<Type> &types,
1197 const std::vector<Var> &arguments,
1199 DeviceAPI device_api = DeviceAPI::Host);
1200 // @}
1201
1202 /** Get the type(s) of the outputs of this Func.
1203 *
1204 * It is not legal to call type() unless the Func has non-Tuple elements.
1205 *
1206 * If the Func isn't yet defined, and was not specified with required types,
1207 * a runtime error will occur.
1208 *
1209 * If the Func isn't yet defined, but *was* specified with required types,
1210 * the requirements will be returned. */
1211 // @{
1212 const Type &type() const;
1213 const std::vector<Type> &types() const;
1214 // @}
1215
1216 /** Get the number of outputs of this Func. Corresponds to the
1217 * size of the Tuple this Func was defined to return.
1218 * If the Func isn't yet defined, but was specified with required types,
1219 * the number of outputs specified in the requirements will be returned. */
1220 int outputs() const;
1221
1222 /** Get the name of the extern function called for an extern
1223 * definition. */
1224 const std::string &extern_function_name() const;
1225
1226 /** The dimensionality (number of arguments) of this function.
1227 * If the Func isn't yet defined, but was specified with required dimensionality,
1228 * the dimensionality specified in the requirements will be returned. */
1229 int dimensions() const;
1230
1231 /** Construct either the left-hand-side of a definition, or a call
1232 * to a functions that happens to only contain vars as
1233 * arguments. If the function has already been defined, and fewer
1234 * arguments are given than the function has dimensions, then
1235 * enough implicit vars are added to the end of the argument list
1236 * to make up the difference (see \ref Var::implicit) */
1237 // @{
1238 FuncRef operator()(std::vector<Var>) const;
1239
1240 template<typename... Args>
1242 operator()(Args &&...args) const {
1243 std::vector<Var> collected_args{std::forward<Args>(args)...};
1244 return this->operator()(collected_args);
1245 }
1246 // @}
1247
1248 /** Either calls to the function, or the left-hand-side of
1249 * an update definition (see \ref RDom). If the function has
1250 * already been defined, and fewer arguments are given than the
1251 * function has dimensions, then enough implicit vars are added to
1252 * the end of the argument list to make up the difference. (see
1253 * \ref Var::implicit)*/
1254 // @{
1255 FuncRef operator()(std::vector<Expr>) const;
1256
1257 template<typename... Args>
1259 operator()(const Expr &x, Args &&...args) const {
1260 std::vector<Expr> collected_args{x, std::forward<Args>(args)...};
1261 return (*this)(collected_args);
1262 }
1263 // @}
1264
1265 /** Creates and returns a new identity Func that wraps this Func. During
1266 * compilation, Halide replaces all calls to this Func done by 'f'
1267 * with calls to the wrapper. If this Func is already wrapped for
1268 * use in 'f', will return the existing wrapper.
1269 *
1270 * For example, g.in(f) would rewrite a pipeline like this:
1271 \code
1272 g(x, y) = ...
1273 f(x, y) = ... g(x, y) ...
1274 \endcode
1275 * into a pipeline like this:
1276 \code
1277 g(x, y) = ...
1278 g_wrap(x, y) = g(x, y)
1279 f(x, y) = ... g_wrap(x, y)
1280 \endcode
1281 *
1282 * This has a variety of uses. You can use it to schedule this
1283 * Func differently in the different places it is used:
1284 \code
1285 g(x, y) = ...
1286 f1(x, y) = ... g(x, y) ...
1287 f2(x, y) = ... g(x, y) ...
1288 g.in(f1).compute_at(f1, y).vectorize(x, 8);
1289 g.in(f2).compute_at(f2, x).unroll(x);
1290 \endcode
1291 *
1292 * You can also use it to stage loads from this Func via some
1293 * intermediate buffer (perhaps on the stack as in
1294 * test/performance/block_transpose.cpp, or in shared GPU memory
1295 * as in test/performance/wrap.cpp). In this we compute the
1296 * wrapper at tiles of the consuming Funcs like so:
1297 \code
1298 g.compute_root()...
1299 g.in(f).compute_at(f, tiles)...
1300 \endcode
1301 *
1302 * Func::in() can also be used to compute pieces of a Func into a
1303 * smaller scratch buffer (perhaps on the GPU) and then copy them
1304 * into a larger output buffer one tile at a time. See
1305 * apps/interpolate/interpolate.cpp for an example of this. In
1306 * this case we compute the Func at tiles of its own wrapper:
1307 \code
1308 f.in(g).compute_root().gpu_tile(...)...
1309 f.compute_at(f.in(g), tiles)...
1310 \endcode
1311 *
1312 * A similar use of Func::in() wrapping Funcs with multiple update
1313 * stages in a pure wrapper. The following code:
1314 \code
1315 f(x, y) = x + y;
1316 f(x, y) += 5;
1317 g(x, y) = f(x, y);
1318 f.compute_root();
1319 \endcode
1320 *
1321 * Is equivalent to:
1322 \code
1323 for y:
1324 for x:
1325 f(x, y) = x + y;
1326 for y:
1327 for x:
1328 f(x, y) += 5
1329 for y:
1330 for x:
1331 g(x, y) = f(x, y)
1332 \endcode
1333 * using Func::in(), we can write:
1334 \code
1335 f(x, y) = x + y;
1336 f(x, y) += 5;
1337 g(x, y) = f(x, y);
1338 f.in(g).compute_root();
1339 \endcode
1340 * which instead produces:
1341 \code
1342 for y:
1343 for x:
1344 f(x, y) = x + y;
1345 f(x, y) += 5
1346 f_wrap(x, y) = f(x, y)
1347 for y:
1348 for x:
1349 g(x, y) = f_wrap(x, y)
1350 \endcode
1351 */
1352 Func in(const Func &f);
1353
1354 /** Create and return an identity wrapper shared by all the Funcs in
1355 * 'fs'. If any of the Funcs in 'fs' already have a custom wrapper,
1356 * this will throw an error. */
1357 Func in(const std::vector<Func> &fs);
1358
1359 /** Create and return a global identity wrapper, which wraps all calls to
1360 * this Func by any other Func. If a global wrapper already exists,
1361 * returns it. The global identity wrapper is only used by callers for
1362 * which no custom wrapper has been specified.
1363 */
1365
1366 /** Similar to \ref Func::in; however, instead of replacing the call to
1367 * this Func with an identity Func that refers to it, this replaces the
1368 * call with a clone of this Func.
1369 *
1370 * For example, f.clone_in(g) would rewrite a pipeline like this:
1371 \code
1372 f(x, y) = x + y;
1373 g(x, y) = f(x, y) + 2;
1374 h(x, y) = f(x, y) - 3;
1375 \endcode
1376 * into a pipeline like this:
1377 \code
1378 f(x, y) = x + y;
1379 f_clone(x, y) = x + y;
1380 g(x, y) = f_clone(x, y) + 2;
1381 h(x, y) = f(x, y) - 3;
1382 \endcode
1383 *
1384 */
1385 //@{
1386 Func clone_in(const Func &f);
1387 Func clone_in(const std::vector<Func> &fs);
1388 //@}
1389
1390 /** Declare that this function should be implemented by a call to
1391 * halide_buffer_copy with the given target device API. Asserts
1392 * that the Func has a pure definition which is a simple call to a
1393 * single input, and no update definitions. The wrapper Funcs
1394 * returned by in() are suitable candidates. Consumes all pure
1395 * variables, and rewrites the Func to have an extern definition
1396 * that calls halide_buffer_copy. */
1398
1399 /** Declare that this function should be implemented by a call to
1400 * halide_buffer_copy with a NULL target device API. Equivalent to
1401 * copy_to_device(DeviceAPI::Host). Asserts that the Func has a
1402 * pure definition which is a simple call to a single input, and
1403 * no update definitions. The wrapper Funcs returned by in() are
1404 * suitable candidates. Consumes all pure variables, and rewrites
1405 * the Func to have an extern definition that calls
1406 * halide_buffer_copy.
1407 *
1408 * Note that if the source Func is already valid in host memory,
1409 * this compiles to code that does the minimum number of calls to
1410 * memcpy.
