Halide 18.0.0
Halide compiler and libraries
Loading...
Searching...
No Matches
LoopNest.h
Go to the documentation of this file.
1/** This file defines the LoopNest, which is our
2 * representation of a Halide schedule, and contains methods to
3 * generate candidates for scheduling as well as extract a
4 * featurization that can be used to cost each candidate. */
5
6#ifndef LOOP_NEST_H
7#define LOOP_NEST_H
8
9#include "ASLog.h"
10#include "CostModel.h"
11#include "FunctionDAG.h"
12#include "GPULoopInfo.h"
13#include "GPUMemInfo.h"
14#include "PerfectHashMap.h"
15#include "SearchSpaceOptions.h"
16#include "Statistics.h"
17#include "ThreadInfo.h"
18#include "Tiling.h"
19#include <set>
20#include <vector>
21
22namespace Halide {
23namespace Internal {
24namespace Autoscheduler {
25
26template<typename T>
28
29template<typename T>
31
32enum class GPU_parallelism {
33 Block,
34 Thread,
35 Serial,
36 Simd,
38 None
39};
40
41std::string stringify(GPU_parallelism label);
42
43// inlined => func is inlined so has no memory store location
44enum class GPUMemoryType {
45 Global,
46 Shared,
47 Local,
50};
51
52bool may_subtile(const Anderson2021Params &params);
53
55
57
59
61 return 128;
62}
63
64int get_unroll_limit(const Target &target);
65
66bool in_range_zero_one(double x);
67
68bool are_valid_thread_extents(const vector<int64_t> &counts);
69
72
73bool all(const vector<int> &v);
74bool accessed_at_constant_indices(const std::vector<int> &unrolled, const FunctionDAG::Edge *e);
75
76// We're going to do a tree search over possible schedules to find an
77// optimal one. A tree search requires a state, and a function that
78// gives you children of the state (with costs). The following struct
79// represents the state, which is a partial schedule.
80//
81// A partial schedule is a tree. Each node is some portion of the for
82// loop nest of some Func. If there are no children, it's the
83// innermost set of loops. If there are children, it's a loop over
84// tiles of that Func.
85struct LoopNest {
86 mutable RefCount ref_count;
87
88 // The extents of this loop. Put another way, the number of tiles,
89 // not the size of each tile.
90 vector<int64_t> size;
91
92 // The nodes inside the loop body
94
95 // Funcs inlined into this inner loop, and the number of times
96 // each is called. Only valid if children is empty.
98
99 // Funcs stored inside this loop
100 std::set<const FunctionDAG::Node *> store_at;
101
102 // The total bounds required of any given Func over all iterations
103 // of this loop. In the paper, this is represented using the
104 // little boxes to the left of the loop nest tree figures.
105 mutable NodeMap<Bound> bounds;
106
107 // The Func this loop nest belongs to
108 const FunctionDAG::Node *node = nullptr;
109
110 // The stage of the Func
111 const FunctionDAG::Node::Stage *stage = nullptr;
112
113 // Is this the innermost loop of this func (the SIMD loop)?
114 bool innermost = false;
115
116 // Are we permitted to tile this loop?
117 bool tileable = false;
118
119 // Is this the parallel outer loop?
120 bool parallel = false;
121
122 // What dimension is this Func vectorized over, in terms of the pure args of the Func?
123 int vector_dim = -1;
124
125 // Which loop corresponds to the innermost storage dimension and will be vectorized. -1 means none of them.
126 int vectorized_loop_index = -1;
127
128 // Apply gpu threads to this loop nest
130
142
143 mutable std::map<uint64_t, StageMap<StageMap<FeatureIntermediates>>> feature_intermediates;
144 mutable std::map<uint64_t, StageMap<ScheduleFeatures>> features;
145
146 bool is_gpu_serial(const Target &target) const {
148 }
149
150 bool is_gpu_thread(const Target &target) const {
152 }
153
154 bool is_gpu_block(const Target &target) const {
156 }
157
158 bool is_scalar() const {
159 return size.empty();
160 }
161
162 // given a newly inserted node f into this LoopNest, get union of thread counts in each dimension
163 // across all siblings of f.