1411 */
1413
1414 /** Split a dimension into inner and outer subdimensions with the
1415 * given names, where the inner dimension iterates from 0 to
1416 * factor-1. The inner and outer subdimensions can then be dealt
1417 * with using the other scheduling calls. It's ok to reuse the old
1418 * variable name as either the inner or outer variable. The final
1419 * argument specifies how the tail should be handled if the split
1420 * factor does not provably divide the extent. */
1421 Func &split(const VarOrRVar &old, const VarOrRVar &outer, const VarOrRVar &inner, const Expr &factor, TailStrategy tail = TailStrategy::Auto);
1422
1423 /** Join two dimensions into a single fused dimension. The fused dimension
1424 * covers the product of the extents of the inner and outer dimensions
1425 * given. The loop type (e.g. parallel, vectorized) of the resulting fused
1426 * dimension is inherited from the first argument. */
1427 Func &fuse(const VarOrRVar &inner, const VarOrRVar &outer, const VarOrRVar &fused);
1428
1429 /** Mark a dimension to be traversed serially. This is the default. */
1430 Func &serial(const VarOrRVar &var);
1431
1432 /** Mark a dimension to be traversed in parallel */
1434
1435 /** Split a dimension by the given task_size, and the parallelize the
1436 * outer dimension. This creates parallel tasks that have size
1437 * task_size. After this call, var refers to the outer dimension of
1438 * the split. The inner dimension has a new anonymous name. If you
1439 * wish to mutate it, or schedule with respect to it, do the split
1440 * manually. */
1442
1443 /** Mark a dimension to be computed all-at-once as a single
1444 * vector. The dimension should have constant extent -
1445 * e.g. because it is the inner dimension following a split by a
1446 * constant factor. For most uses of vectorize you want the two
1447 * argument form. The variable to be vectorized should be the
1448 * innermost one. */
1450
1451 /** Mark a dimension to be completely unrolled. The dimension
1452 * should have constant extent - e.g. because it is the inner
1453 * dimension following a split by a constant factor. For most uses
1454 * of unroll you want the two-argument form. */
1455 Func &unroll(const VarOrRVar &var);
1456
1457 /** Split a dimension by the given factor, then vectorize the
1458 * inner dimension. This is how you vectorize a loop of unknown
1459 * size. The variable to be vectorized should be the innermost
1460 * one. After this call, var refers to the outer dimension of the
1461 * split. 'factor' must be an integer. */
1462 Func &vectorize(const VarOrRVar &var, const Expr &factor, TailStrategy tail = TailStrategy::Auto);
1463
1464 /** Split a dimension by the given factor, then unroll the inner
1465 * dimension. This is how you unroll a loop of unknown size by
1466 * some constant factor. After this call, var refers to the outer
1467 * dimension of the split. 'factor' must be an integer. */
1468 Func &unroll(const VarOrRVar &var, const Expr &factor, TailStrategy tail = TailStrategy::Auto);
1469
1470 /** Set the loop partition policy. Loop partitioning can be useful to
1471 * optimize boundary conditions (such as clamp_edge). Loop partitioning
1472 * splits a for loop into three for loops: a prologue, a steady-state,
1473 * and an epilogue.
1474 * The default policy is Auto. */
1475 Func &partition(const VarOrRVar &var, Partition partition_policy);
1476
1477 /** Set the loop partition policy to Never for a vector of Vars and
1478 * RVars. */
1479 Func &never_partition(const std::vector<VarOrRVar> &vars);
1480
1481 /** Set the loop partition policy to Never for some number of Vars and RVars. */
1482 template<typename... Args>
1485 std::vector<VarOrRVar> collected_args{x, std::forward<Args>(args)...};
1487 }
1488
1489 /** Set the loop partition policy to Never for all Vars and RVar of the
1490 * initial definition of the Func. It must be called separately on any
1491 * update definitions. */
1493
1494 /** Set the loop partition policy to Always for a vector of Vars and
1495 * RVars. */
1496 Func &always_partition(const std::vector<VarOrRVar> &vars);
1497
1498 /** Set the loop partition policy to Always for some number of Vars and RVars. */
1499 template<typename... Args>
1502 std::vector<VarOrRVar> collected_args{x, std::forward<Args>(args)...};
1504 }
1505
1506 /** Set the loop partition policy to Always for all Vars and RVar of the
1507 * initial definition of the Func. It must be called separately on any
1508 * update definitions. */
1510
1511 /** Statically declare that the range over which a function should
1512 * be evaluated is given by the second and third arguments. This
1513 * can let Halide perform some optimizations. E.g. if you know
1514 * there are going to be 4 color channels, you can completely
1515 * vectorize the color channel dimension without the overhead of
1516 * splitting it up. If bounds inference decides that it requires
1517 * more of this function than the bounds you have stated, a
1518 * runtime error will occur when you try to run your pipeline. */
1519 Func &bound(const Var &var, Expr min, Expr extent);
1520
1521 /** Statically declare the range over which the function will be
1522 * evaluated in the general case. This provides a basis for the auto
1523 * scheduler to make trade-offs and scheduling decisions. The auto
1524 * generated schedules might break when the sizes of the dimensions are
1525 * very different from the estimates specified. These estimates are used
1526 * only by the auto scheduler if the function is a pipeline output. */
1527 Func &set_estimate(const Var &var, const Expr &min, const Expr &extent);
1528
1529 /** Set (min, extent) estimates for all dimensions in the Func
1530 * at once; this is equivalent to calling `set_estimate(args()[n], min, extent)`
1531 * repeatedly, but slightly terser. The size of the estimates vector
1532 * must match the dimensionality of the Func. */
1533 Func &set_estimates(const Region &estimates);
1534
1535 /** Expand the region computed so that the min coordinates is
1536 * congruent to 'remainder' modulo 'modulus', and the extent is a
1537 * multiple of 'modulus'. For example, f.align_bounds(x, 2) forces
1538 * the min and extent realized to be even, and calling
1539 * f.align_bounds(x, 2, 1) forces the min to be odd and the extent
1540 * to be even. The region computed always contains the region that
1541 * would have been computed without this directive, so no
1542 * assertions are injected.
1543 */
1544 Func &align_bounds(const Var &var, Expr modulus, Expr remainder = 0);
1545
1546 /** Expand the region computed so that the extent is a
1547 * multiple of 'modulus'. For example, f.align_extent(x, 2) forces
1548 * the extent realized to be even. The region computed always contains the
1549 * region that would have been computed without this directive, so no
1550 * assertions are injected. (This is essentially equivalent to align_bounds(),
1551 * but always leaving the min untouched.)
1552 */
1553 Func &align_extent(const Var &var, Expr modulus);
1554
1555 /** Bound the extent of a Func's realization, but not its
1556 * min. This means the dimension can be unrolled or vectorized
1557 * even when its min is not fixed (for example because it is
1558 * compute_at tiles of another Func). This can also be useful for
1559 * forcing a function's allocation to be a fixed size, which often
1560 * means it can go on the stack. */
1561 Func &bound_extent(const Var &var, Expr extent);
1562
1563 /** Split two dimensions at once by the given factors, and then
1564 * reorder the resulting dimensions to be xi, yi, xo, yo from
1565 * innermost outwards. This gives a tiled traversal. */
1566 Func &tile(const VarOrRVar &x, const VarOrRVar &y,
1567 const VarOrRVar &xo, const VarOrRVar &yo,
1568 const VarOrRVar &xi, const VarOrRVar &yi,
1569 const Expr &xfactor, const Expr &yfactor,
1571
1572 /** A shorter form of tile, which reuses the old variable names as
1573 * the new outer dimensions */
1574 Func &tile(const VarOrRVar &x, const VarOrRVar &y,
1575 const VarOrRVar &xi, const VarOrRVar &yi,
1576 const Expr &xfactor, const Expr &yfactor,
1578
1579 /** A more general form of tile, which defines tiles of any dimensionality. */
1580 Func &tile(const std::vector<VarOrRVar> &previous,
1581 const std::vector<VarOrRVar> &outers,
1582 const std::vector<VarOrRVar> &inners,
1583 const std::vector<Expr> &factors,
1584 const std::vector<TailStrategy> &tails);
1585
1586 /** The generalized tile, with a single tail strategy to apply to all vars. */
1587 Func &tile(const std::vector<VarOrRVar> &previous,
1588 const std::vector<VarOrRVar> &outers,
1589 const std::vector<VarOrRVar> &inners,
1590 const std::vector<Expr> &factors,
1592
1593 /** Generalized tiling, reusing the previous names as the outer names. */
1594 Func &tile(const std::vector<VarOrRVar> &previous,
1595 const std::vector<VarOrRVar> &inners,
1596 const std::vector<Expr> &factors,
1598
1599 /** Reorder variables to have the given nesting order, from
1600 * innermost out */
1601 Func &reorder(const std::vector<VarOrRVar> &vars);
1602
1603 template<typename... Args>
1605 reorder(const VarOrRVar &x, const VarOrRVar &y, Args &&...args) {
1606 std::vector<VarOrRVar> collected_args{x, y, std::forward<Args>(args)...};
1607 return reorder(collected_args);
1608 }
1609
1610 /** Rename a dimension. Equivalent to split with a inner size of one. */
1612
1613 /** Specify that race conditions are permitted for this Func,
1614 * which enables parallelizing over RVars even when Halide cannot
1615 * prove that it is safe to do so. Use this with great caution,
1616 * and only if you can prove to yourself that this is safe, as it
1617 * may result in a non-deterministic routine that returns
1618 * different values at different times or on different machines. */
1620
1621 /** Issue atomic updates for this Func. This allows parallelization
1622 * on associative RVars. The function throws a compile error when
1623 * Halide fails to prove associativity. Use override_associativity_test
1624 * to disable the associativity test if you believe the function is
1625 * associative or the order of reduction variable execution does not
1626 * matter.
1627 * Halide compiles this into hardware atomic operations whenever possible,
1628 * and falls back to a mutex lock per storage element if it is impossible
1629 * to atomically update.
1630 * There are three possible outcomes of the compiled code:
1631 * atomic add, compare-and-swap loop, and mutex lock.
1632 * For example:
1633 *
1634 * hist(x) = 0;
1635 * hist(im(r)) += 1;
1636 * hist.compute_root();
1637 * hist.update().atomic().parallel();
1638 *
1639 * will be compiled to atomic add operations.
1640 *
1641 * hist(x) = 0;
1642 * hist(im(r)) = min(hist(im(r)) + 1, 100);
1643 * hist.compute_root();
1644 * hist.update().atomic().parallel();
1645 *
1646 * will be compiled to compare-and-swap loops.
1647 *
1648 * arg_max() = {0, im(0)};
1649 * Expr old_index = arg_max()[0];
1650 * Expr old_max = arg_max()[1];
1651 * Expr new_index = select(old_max < im(r), r, old_index);
1652 * Expr new_max = max(im(r), old_max);
1653 * arg_max() = {new_index, new_max};
1654 * arg_max.compute_root();
1655 * arg_max.update().atomic().parallel();
1656 *
1657 * will be compiled to updates guarded by a mutex lock,
1658 * since it is impossible to atomically update two different locations.
1659 *
1660 * Currently the atomic operation is supported by x86, CUDA, and OpenCL backends.
1661 * Compiling to other backends results in a compile error.
1662 * If an operation is compiled into a mutex lock, and is vectorized or is
1663 * compiled to CUDA or OpenCL, it also results in a compile error,
1664 * since per-element mutex lock on vectorized operation leads to a
1665 * deadlock.
1666 * Vectorization of predicated RVars (through rdom.where()) on CPU
1667 * is also unsupported yet (see https://github.com/halide/Halide/issues/4298).
1668 * 8-bit and 16-bit atomics on GPU are also not supported. */
1670
1671 /** Specialize a Func. This creates a special-case version of the
1672 * Func where the given condition is true. The most effective
1673 * conditions are those of the form param == value, and boolean
1674 * Params. Consider a simple example:
1675 \code
1676 f(x) = x + select(cond, 0, 1);
1677 f.compute_root();
1678 \endcode
1679 * This is equivalent to:
1680 \code
1681 for (int x = 0; x < width; x++) {
1682 f[x] = x + (cond ? 0 : 1);
1683 }
1684 \endcode
1685 * Adding the scheduling directive:
1686 \code
1687 f.specialize(cond)
1688 \endcode
1689 * makes it equivalent to:
1690 \code
1691 if (cond) {
1692 for (int x = 0; x < width; x++) {
1693 f[x] = x;
1694 }
1695 } else {
1696 for (int x = 0; x < width; x++) {
1697 f[x] = x + 1;
1698 }
1699 }
1700 \endcode
1701 * Note that the inner loops have been simplified. In the first
1702 * path Halide knows that cond is true, and in the second path
1703 * Halide knows that it is false.
1704 *
1705 * The specialized version gets its own schedule, which inherits
1706 * every directive made about the parent Func's schedule so far
1707 * except for its specializations. This method returns a handle to
1708 * the new schedule. If you wish to retrieve the specialized
1709 * sub-schedule again later, you can call this method with the
1710 * same condition. Consider the following example of scheduling
1711 * the specialized version:
1712 *
1713 \code
1714 f(x) = x;
1715 f.compute_root();
1716 f.specialize(width > 1).unroll(x, 2);
1717 \endcode
1718 * Assuming for simplicity that width is even, this is equivalent to:
1719 \code
1720 if (width > 1) {
1721 for (int x = 0; x < width/2; x++) {
1722 f[2*x] = 2*x;
1723 f[2*x + 1] = 2*x + 1;
1724 }
1725 } else {
1726 for (int x = 0; x < width/2; x++) {
1727 f[x] = x;
1728 }
1729 }
1730 \endcode
1731 * For this case, it may be better to schedule the un-specialized
1732 * case instead:
1733 \code
1734 f(x) = x;
1735 f.compute_root();
1736 f.specialize(width == 1); // Creates a copy of the schedule so far.
1737 f.unroll(x, 2); // Only applies to the unspecialized case.
1738 \endcode
1739 * This is equivalent to:
1740 \code
1741 if (width == 1) {
1742 f[0] = 0;
1743 } else {
1744 for (int x = 0; x < width/2; x++) {
1745 f[2*x] = 2*x;
1746 f[2*x + 1] = 2*x + 1;
1747 }
1748 }
1749 \endcode
1750 * This can be a good way to write a pipeline that splits,
1751 * vectorizes, or tiles, but can still handle small inputs.
1752 *
1753 * If a Func has several specializations, the first matching one
1754 * will be used, so the order in which you define specializations
1755 * is significant. For example:
1756 *
1757 \code
1758 f(x) = x + select(cond1, a, b) - select(cond2, c, d);
1759 f.specialize(cond1);
1760 f.specialize(cond2);
1761 \endcode
1762 * is equivalent to:
1763 \code
1764 if (cond1) {
1765 for (int x = 0; x < width; x++) {
1766 f[x] = x + a - (cond2 ? c : d);
1767 }
1768 } else if (cond2) {
1769 for (int x = 0; x < width; x++) {
1770 f[x] = x + b - c;
1771 }
1772 } else {
1773 for (int x = 0; x < width; x++) {
1774 f[x] = x + b - d;
1775 }
1776 }
1777 \endcode
1778 *
1779 * Specializations may in turn be specialized, which creates a
1780 * nested if statement in the generated code.
1781 *
1782 \code
1783 f(x) = x + select(cond1, a, b) - select(cond2, c, d);
1784 f.specialize(cond1).specialize(cond2);
1785 \endcode
1786 * This is equivalent to:
1787 \code
1788 if (cond1) {
1789 if (cond2) {
1790 for (int x = 0; x < width; x++) {
1791 f[x] = x + a - c;
1792 }
1793 } else {
1794 for (int x = 0; x < width; x++) {
1795 f[x] = x + a - d;
1796 }
1797 }
1798 } else {
1799 for (int x = 0; x < width; x++) {
1800 f[x] = x + b - (cond2 ? c : d);
1801 }
1802 }
1803 \endcode
1804 * To create a 4-way if statement that simplifies away all of the
1805 * ternary operators above, you could say:
1806 \code
1807 f.specialize(cond1).specialize(cond2);
1808 f.specialize(cond2);
1809 \endcode
1810 * or
1811 \code
1812 f.specialize(cond1 && cond2);
1813 f.specialize(cond1);
1814 f.specialize(cond2);
1815 \endcode
1816 *
1817 * Any prior Func which is compute_at some variable of this Func
1818 * gets separately included in all paths of the generated if
1819 * statement. The Var in the compute_at call to must exist in all
1820 * paths, but it may have been generated via a different path of
1821 * splits, fuses, and renames. This can be used somewhat
1822 * creatively. Consider the following code:
1823 \code
1824 g(x, y) = 8*x;
1825 f(x, y) = g(x, y) + 1;
1826 f.compute_root().specialize(cond);
1827 Var g_loop;
1828 f.specialize(cond).rename(y, g_loop);
1829 f.rename(x, g_loop);
1830 g.compute_at(f, g_loop);
1831 \endcode
1832 * When cond is true, this is equivalent to g.compute_at(f,y).
1833 * When it is false, this is equivalent to g.compute_at(f,x).
1834 */
1835 Stage specialize(const Expr &condition);
1836
1837 /** Add a specialization to a Func that always terminates execution
1838 * with a call to halide_error(). By itself, this is of limited use,
1839 * but can be useful to terminate chains of specialize() calls where
1840 * no "default" case is expected (thus avoiding unnecessary code generation).
1841 *
1842 * For instance, say we want to optimize a pipeline to process images
1843 * in planar and interleaved format; we might typically do something like:
1844 \code
1845 ImageParam im(UInt(8), 3);
1846 Func f = do_something_with(im);
1847 f.specialize(im.dim(0).stride() == 1).vectorize(x, 8); // planar
1848 f.specialize(im.dim(2).stride() == 1).reorder(c, x, y).vectorize(c); // interleaved
1849 \endcode
1850 * This code will vectorize along rows for the planar case, and across pixel
1851 * components for the interleaved case... but there is an implicit "else"
1852 * for the unhandled cases, which generates unoptimized code. If we never
1853 * anticipate passing any other sort of images to this, we code streamline
1854 * our code by adding specialize_fail():
1855 \code
1856 ImageParam im(UInt(8), 3);
1857 Func f = do_something(im);
1858 f.specialize(im.dim(0).stride() == 1).vectorize(x, 8); // planar
1859 f.specialize(im.dim(2).stride() == 1).reorder(c, x, y).vectorize(c); // interleaved
1860 f.specialize_fail("Unhandled image format");
1861 \endcode
1862 * Conceptually, this produces codes like:
1863 \code
1864 if (im.dim(0).stride() == 1) {
1865 do_something_planar();
1866 } else if (im.dim(2).stride() == 1) {
1867 do_something_interleaved();
1868 } else {
1869 halide_error("Unhandled image format");
1870 }
1871 \endcode
1872 *
1873 * Note that calling specialize_fail() terminates the specialization chain
1874 * for a given Func; you cannot create new specializations for the Func
1875 * afterwards (though you can retrieve handles to previous specializations).