164 vector<int64_t> get_union_thread_counts(const FunctionDAG::Node *f) const;
165
166 // given a newly inserted node f into this LoopNest, gets the size of
167 // all of f's stages and their pure_dim indices
169 vector<vector<int64_t>> &stage_sizes,
170 vector<vector<int>> &pure_dims,
171 vector<int> &vectorized_indices) const;
172
173 // given the loop nest of a stage to parallelize at root, figure out if using odd tile sizes
174 // for the vectorized dimension will allow the resulting thread tiles to be multiples of 32
175 // if so, we will include these in the serial loop sizes
177
178 // get the loop nests of a newly inserted node, f, that is marked GPU threads. Tiles
179 // the newly inserted loop nests of f into a threads loop outside a serial loop.
180 // V is the vectorized dimension of f. Adds loopnests created from each tiling option in result.
182 const Anderson2021Params &params,
183 const Target &target,
184 int v,
185 vector<IntrusivePtr<const LoopNest>> &result,
186 const vector<int64_t> &max_size);
187
188 void copy_from(const LoopNest &n);
190
191 static void hash_combine(uint64_t &h, uint64_t next) {
192 // From boost
193 h ^= (next + 0x9e3779b9 + (h << 6) + (h >> 2));
194 }
195
196 // Hash the loop structure and sizes up to a fixed depth. This is
197 // used as the hash function for the coarse-to-fine beam search in
198 // the paper.
199 void structural_hash(uint64_t &h, int depth) const;
200
201 // How many funcs are scheduled inside this loop level. Used in
202 // the structural hash.
204 size_t count = inlined.size() + store_at.size();
205 for (const auto &c : children) {
206 count += c->funcs_realized_or_inlined();
207 }
208 return count;
209 }
210
211 // All of a stage's interesting locations in the loop nest. Used to help compute the featurization of a stage.
212 struct Sites {
213 const LoopNest *compute = nullptr; // Its containing compute_at site
214 const LoopNest *store = nullptr; // Its containing store_at site
215 const LoopNest *produce = nullptr; // Its own outermost node
216 const LoopNest *innermost = nullptr; // Its innermost node - usually a SIMD loop
217 const LoopNest *task = nullptr; // The parallel for loop it belongs to
218 const LoopNest *thread = nullptr; // Its containing gpu_thread loop
219 GPUMemoryType gpu_store_memory_type; // global, local, shared?
220 int64_t allocation_size = 0; // Allocation size in bytes
221 bool is_constant_allocation = false; // Does the allocation have constant size?
222 int64_t num_realizations = 0; // Number of times this stage is realized. Only valid for unscheduled producers
223 bool inlined = false; // Is the Func inlined?
224 std::vector<const LoopNest *> inlined_innermosts; // Is the Func inlined?
226
239 };
240
242 bool in_thread,
243 bool is_inlined = false) const;
244
245 std::vector<int> unrolled_loops(const Target &target,
246 const LoopNest *parent,
247 const LoopNest *grandparent) const;
248
252 const LoopNest *grandparent,
253 const LoopNest *parent) const;
254
256 StageMap<Sites> &sites) const;
257
258 // Compute all the sites of interest for each pipeline stage
259 void get_sites(const Target &target,
262 const LoopNest *task = nullptr,
263 const LoopNest *parent = nullptr,
264 const LoopNest *current_thread_loop = nullptr) const;
265
266 // A helper for the working_set_at_task feature. Most features are
267 // computed in the recursive pass 'compute_features' below, but
268 // this one must be done in a second separate recursive pass.