1876 */
1877 void specialize_fail(const std::string &message);
1878
1879 /** Tell Halide that the following dimensions correspond to GPU
1880 * thread indices. This is useful if you compute a producer
1881 * function within the block indices of a consumer function, and
1882 * want to control how that function's dimensions map to GPU
1883 * threads. If the selected target is not an appropriate GPU, this
1884 * just marks those dimensions as parallel. */
1885 // @{
1889 // @}
1890
1891 /** The given dimension corresponds to the lanes in a GPU
1892 * warp. GPU warp lanes are distinguished from GPU threads by the
1893 * fact that all warp lanes run together in lockstep, which
1894 * permits lightweight communication of data from one lane to
1895 * another. */
1897
1898 /** Tell Halide to run this stage using a single gpu thread and
1899 * block. This is not an efficient use of your GPU, but it can be
1900 * useful to avoid copy-back for intermediate update stages that
1901 * touch a very small part of your Func. */
1903
1904 /** Tell Halide that the following dimensions correspond to GPU
1905 * block indices. This is useful for scheduling stages that will
1906 * run serially within each GPU block. If the selected target is
1907 * not ptx, this just marks those dimensions as parallel. */
1908 // @{
1912 // @}
1913
1914 /** Tell Halide that the following dimensions correspond to GPU
1915 * block indices and thread indices. If the selected target is not
1916 * ptx, these just mark the given dimensions as parallel. The
1917 * dimensions are consumed by this call, so do all other
1918 * unrolling, reordering, etc first. */
1919 // @{
1925 // @}
1926
1927 /** Short-hand for tiling a domain and mapping the tile indices
1928 * to GPU block indices and the coordinates within each tile to
1929 * GPU thread indices. Consumes the variables given, so do all
1930 * other scheduling first. */
1931 // @{
1932 Func &gpu_tile(const VarOrRVar &x, const VarOrRVar &bx, const VarOrRVar &tx, const Expr &x_size,
1934 DeviceAPI device_api = DeviceAPI::Default_GPU);
1935
1936 Func &gpu_tile(const VarOrRVar &x, const VarOrRVar &tx, const Expr &x_size,
1938 DeviceAPI device_api = DeviceAPI::Default_GPU);
1939 Func &gpu_tile(const VarOrRVar &x, const VarOrRVar &y,
1940 const VarOrRVar &bx, const VarOrRVar &by,
1941 const VarOrRVar &tx, const VarOrRVar &ty,
1942 const Expr &x_size, const Expr &y_size,
1944 DeviceAPI device_api = DeviceAPI::Default_GPU);
1945
1946 Func &gpu_tile(const VarOrRVar &x, const VarOrRVar &y,
1947 const VarOrRVar &tx, const VarOrRVar &ty,
1948 const Expr &x_size, const Expr &y_size,
1950 DeviceAPI device_api = DeviceAPI::Default_GPU);
1951
1952 Func &gpu_tile(const VarOrRVar &x, const VarOrRVar &y, const VarOrRVar &z,
1953 const VarOrRVar &bx, const VarOrRVar &by, const VarOrRVar &bz,
1954 const VarOrRVar &tx, const VarOrRVar &ty, const VarOrRVar &tz,
1955 const Expr &x_size, const Expr &y_size, const Expr &z_size,
1957 DeviceAPI device_api = DeviceAPI::Default_GPU);
1958 Func &gpu_tile(const VarOrRVar &x, const VarOrRVar &y, const VarOrRVar &z,
1959 const VarOrRVar &tx, const VarOrRVar &ty, const VarOrRVar &tz,
1960 const Expr &x_size, const Expr &y_size, const Expr &z_size,
1962 DeviceAPI device_api = DeviceAPI::Default_GPU);
1963 // @}
1964
1965 /** Schedule for execution on Hexagon. When a loop is marked with
1966 * Hexagon, that loop is executed on a Hexagon DSP. */
1968
1969 /** Prefetch data written to or read from a Func or an ImageParam by a
1970 * subsequent loop iteration, at an optionally specified iteration offset. You may specify
1971 * specification of different vars for the location of the prefetch() instruction
1972 * vs. the location that is being prefetched:
1973 *
1974 * - the first var specified, 'at', indicates the loop in which the prefetch will be placed
1975 * - the second var specified, 'from', determines the var used to find the bounds to prefetch
1976 * (in conjunction with 'offset')
1977 *
1978 * If 'at' and 'from' are distinct vars, then 'from' must be at a nesting level outside 'at.'
1979 * Note that the value for 'offset' applies only to 'from', not 'at'.
1980 *
1981 * The final argument specifies how prefetch of region outside bounds
1982 * should be handled.
1983 *
1984 * For example, consider this pipeline:
1985 \code
1986 Func f, g;
1987 Var x, y, z;
1988 f(x, y) = x + y;
1989 g(x, y) = 2 * f(x, y);
1990 h(x, y) = 3 * f(x, y);
1991 \endcode
1992 *
1993 * The following schedule:
1994 \code
1995 f.compute_root();
1996 g.prefetch(f, x, x, 2, PrefetchBoundStrategy::NonFaulting);
1997 h.prefetch(f, x, y, 2, PrefetchBoundStrategy::NonFaulting);
1998 \endcode
1999 *
2000 * will inject prefetch call at the innermost loop of 'g' and 'h' and generate
2001 * the following loop nest:
2002 \code
2003 for y = ...
2004 for x = ...
2005 f(x, y) = x + y
2006 for y = ..
2007 for x = ...
2008 prefetch(&f[x + 2, y], 1, 16);
2009 g(x, y) = 2 * f(x, y)
2010 for y = ..
2011 for x = ...
2012 prefetch(&f[x, y + 2], 1, 16);
2013 h(x, y) = 3 * f(x, y)
2014 \endcode
2015 *
2016 * Note that the 'from' nesting level need not be adjacent to 'at':
2017 \code
2018 Func f, g;
2019 Var x, y, z, w;
2020 f(x, y, z, w) = x + y + z + w;
2021 g(x, y, z, w) = 2 * f(x, y, z, w);
2022 \endcode
2023 *
2024 * The following schedule:
2025 \code
2026 f.compute_root();
2027 g.prefetch(f, y, w, 2, PrefetchBoundStrategy::NonFaulting);
2028 \endcode
2029 *
2030 * will produce code that prefetches a tile of data:
2031 \code
2032 for w = ...
2033 for z = ...
2034 for y = ...
2035 for x = ...
2036 f(x, y, z, w) = x + y + z + w
2037 for w = ...
2038 for z = ...
2039 for y = ...
2040 for x0 = ...
2041 prefetch(&f[x0, y, z, w + 2], 1, 16);
2042 for x = ...
2043 g(x, y, z, w) = 2 * f(x, y, z, w)
2044 \endcode
2045 *
2046 * Note that calling prefetch() with the same var for both 'at' and 'from'
2047 * is equivalent to calling prefetch() with that var.
2048 */
2049 // @{
2050 Func &prefetch(const Func &f, const VarOrRVar &at, const VarOrRVar &from, Expr offset = 1,
2052 Func &prefetch(const Parameter &param, const VarOrRVar &at, const VarOrRVar &from, Expr offset = 1,
2054 template<typename T>
2055 Func &prefetch(const T &image, const VarOrRVar &at, const VarOrRVar &from, Expr offset = 1,
2057 return prefetch(image.parameter(), at, from, std::move(offset), strategy);
2058 }
2059 // @}
2060
2061 /** Specify how the storage for the function is laid out. These
2062 * calls let you specify the nesting order of the dimensions. For
2063 * example, foo.reorder_storage(y, x) tells Halide to use
2064 * column-major storage for any realizations of foo, without
2065 * changing how you refer to foo in the code. You may want to do
2066 * this if you intend to vectorize across y. When representing
2067 * color images, foo.reorder_storage(c, x, y) specifies packed
2068 * storage (red, green, and blue values adjacent in memory), and
2069 * foo.reorder_storage(x, y, c) specifies planar storage (entire
2070 * red, green, and blue images one after the other in memory).
2071 *
2072 * If you leave out some dimensions, those remain in the same
2073 * positions in the nesting order while the specified variables
2074 * are reordered around them. */
2075 // @{
2076 Func &reorder_storage(const std::vector<Var> &dims);
2077
2078 Func &reorder_storage(const Var &x, const Var &y);
2079 template<typename... Args>
2081 reorder_storage(const Var &x, const Var &y, Args &&...args) {
2082 std::vector<Var> collected_args{x, y, std::forward<Args>(args)...};
2083 return reorder_storage(collected_args);
2084 }
2085 // @}
2086
2087 /** Pad the storage extent of a particular dimension of
2088 * realizations of this function up to be a multiple of the
2089 * specified alignment. This guarantees that the strides for the
2090 * dimensions stored outside of dim will be multiples of the
2091 * specified alignment, where the strides and alignment are
2092 * measured in numbers of elements.
2093 *
2094 * For example, to guarantee that a function foo(x, y, c)
2095 * representing an image has scanlines starting on offsets
2096 * aligned to multiples of 16, use foo.align_storage(x, 16). */
2097 Func &align_storage(const Var &dim, const Expr &alignment);
2098
2099 /** Store realizations of this function in a circular buffer of a
2100 * given extent. This is more efficient when the extent of the
2101 * circular buffer is a power of 2. If the fold factor is too
2102 * small, or the dimension is not accessed monotonically, the
2103 * pipeline will generate an error at runtime.
2104 *
2105 * The fold_forward option indicates that the new values of the
2106 * producer are accessed by the consumer in a monotonically
2107 * increasing order. Folding storage of producers is also
2108 * supported if the new values are accessed in a monotonically
2109 * decreasing order by setting fold_forward to false.
2110 *
2111 * For example, consider the pipeline:
2112 \code
2113 Func f, g;
2114 Var x, y;
2115 g(x, y) = x*y;
2116 f(x, y) = g(x, y) + g(x, y+1);
2117 \endcode
2118 *
2119 * If we schedule f like so:
2120 *
2121 \code
2122 g.compute_at(f, y).store_root().fold_storage(y, 2);
2123 \endcode
2124 *
2125 * Then g will be computed at each row of f and stored in a buffer
2126 * with an extent in y of 2, alternately storing each computed row
2127 * of g in row y=0 or y=1.
2128 */
2129 Func &fold_storage(const Var &dim, const Expr &extent, bool fold_forward = true);
2130
2131 /** Compute this function as needed for each unique value of the
2132 * given var for the given calling function f.
2133 *
2134 * For example, consider the simple pipeline:
2135 \code
2136 Func f, g;
2137 Var x, y;
2138 g(x, y) = x*y;
2139 f(x, y) = g(x, y) + g(x, y+1) + g(x+1, y) + g(x+1, y+1);
2140 \endcode
2141 *
2142 * If we schedule f like so:
2143 *
2144 \code
2145 g.compute_at(f, x);
2146 \endcode
2147 *
2148 * Then the C code equivalent to this pipeline will look like this
2149 *
2150 \code
2151
2152 int f[height][width];
2153 for (int y = 0; y < height; y++) {
2154 for (int x = 0; x < width; x++) {
2155 int g[2][2];
2156 g[0][0] = x*y;
2157 g[0][1] = (x+1)*y;
2158 g[1][0] = x*(y+1);
2159 g[1][1] = (x+1)*(y+1);
2160 f[y][x] = g[0][0] + g[1][0] + g[0][1] + g[1][1];
2161 }
2162 }
2163
2164 \endcode
2165 *
2166 * The allocation and computation of g is within f's loop over x,
2167 * and enough of g is computed to satisfy all that f will need for
2168 * that iteration. This has excellent locality - values of g are
2169 * used as soon as they are computed, but it does redundant
2170 * work. Each value of g ends up getting computed four times. If
2171 * we instead schedule f like so:
2172 *
2173 \code
2174 g.compute_at(f, y);
2175 \endcode
2176 *
2177 * The equivalent C code is:
2178 *
2179 \code
2180 int f[height][width];
2181 for (int y = 0; y < height; y++) {
2182 int g[2][width+1];
2183 for (int x = 0; x < width; x++) {
2184 g[0][x] = x*y;
2185 g[1][x] = x*(y+1);
2186 }
2187 for (int x = 0; x < width; x++) {
2188 f[y][x] = g[0][x] + g[1][x] + g[0][x+1] + g[1][x+1];
2189 }
2190 }
2191 \endcode
2192 *
2193 * The allocation and computation of g is within f's loop over y,
2194 * and enough of g is computed to satisfy all that f will need for
2195 * that iteration. This does less redundant work (each point in g
2196 * ends up being evaluated twice), but the locality is not quite
2197 * as good, and we have to allocate more temporary memory to store
2198 * g.
2199 */
2200 Func &compute_at(const Func &f, const Var &var);
2201
2202 /** Schedule a function to be computed within the iteration over
2203 * some dimension of an update domain. Produces equivalent code
2204 * to the version of compute_at that takes a Var. */
2205 Func &compute_at(const Func &f, const RVar &var);
2206
2207 /** Schedule a function to be computed within the iteration over
2208 * a given LoopLevel. */
2210
2211 /** Schedule the iteration over the initial definition of this function
2212 * to be fused with another stage 's' from outermost loop to a
2213 * given LoopLevel. */
2214 // @{
2215 Func &compute_with(const Stage &s, const VarOrRVar &var, const std::vector<std::pair<VarOrRVar, LoopAlignStrategy>> &align);
2217 Func &compute_with(LoopLevel loop_level, const std::vector<std::pair<VarOrRVar, LoopAlignStrategy>> &align);
2219
2220 /** Compute all of this function once ahead of time. Reusing
2221 * the example in \ref Func::compute_at :
2222 *
2223 \code
2224 Func f, g;
2225 Var x, y;
2226 g(x, y) = x*y;
2227 f(x, y) = g(x, y) + g(x, y+1) + g(x+1, y) + g(x+1, y+1);
2228
2229 g.compute_root();
2230 \endcode
2231 *
2232 * is equivalent to
2233 *
2234 \code
2235 int f[height][width];
2236 int g[height+1][width+1];
2237 for (int y = 0; y < height+1; y++) {
2238 for (int x = 0; x < width+1; x++) {
2239 g[y][x] = x*y;
2240 }
2241 }
2242 for (int y = 0; y < height; y++) {
2243 for (int x = 0; x < width; x++) {
2244 f[y][x] = g[y][x] + g[y+1][x] + g[y][x+1] + g[y+1][x+1];
2245 }
2246 }
2247 \endcode
2248 *
2249 * g is computed once ahead of time, and enough is computed to
2250 * satisfy all uses of it. This does no redundant work (each point
2251 * in g is evaluated once), but has poor locality (values of g are
2252 * probably not still in cache when they are used by f), and
2253 * allocates lots of temporary memory to store g.
2254 */
2256
2257 /** Use the halide_memoization_cache_... interface to store a
2258 * computed version of this function across invocations of the
2259 * Func.
2260 *
2261 * If an eviction_key is provided, it must be constructed with
2262 * Expr of integer or handle type. The key Expr will be promoted
2263 * to a uint64_t and can be used with halide_memoization_cache_evict
2264 * to remove memoized entries using this eviction key from the
2265 * cache. Memoized computations that do not provide an eviction
2266 * key will never be evicted by this mechanism.
2267 */
2269
2270 /** Produce this Func asynchronously in a separate
2271 * thread. Consumers will be run by the task system when the
2272 * production is complete. If this Func's store level is different
2273 * to its compute level, consumers will be run concurrently,
2274 * blocking as necessary to prevent reading ahead of what the
2275 * producer has computed. If storage is folded, then the producer
2276 * will additionally not be permitted to run too far ahead of the
2277 * consumer, to avoid clobbering data that has not yet been
2278 * used.
2279 *
2280 * Take special care when combining this with custom thread pool
2281 * implementations, as avoiding deadlock with producer-consumer
2282 * parallelism requires a much more sophisticated parallel runtime
2283 * than with data parallelism alone. It is strongly recommended
2284 * you just use Halide's default thread pool, which guarantees no
2285 * deadlock and a bound on the number of threads launched.
2286 */
2288
2289 /** Expands the storage of the function by an extra dimension
2290 * to enable ring buffering. For this to be useful the storage
2291 * of the function has to be hoisted to an upper loop level using
2292 * \ref Func::hoist_storage. The index for the new ring buffer dimension
2293 * is calculated implicitly based on a linear combination of the all of
2294 * the loop variables between hoist_storage and compute_at/store_at
2295 * loop levels. Scheduling a function with ring_buffer increases the
2296 * amount of memory required for this function by an *extent* times.
2297 * ring_buffer is especially useful in combination with \ref Func::async,
2298 * but can be used without it.
2299 *
2300 * The extent is expected to be a positive integer.
2301 */
2303
2304 /** Bound the extent of a Func's storage, but not extent of its
2305 * compute. This can be useful for forcing a function's allocation
2306 * to be a fixed size, which often means it can go on the stack.
2307 * If bounds inference decides that it requires more storage for
2308 * this function than the allocation size you have stated, a runtime
2309 * error will occur when you try to run the pipeline. */
2310 Func &bound_storage(const Var &dim, const Expr &bound);
2311
2312 /** Allocate storage for this function within f's loop over
2313 * var. Scheduling storage is optional, and can be used to
2314 * separate the loop level at which storage occurs from the loop
2315 * level at which computation occurs to trade off between locality
2316 * and redundant work. This can open the door for two types of
2317 * optimization.
2318 *
2319 * Consider again the pipeline from \ref Func::compute_at :
2320 \code
2321 Func f, g;
2322 Var x, y;
2323 g(x, y) = x*y;
2324 f(x, y) = g(x, y) + g(x+1, y) + g(x, y+1) + g(x+1, y+1);
2325 \endcode
2326 *
2327 * If we schedule it like so:
2328 *
2329 \code
2330 g.compute_at(f, x).store_at(f, y);
2331 \endcode
2332 *
2333 * Then the computation of g takes place within the loop over x,
2334 * but the storage takes place within the loop over y:
2335 *
2336 \code
2337 int f[height][width];
2338 for (int y = 0; y < height; y++) {
2339 int g[2][width+1];
2340 for (int x = 0; x < width; x++) {
2341 g[0][x] = x*y;
2342 g[0][x+1] = (x+1)*y;
2343 g[1][x] = x*(y+1);
2344 g[1][x+1] = (x+1)*(y+1);
2345 f[y][x] = g[0][x] + g[1][x] + g[0][x+1] + g[1][x+1];
2346 }
2347 }
2348 \endcode
2349 *
2350 * Provided the for loop over x is serial, halide then
2351 * automatically performs the following sliding window
2352 * optimization:
2353 *
2354 \code
2355 int f[height][width];
2356 for (int y = 0; y < height; y++) {
2357 int g[2][width+1];
2358 for (int x = 0; x < width; x++) {
2359 if (x == 0) {
2360 g[0][x] = x*y;
2361 g[1][x] = x*(y+1);
2362 }
2363 g[0][x+1] = (x+1)*y;
2364 g[1][x+1] = (x+1)*(y+1);
2365 f[y][x] = g[0][x] + g[1][x] + g[0][x+1] + g[1][x+1];
2366 }
2367 }
2368 \endcode
2369 *
2370 * Two of the assignments to g only need to be done when x is
2371 * zero. The rest of the time, those sites have already been
2372 * filled in by a previous iteration. This version has the
2373 * locality of compute_at(f, x), but allocates more memory and
2374 * does much less redundant work.
2375 *
2376 * Halide then further optimizes this pipeline like so:
2377 *
2378 \code
2379 int f[height][width];
2380 for (int y = 0; y < height; y++) {
2381 int g[2][2];
2382 for (int x = 0; x < width; x++) {
2383 if (x == 0) {
2384 g[0][0] = x*y;
2385 g[1][0] = x*(y+1);
2386 }
2387 g[0][(x+1)%2] = (x+1)*y;
2388 g[1][(x+1)%2] = (x+1)*(y+1);
2389 f[y][x] = g[0][x%2] + g[1][x%2] + g[0][(x+1)%2] + g[1][(x+1)%2];
2390 }
2391 }
2392 \endcode
2393 *
2394 * Halide has detected that it's possible to use a circular buffer
2395 * to represent g, and has reduced all accesses to g modulo 2 in
2396 * the x dimension. This optimization only triggers if the for
2397 * loop over x is serial, and if halide can statically determine
2398 * some power of two large enough to cover the range needed. For
2399 * powers of two, the modulo operator compiles to more efficient
2400 * bit-masking. This optimization reduces memory usage, and also
2401 * improves locality by reusing recently-accessed memory instead
2402 * of pulling new memory into cache.