271 for (const auto &c : children) {
272 c->set_working_set_at_task_feature(working_set, features);
273 features->get(c->stage).working_set_at_task = working_set;
274 }
275 }
276
278 const LoopNest *parent,
279 bool in_threads_loop) const;
280
282
284
286
288
290
291 // Get the stride over "node's" storage for a unit increment in the vectorized loop's
292 // index
296 const Bound &store_bounds,
297 const LoopNest &root) const;
298
302 const Bound &store_bounds,
303 const ThreadInfo *thread_info,
304 bool verbose = false) const;
305
308 const LoopNest &root) const;
309
311
314 const FunctionDAG::Node *node,
317 const std::vector<int64_t> &inner_serial_loop_extents,
318 const Sites &consumer_site,
320 const LoopNest *parent,
321 const LoopNest &root,
325 bool verbose = false) const;
326
329 int innermost_dim,
330 int loop_index) const;
331
335 int innermost_dim,
336 int loop_index,
337 const GPUMemoryType &mem_type) const;
338
342 int innermost_dim,
343 const GPUMemoryType &mem_type,
344 bool verbose = false) const;
345
347 bool verbose = false) const;
348
349 template<typename T>
351 const FunctionDAG::Node *node,
352 const Bound &store_bounds,
353 const ThreadInfo *thread_info,
354 int innermost_dim,
357 bool verbose = false) const;
358
359 std::pair<double, double> compute_local_mem_store_features(const LoadJacobian &jac,
361 const FunctionDAG::Node *node,
363 const LoopNest &root,
364 double serial_loop_extents) const;
365
366 template<typename T>
369 const FunctionDAG::Node *node,
371 const ThreadInfo *thread_info,
372 double serial_loop_extents,
373 bool verbose) const;
374
375 template<typename T>
378 const FunctionDAG::Node *node,
381 const ThreadInfo *thread_info,
383 double serial_loop_extents,
384 bool verbose = false) const;
385
386 double compute_local_mem_stride(double stride,
387 double bytes) const;
388
389 // Assumes block, serial, thread or block, thread nesting
391 const LoopNest *grandparent) const;
392
393 std::pair<int64_t, int64_t> get_block_and_serial_extents(const LoopNest *block) const;
394
396
398
400 const GPULoopInfo &gpu_loop_info) const;
401
402 // Assume that when a block is active, all its warps are active
405 const GPULoopInfo &gpu_loop_info) const;
406
408 const Target &target,
410 ScheduleFeatures &feat) const;
411
412 std::pair<const LoopNest *, const LoopNest *> find_innermost_and_parent() const;
413
415 const Target &target,
417 const std::vector<const FunctionDAG::Edge *> &edge_chain,
418 const LoadJacobian &jac,
419 const LoopNest *parent,
420 const LoopNest *grandparent,
421 int64_t n,
422 const ScheduleFeatures &feat,
426 const GPUMemoryType &mem_type,
427 bool verbose) const;
428
430 const LoopNest *parent,
431 const ScheduleFeatures &feat,
432 const LoadJacobian &jac,
433 int producer_dims) const;
434
437
439
441
442 void collect_stages(std::set<const FunctionDAG::Node::Stage *> &stages) const;
443
446
449
452
453 std::pair<int64_t, bool> compute_alloc_size_of_node_here(const FunctionDAG::Node *f) const;
454
455 // Do a recursive walk over the loop nest computing features to feed the cost model.
457 const Anderson2021Params &params,
458 const Target &target,
459 const StageMap<Sites> &sites,
460 int64_t instances,
461 int64_t parallelism,
462 const LoopNest *parent,
463 const LoopNest *grandparent,
464 const LoopNest &root,
468 int64_t *working_set,
472 Statistics &stats,
473 bool verbose = false) const;
474
475 bool is_root() const {
476 // The root is the sole node without a Func associated with
477 // it.
478 return node == nullptr;
479 }
480
481 // Set the region required of a Func at this site.
482 const Bound &set_bounds(const FunctionDAG::Node *f, BoundContents *b) const {
483 return bounds.emplace(f, b);
484 }
485
486 // Get the region required of a Func at this site, from which we
487 // know what region would be computed if it were scheduled here,
488 // and what its loop nest would be.