2403 *
2404 */
2405 Func &store_at(const Func &f, const Var &var);
2406
2407 /** Equivalent to the version of store_at that takes a Var, but
2408 * schedules storage within the loop over a dimension of a
2409 * reduction domain */
2410 Func &store_at(const Func &f, const RVar &var);
2411
2412 /** Equivalent to the version of store_at that takes a Var, but
2413 * schedules storage at a given LoopLevel. */
2415
2416 /** Equivalent to \ref Func::store_at, but schedules storage
2417 * outside the outermost loop. */
2419
2420 /** Hoist storage for this function within f's loop over
2421 * var. This is different from \ref Func::store_at, because hoist_storage
2422 * simply moves an actual allocation to a given loop level and
2423 * doesn't trigger any of the optimizations such as sliding window.
2424 * Hoisting storage is optional and can be used as an optimization
2425 * to avoid unnecessary allocations by moving it out from an inner
2426 * loop.
2427 *
2428 * Consider again the pipeline from \ref Func::compute_at :
2429 \code
2430 Func f, g;
2431 Var x, y;
2432 g(x, y) = x*y;
2433 f(x, y) = g(x, y) + g(x, y+1) + g(x+1, y) + g(x+1, y+1);
2434 \endcode
2435 *
2436 * If we schedule f like so:
2437 *
2438 \code
2439 g.compute_at(f, x);
2440 \endcode
2441 *
2442 * Then the C code equivalent to this pipeline will look like this
2443 *
2444 \code
2445
2446 int f[height][width];
2447 for (int y = 0; y < height; y++) {
2448 for (int x = 0; x < width; x++) {
2449 int g[2][2];
2450 g[0][0] = x*y;
2451 g[0][1] = (x+1)*y;
2452 g[1][0] = x*(y+1);
2453 g[1][1] = (x+1)*(y+1);
2454 f[y][x] = g[0][0] + g[1][0] + g[0][1] + g[1][1];
2455 }
2456 }
2457
2458 \endcode
2459 *
2460 * Note the allocation for g inside of the loop over variable x which
2461 * can happen for each iteration of the inner loop (in total height * width times).
2462 * In some cases allocation can be expensive, so it might be better to do it once
2463 * and reuse allocated memory across all iterations of the loop.
2464 *
2465 * This can be done by scheduling g like so:
2466 *
2467 \code
2468 g.compute_at(f, x).hoist_storage(f, Var::outermost());
2469 \endcode
2470 *
2471 * Then the C code equivalent to this pipeline will look like this
2472 *
2473 \code
2474
2475 int f[height][width];
2476 int g[2][2];
2477 for (int y = 0; y < height; y++) {
2478 for (int x = 0; x < width; x++) {
2479 g[0][0] = x*y;
2480 g[0][1] = (x+1)*y;
2481 g[1][0] = x*(y+1);
2482 g[1][1] = (x+1)*(y+1);
2483 f[y][x] = g[0][0] + g[1][0] + g[0][1] + g[1][1];
2484 }
2485 }
2486
2487 \endcode
2488 *
2489 * hoist_storage can be used together with \ref Func::store_at and
2490 * \ref Func::fold_storage (for example, to hoist the storage allocated
2491 * after sliding window optimization).
2492 *
2493 */
2494 Func &hoist_storage(const Func &f, const Var &var);
2495
2496 /** Equivalent to the version of hoist_storage that takes a Var, but
2497 * schedules storage within the loop over a dimension of a
2498 * reduction domain */
2499 Func &hoist_storage(const Func &f, const RVar &var);
2500
2501 /** Equivalent to the version of hoist_storage that takes a Var, but
2502 * schedules storage at a given LoopLevel. */
2504
2505 /** Equivalent to \ref Func::hoist_storage_root, but schedules storage
2506 * outside the outermost loop. */
2508
2509 /** Aggressively inline all uses of this function. This is the
2510 * default schedule, so you're unlikely to need to call this. For
2511 * a Func with an update definition, that means it gets computed
2512 * as close to the innermost loop as possible.
2513 *
2514 * Consider once more the pipeline from \ref Func::compute_at :
2515 *
2516 \code
2517 Func f, g;
2518 Var x, y;
2519 g(x, y) = x*y;
2520 f(x, y) = g(x, y) + g(x+1, y) + g(x, y+1) + g(x+1, y+1);
2521 \endcode
2522 *
2523 * Leaving g as inline, this compiles to code equivalent to the following C:
2524 *
2525 \code
2526 int f[height][width];
2527 for (int y = 0; y < height; y++) {
2528 for (int x = 0; x < width; x++) {
2529 f[y][x] = x*y + x*(y+1) + (x+1)*y + (x+1)*(y+1);
2530 }
2531 }
2532 \endcode
2533 */
2535
2536 /** Get a handle on an update step for the purposes of scheduling
2537 * it. */
2538 Stage update(int idx = 0);
2539
2540 /** Set the type of memory this Func should be stored in. Controls
2541 * whether allocations go on the stack or the heap on the CPU, and
2542 * in global vs shared vs local on the GPU. See the documentation
2543 * on MemoryType for more detail. */
2544 Func &store_in(MemoryType memory_type);
2545
2546 /** Trace all loads from this Func by emitting calls to
2547 * halide_trace. If the Func is inlined, this has no
2548 * effect. */
2550
2551 /** Trace all stores to the buffer backing this Func by emitting
2552 * calls to halide_trace. If the Func is inlined, this call
2553 * has no effect. */
2555
2556 /** Trace all realizations of this Func by emitting calls to
2557 * halide_trace. */
2559
2560 /** Add a string of arbitrary text that will be passed thru to trace
2561 * inspection code if the Func is realized in trace mode. (Funcs that are
2562 * inlined won't have their tags emitted.) Ignored entirely if
2563 * tracing is not enabled for the Func (or globally).
2564 */
2565 Func &add_trace_tag(const std::string &trace_tag);
2566
2567 /** Marks this function as a function that should not be profiled
2568 * when using the target feature Profile or ProfileByTimer.
2569 * This is useful when this function is does too little work at once
2570 * such that the overhead of setting the profiling token might
2571 * become significant, or that the measured time is not representative
2572 * due to modern processors (instruction level parallelism, out-of-order
2573 * execution). */
2575
2576 /** Get a handle on the internal halide function that this Func
2577 * represents. Useful if you want to do introspection on Halide
2578 * functions */
2580 return func;
2581 }
2582
2583 /** You can cast a Func to its pure stage for the purposes of
2584 * scheduling it. */
2585 operator Stage() const;
2586
2587 /** Get a handle on the output buffer for this Func. Only relevant
2588 * if this is the output Func in a pipeline. Useful for making
2589 * static promises about strides, mins, and extents. */
2590 // @{
2592 std::vector<OutputImageParam> output_buffers() const;
2593 // @}
2594
2595 /** Use a Func as an argument to an external stage. */
2596 operator ExternFuncArgument() const;
2597
2598 /** Infer the arguments to the Func, sorted into a canonical order:
2599 * all buffers (sorted alphabetically by name), followed by all non-buffers
2600 * (sorted alphabetically by name).
2601 This lets you write things like:
2602 \code
2603 func.compile_to_assembly("/dev/stdout", func.infer_arguments());
2604 \endcode
2605 */
2606 std::vector<Argument> infer_arguments() const;
2607
2608 /** Return the current StageSchedule associated with this initial
2609 * Stage of this Func. For introspection only: to modify schedule,
2610 * use the Func interface. */
2612 return Stage(*this).get_schedule();
2613 }
2614};
2615
2616namespace Internal {
2617
2618template<typename Last>
2619inline void check_types(const Tuple &t, int idx) {
2620 using T = typename std::remove_pointer<typename std::remove_reference<Last>::type>::type;
2621 user_assert(t[idx].type() == type_of<T>())
2622 << "Can't evaluate expression "
2623 << t[idx] << " of type " << t[idx].type()
2624 << " as a scalar of type " << type_of<T>() << "\n";
2625}
2626
2627template<typename First, typename Second, typename... Rest>
2628inline void check_types(const Tuple &t, int idx) {
2629 check_types<First>(t, idx);
2630 check_types<Second, Rest...>(t, idx + 1);
2631}
2632
2633template<typename Last>
2634inline void assign_results(Realization &r, int idx, Last last) {
2635 using T = typename std::remove_pointer<typename std::remove_reference<Last>::type>::type;
2636 *last = Buffer<T>(r[idx])();
2637}
2638
2639template<typename First, typename Second, typename... Rest>
2640inline void assign_results(Realization &r, int idx, First first, Second second, Rest &&...rest) {
2641 assign_results<First>(r, idx, first);
2642 assign_results<Second, Rest...>(r, idx + 1, second, rest...);
2643}
2644
2645} // namespace Internal
2646
2647/** JIT-Compile and run enough code to evaluate a Halide
2648 * expression. This can be thought of as a scalar version of
2649 * \ref Func::realize */
2650template<typename T>
2652 user_assert(e.type() == type_of<T>())
2653 << "Can't evaluate expression "
2654 << e << " of type " << e.type()
2655 << " as a scalar of type " << type_of<T>() << "\n";
2656 Func f;
2657 f() = e;
2659 return im();
2660}
2661
2662/** evaluate with a default user context */
2663template<typename T>
2665 return evaluate<T>(nullptr, e);
2666}
2667
2668/** JIT-compile and run enough code to evaluate a Halide Tuple. */
2669template<typename First, typename... Rest>
2671 Internal::check_types<First, Rest...>(t, 0);
2672
2673 Func f;
2674 f() = t;
2675 Realization r = f.realize(ctx);
2676 Internal::assign_results(r, 0, first, rest...);
2677}
2678
2679/** JIT-compile and run enough code to evaluate a Halide Tuple. */
2680template<typename First, typename... Rest>
2682 evaluate<First, Rest...>(nullptr, std::move(t), std::forward<First>(first), std::forward<Rest...>(rest...));
2683}
2684
2685namespace Internal {
2686
2687inline void schedule_scalar(Func f) {
2689 if (t.has_gpu_feature()) {
2691 }
2692 if (t.has_feature(Target::HVX)) {
2693 f.hexagon();
2694 }
2695}
2696
2697} // namespace Internal
2698
2699/** JIT-Compile and run enough code to evaluate a Halide
2700 * expression. This can be thought of as a scalar version of
2701 * \ref Func::realize. Can use GPU if jit target from environment
2702 * specifies one.
2703 */
2704template<typename T>
2706 user_assert(e.type() == type_of<T>())
2707 << "Can't evaluate expression "
2708 << e << " of type " << e.type()
2709 << " as a scalar of type " << type_of<T>() << "\n";
2710 Func f;
2711 f() = e;
2713 Buffer<T, 0> im = f.realize();
2714 return im();
2715}
2716
2717/** JIT-compile and run enough code to evaluate a Halide Tuple. Can
2718 * use GPU if jit target from environment specifies one. */
2719// @{
2720template<typename First, typename... Rest>
2722 Internal::check_types<First, Rest...>(t, 0);
2723
2724 Func f;
2725 f() = t;
2727 Realization r = f.realize();
2728 Internal::assign_results(r, 0, first, rest...);
2729}
2730// @}
2731
2732} // namespace Halide
2733
2734#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:688
EvictionKey(const Expr &expr=Expr())
Definition Func.h:694
A halide function.
Definition Func.h:703
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 & gpu_blocks(const VarOrRVar &block_x, DeviceAPI device_api=DeviceAPI::Default_GPU)
Tell Halide that the following dimensions correspond to GPU block indices.
Func & bound_extent(const Var &var, Expr extent)
Bound the extent of a Func's realization, but not its min.
void print_loop_nest()
Write out the loop nests specified by the schedule for this Function.
Func & unroll(const VarOrRVar &var, const Expr &factor, TailStrategy tail=TailStrategy::Auto)
Split a dimension by the given factor, then unroll the inner dimension.
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 & 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)
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 & memoize(const EvictionKey &eviction_key=EvictionKey())
Use the halide_memoization_cache_... interface to store a computed version of this function across in...
Func & partition(const VarOrRVar &var, Partition partition_policy)
Set the loop partition policy.
Func & trace_stores()
Trace all stores to the buffer backing this Func by emitting calls to halide_trace.
Func & trace_loads()
Trace all loads from this Func by emitting calls to halide_trace.
Func & never_partition_all()
Set the loop partition policy to Never for all Vars and RVar of the initial definition of the Func.
Func & prefetch(const Parameter &param, const VarOrRVar &at, const VarOrRVar &from, Expr offset=1, PrefetchBoundStrategy strategy=PrefetchBoundStrategy::GuardWithIf)
void specialize_fail(const std::string &message)
Add a specialization to a Func that always terminates execution with a call to halide_error().
Func & vectorize(const VarOrRVar &var, const Expr &factor, TailStrategy tail=TailStrategy::Auto)
Split a dimension by the given factor, then vectorize the inner dimension.
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 & 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.
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_...
Internal::Function function() const
Get a handle on the internal halide function that this Func represents.
Definition Func.h:2579
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 & compute_with(LoopLevel loop_level, const std::vector< std::pair< VarOrRVar, LoopAlignStrategy > > &align)
Func()
Declare a new undefined function with an automatically-generated unique name.
Func & store_in(MemoryType memory_type)
Set the type of memory this Func should be stored in.
const Type & type() const
Get the type(s) of the outputs of this Func.
void compile_to_bitcode(const std::string &filename, const std::vector< Argument > &, const Target &target=get_target_from_environment())
void infer_input_bounds(Pipeline::RealizationArg outputs, const Target &target=get_jit_target_from_environment())
Func & async()
Produce this Func asynchronously in a separate thread.
Func & reorder(const std::vector< VarOrRVar > &vars)
Reorder variables to have the given nesting order, from innermost out.
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:1259
Func & gpu_blocks(const VarOrRVar &block_x, const VarOrRVar &block_y, const VarOrRVar &block_z, DeviceAPI device_api=DeviceAPI::Default_GPU)
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 & gpu_lanes(const VarOrRVar &thread_x, DeviceAPI device_api=DeviceAPI::Default_GPU)
The given dimension corresponds to the lanes in a GPU warp.
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.
Stage update(int idx=0)
Get a handle on an update step for the purposes of scheduling it.
Func & gpu_blocks(const VarOrRVar &block_x, const VarOrRVar &block_y, DeviceAPI device_api=DeviceAPI::Default_GPU)
Func & gpu_threads(const VarOrRVar &thread_x, const VarOrRVar &thread_y, DeviceAPI device_api=DeviceAPI::Default_GPU)
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 & 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...
Func copy_to_host()
Declare that this function should be implemented by a call to halide_buffer_copy with a NULL target d...
void infer_input_bounds(JITUserContext *context, Pipeline::RealizationArg outputs, const Target &target=get_jit_target_from_environment())
Func & gpu_tile(const VarOrRVar &x, const VarOrRVar &tx, const Expr &x_size, TailStrategy tail=TailStrategy::Auto, DeviceAPI device_api=DeviceAPI::Default_GPU)
Func & serial(const VarOrRVar &var)
Mark a dimension to be traversed serially.
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.
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',...
Func & reorder_storage(const Var &x, const Var &y)
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 ...
Func(const Expr &e)
Declare a new function with an automatically-generated unique name, and define it to return the given...
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...
int dimensions() const
The dimensionality (number of arguments) of this function.
Func & gpu(const VarOrRVar &block_x, const VarOrRVar &block_y, const VarOrRVar &thread_x, const VarOrRVar &thread_y, DeviceAPI device_api=DeviceAPI::Default_GPU)
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...
const std::vector< Type > & types() const
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 & 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)
const std::vector< Expr > & update_args(int idx=0) const
Get the left-hand-side of the update definition.
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.
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:2081
Func & compute_root()
Compute all of this function once ahead of time.
Func & compute_with(LoopLevel loop_level, LoopAlignStrategy align=LoopAlignStrategy::Auto)
Func & trace_realizations()
Trace all realizations of this Func by emitting calls to halide_trace.
JITHandlers & jit_handlers()
Get a struct containing the currently set custom functions used by JIT.
std::vector< Var > args() const
Get the pure arguments.
Tuple update_values(int idx=0) const
Get the right-hand-side of an update definition for functions that returns multiple values.
Func & allow_race_conditions()
Specify that race conditions are permitted for this Func, which enables parallelizing over RVars even...
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 ...
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:1605
int num_update_definitions() const
How many update definitions does this function have?
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 & never_partition(const std::vector< VarOrRVar > &vars)
Set the loop partition policy to Never for a vector of Vars and RVars.
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 & 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 & store_at(LoopLevel loop_level)
Equivalent to the version of store_at that takes a Var, but schedules storage at a given LoopLevel.
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)
HALIDE_NO_USER_CODE_INLINE Func(Buffer< T, Dims > &im)
Construct a new Func to wrap a Buffer.
Definition Func.h:762
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 & 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...
Expr value() const
The right-hand-side value of the pure definition of this function.
Func & align_extent(const Var &var, Expr modulus)
Expand the region computed so that the extent is a multiple of 'modulus'.
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 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...
Func & atomic(bool override_associativity_test=false)
Issue atomic updates for this Func.
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:1175
void realize(Pipeline::RealizationArg outputs, const Target &target=Target())
Evaluate this function into an existing allocated buffer or buffers.
Func & unroll(const VarOrRVar &var)
Mark a dimension to be completely unrolled.
Func & set_estimates(const Region &estimates)
Set (min, extent) estimates for all dimensions in the Func at once; this is equivalent to calling set...
Func in()
Create and return a global identity wrapper, which wraps all calls to this Func by any other Func.
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.
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.
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_llvm_assembly(const std::string &filename, const std::vector< Argument > &, const Target &target=get_target_from_environment())
Func & store_at(const Func &f, const Var &var)
Allocate storage for this function within f's loop over var.
void add_custom_lowering_pass(T *pass)
Add a custom pass to be used during lowering.
Definition Func.h:1062
Func in(const std::vector< Func > &fs)
Create and return an identity wrapper shared by all the Funcs in 'fs'.
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 & hoist_storage_root()
Equivalent to Func::hoist_storage_root, but schedules storage outside the outermost loop.
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 & compute_at(LoopLevel loop_level)
Schedule a function to be computed within the iteration over a given LoopLevel.
Func & gpu_threads(const VarOrRVar &thread_x, DeviceAPI device_api=DeviceAPI::Default_GPU)
Tell Halide that the following dimensions correspond to GPU thread indices.
Func & always_partition(const std::vector< VarOrRVar > &vars)
Set the loop partition policy to Always for a vector of Vars and RVars.
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:1157
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.
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...
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...
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 ...
void clear_custom_lowering_passes()
Remove all previously-set custom lowering passes.
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...
const std::string & name() const
The name of this function, either given during construction, or automatically generated.
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 & hexagon(const VarOrRVar &x=Var::outermost())
Schedule for execution on Hexagon.
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 & store_root()
Equivalent to Func::store_at, but schedules storage outside the outermost loop.
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 & 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,...
void compile_to_assembly(const std::string &filename, const std::vector< Argument > &, const Target &target=get_target_from_environment())
std::vector< RVar > rvars(int idx=0) const
Get the RVars of the reduction domain for an update definition, if there is one.
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...
const std::vector< CustomLoweringPass > & custom_lowering_passes()
Get the custom lowering passes.
HALIDE_NO_USER_CODE_INLINE std::enable_if< Internal::all_are_convertible< Var, Args... >::value, FuncRef >::type operator()(Args &&...args) const
Definition Func.h:1242
Func & hoist_storage(const Func &f, const Var &var)
Hoist storage for this function within f's loop over var.
Func & compute_inline()
Aggressively inline all uses of this function.
Func(Internal::Function f)
Construct a new Func to wrap an existing, already-define Function object.
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...
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:1484
Func & bound_storage(const Var &dim, const Expr &bound)
Bound the extent of a Func's storage, but not extent of its compute.
Func & rename(const VarOrRVar &old_name, const VarOrRVar &new_name)
Rename a dimension.
Tuple values() const
The values returned by this function.
const std::string & extern_function_name() const
Get the name of the extern function called for an extern definition.
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 & parallel(const VarOrRVar &var)
Mark a dimension to be traversed in parallel.
Func & gpu_threads(const VarOrRVar &thread_x, const VarOrRVar &thread_y, const VarOrRVar &thread_z, 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())
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 & reorder_storage(const std::vector< Var > &dims)
Specify how the storage for the function is laid out.
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 & no_profiling()
Marks this function as a function that should not be profiled when using the target feature Profile o...
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:1185
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.
Func & vectorize(const VarOrRVar &var)
Mark a dimension to be computed all-at-once as a single vector.
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 & 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.
Func & fuse(const VarOrRVar &inner, const VarOrRVar &outer, const VarOrRVar &fused)
Join two dimensions into a single fused dimension.
Func in(const Func &f)
Creates and returns a new identity Func that wraps this Func.
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...
std::vector< OutputImageParam > output_buffers() const
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:1501
Func & prefetch(const T &image, const VarOrRVar &at, const VarOrRVar &from, Expr offset=1, PrefetchBoundStrategy strategy=PrefetchBoundStrategy::GuardWithIf)
Definition Func.h:2055
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 & 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 & 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...
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...
Func & always_partition_all()
Set the loop partition policy to Always for all Vars and RVar of the initial definition of the Func.
const Internal::StageSchedule & get_schedule() const
Return the current StageSchedule associated with this initial Stage of this Func.
Definition Func.h:2611
Func & gpu_single_thread(DeviceAPI device_api=DeviceAPI::Default_GPU)
Tell Halide to run this stage using a single gpu thread and block.
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:1167
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 ...
A fragment of front-end syntax of the form f(x, y, z), where x, y, z are Vars or Exprs.
Definition Func.h:494
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 Expr &)
Define a stage that divides this Func by the given expression.
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:591
Stage operator-=(const Tuple &)
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 &)
Stage operator-=(const Expr &)
Define a stage that adds the negative of the given expression to this Func.
Stage operator*=(const Expr &)
Define a stage that multiplies this Func by the given expression.
Stage operator+=(const Expr &)
Define a stage that adds the given expression to this Func.
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:613
int index() const
Return index to the function outputs.
Definition Func.h:677
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)
Internal::Function function() const
What function is this calling?
Definition Func.h:672
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 std::vector< Expr > & args() const
Get the default (no-specialization) arguments (left-hand-side) of the definition.
const StageSchedule & schedule() const
Get the default (no-specialization) stage-specific schedule associated with this definition.
bool defined() const
Definition objects are nullable.
const std::vector< StorageDim > & storage_dims() const
The list and order of dimensions used to store this function.
A reference-counted handle to Halide's internal representation of a function.
Definition Function.h:39
FuncSchedule & schedule()
Get a handle to the function-specific schedule for the purpose of modifying it.
const std::vector< std::string > & args() const
Get the pure arguments.
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:680
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:70
Stage & prefetch(const Func &f, const VarOrRVar &at, const VarOrRVar &from, Expr offset=1, PrefetchBoundStrategy strategy=PrefetchBoundStrategy::GuardWithIf)
std::string name() const
Return the name of this stage, e.g.
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 & tile(const std::vector< VarOrRVar > &previous, 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)
Stage & gpu_threads(const VarOrRVar &thread_x, const VarOrRVar &thread_y, 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:386
Stage & compute_with(const Stage &s, const VarOrRVar &var, const std::vector< std::pair< VarOrRVar, LoopAlignStrategy > > &align)
Func rfactor(const RVar &r, const Var &v)
Stage & gpu(const VarOrRVar &block_x, const VarOrRVar &thread_x, DeviceAPI device_api=DeviceAPI::Default_GPU)
Stage & vectorize(const VarOrRVar &var)
Stage & never_partition(const std::vector< VarOrRVar > &vars)
Stage & gpu_lanes(const VarOrRVar &thread_x, DeviceAPI device_api=DeviceAPI::Default_GPU)
Stage & unroll(const VarOrRVar &var)
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 & 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 & 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 & allow_race_conditions()
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 & parallel(const VarOrRVar &var, const Expr &task_size, TailStrategy tail=TailStrategy::Auto)
Stage & rename(const VarOrRVar &old_name, const VarOrRVar &new_name)
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_single_thread(DeviceAPI device_api=DeviceAPI::Default_GPU)
Stage & partition(const VarOrRVar &var, Partition partition_policy)
Stage & unroll(const VarOrRVar &var, const Expr &factor, TailStrategy tail=TailStrategy::Auto)
Stage specialize(const Expr &condition)
Stage & prefetch(const T &image, const VarOrRVar &at, const VarOrRVar &from, Expr offset=1, PrefetchBoundStrategy strategy=PrefetchBoundStrategy::GuardWithIf)
Definition Func.h:471
Stage & gpu_threads(const VarOrRVar &thread_x, const VarOrRVar &thread_y, const VarOrRVar &thread_z, DeviceAPI device_api=DeviceAPI::Default_GPU)
Stage & prefetch(const Parameter &param, const VarOrRVar &at, const VarOrRVar &from, Expr offset=1, PrefetchBoundStrategy strategy=PrefetchBoundStrategy::GuardWithIf)
Stage & reorder(const std::vector< VarOrRVar > &vars)
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(Internal::Function f, Internal::Definition d, size_t stage_index)
Definition Func.h:96
Stage & gpu_blocks(const VarOrRVar &block_x, const VarOrRVar &block_y, const VarOrRVar &block_z, 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 & compute_with(LoopLevel loop_level, LoopAlignStrategy align=LoopAlignStrategy::Auto)
Stage & parallel(const VarOrRVar &var)
const Internal::StageSchedule & get_schedule() const
Return the current StageSchedule associated with this Stage.
Definition Func.h:109
Stage & serial(const VarOrRVar &var)
Stage & gpu_blocks(const VarOrRVar &block_x, DeviceAPI device_api=DeviceAPI::Default_GPU)
Stage & fuse(const VarOrRVar &inner, const VarOrRVar &outer, const VarOrRVar &fused)
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:393
Stage & vectorize(const VarOrRVar &var, const Expr &factor, TailStrategy tail=TailStrategy::Auto)
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 & compute_with(const Stage &s, const VarOrRVar &var, LoopAlignStrategy align=LoopAlignStrategy::Auto)
void specialize_fail(const std::string &message)
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:400
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 & gpu_blocks(const VarOrRVar &block_x, const VarOrRVar &block_y, DeviceAPI device_api=DeviceAPI::Default_GPU)
Stage & hexagon(const VarOrRVar &x=Var::outermost())
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 & always_partition_all()
Stage & never_partition_all()
Stage & atomic(bool override_associativity_test=false)
Stage & gpu_threads(const VarOrRVar &thread_x, DeviceAPI device_api=DeviceAPI::Default_GPU)
Func rfactor(const std::vector< std::pair< RVar, Var > > &preserved)
Calling rfactor() on an associative update definition a Func will split the update into an intermedia...
std::string dump_argument_list() const
Return a string describing the current var list taking into account all the splits,...
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)
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)
void unscheduled()
Assert that this stage has intentionally been given no schedule, and suppress the warning about unsch...
Stage & always_partition(const std::vector< VarOrRVar > &vars)
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:162
void schedule_scalar(Func f)
Definition Func.h:2687
std::vector< Var > make_argument_list(int dimensionality)
Make a list of unique arguments for definitions with unnamed arguments.
void assign_results(Realization &r, int idx, Last last)
Definition Func.h:2634
void check_types(const Tuple &t, int idx)
Definition Func.h:2619
ForType
An enum describing a type of loop traversal.
Definition Expr.h:406
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:2705
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.
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.
Internal::ConstantInterval cast(Type t, const Internal::ConstantInterval &a)
Cast operators for ConstantIntervals.
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:485
std::vector< Range > Region
A multi-dimensional box.
Definition Expr.h:350
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:2651
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.
Represent the equivalent associative op of an update definition.
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
VarOrRVar(const Var &v)
Definition Func.h:33
VarOrRVar(const RVar &r)
Definition Func.h:36
VarOrRVar(const std::string &n, bool r)
Definition Func.h:30
VarOrRVar(const ImplicitVar< N > &u)
Definition Func.h:43
const std::string & name() const
Definition Func.h:47
VarOrRVar(const RDom &r)
Definition Func.h:39
#define user_assert(c)
Definition test.h:10