489 const Bound &get_bounds(const FunctionDAG::Node *f) const;
490
491 // Get the region required of a Func at this site (but only to satisfy the
492 // consumers along the given edge chain), from which we know what region
493 // would be computed if it were scheduled here and what its loop nest
494 // would be.
497
498 void dump() const;
499
500 std::string to_string() const;
501
502 // Recursively print a loop nest representation to stderr
503 template<typename T>
504 void dump(T &stream, string prefix, const LoopNest *parent) const;
505
506 // Does this loop nest access the given Func
507 bool calls(const FunctionDAG::Node *f) const;
508
509 // What is the maximum number of inlined calls to a Func that
510 // occur within this loop. Used to prune states that would
511 // generate too much code.
513
514 // Does this loop nest access an input buffer? Used to select
515 // trail strategies when splitting loops. We don't want to read
516 // out of bounds on inputs, even if we don't intend to use the
517 // values read. It could create annoying assertion failures for
518 // the user. It's OK to read out of range of the values computed
519 // on internal Funcs though. Allocation bounds inference just pads
520 // out the bounds so that it won't fault.
522
523 // Does this loop nest contain a computation of the given Func.
524 bool computes(const FunctionDAG::Node *f) const;
525
526 // Above here most methods query the loop nest. Below we have
527 // methods that mutate the loop nest.
528
529 // Inline a Func into all consumers within this loop.
531
532 // Compute a Func at this site.
534 bool tileable,
535 int v,
536 bool in_threads_loop,
537 const Anderson2021Params &params,
538 const Target &target);
539
540 // Parallelize this loop according to the given tiling.
542 const LoopNest *parent,
543 const Anderson2021Params &params,
544 const Target &target,
545 bool inner_tiling,
546 bool adjust_tiling,
547 bool move_all_rvars_inward = true,
548 const vector<int> &rvars_to_move_inward = {}) const;
549
551 bool in_threads_loop = false) const;
553
554 // All store ats further in than the block level must be fixed
555 // sized allocations. This method checks if f will require a dynamic
556 // allocation
558 const Target &target,
559 bool in_threads_loop) const;
560
561 // Return all possible ways to compute f in tiles somewhere within
562 // this loop nest.
563 // in_threads_loop tracks whether or not function is going to be placed inside a
564 // loop marked gpu_threads, in which case f's loops cannot be gpu_threads
566 const LoopNest *parent,
567 const Anderson2021Params &params,
568 const Target &target,
569 const SearchSpaceOptions &search_space_options,
570 int v,
571 bool in_realization,
572 bool in_threads_loop,
573 bool is_pre_pass,
574 vector<int64_t> union_counts = vector<int64_t>()) const;
575
576 // Below here we have methods that apply a schedule to a Halide pipeline.
577
578 // A model of the state of the loop nest of a Func while applying
579 // Halide's scheduling directives.
580
581 // Note that StageScheduleState is movable-but-not-copyable thanks to its ostringstream member.
582 struct StageScheduleState {
583 // How much parallelism do we need to exploit with this Func?
584 double num_cores = 0;
585
586 // Which storage dimension is vectorized? We need to reorder it innermost
587 int vector_dim = -1;
588 int vectorized_loop_index = -1;
589
590 // The various Vars and RVars used for scheduling a Func.
591 struct FuncVar {
592 // The top-level var or rvar this was split off from
594
595 // This var.
597
598 // Source code to access this Var/RVar. Used for printing
599 // valid Halide source for this schedule.
600 string accessor;
601
602 // Our estimate of the extent of this var. This is exact
603 // when constant_extent flag is true.
604 int64_t extent = 0;
605
606 // Which index in the symbolic loop nest does this var
607 // belong to.
608 size_t index = 0;
609
610 // Some flags.
611 bool innermost_pure_dim = false;
612 bool outermost = false;
613 bool parallel = false;
614 bool exists = false;
615 bool pure = false;
616 bool constant_extent = false;
617
618 bool vectorized = false;
619 bool gpu_threads = false;
620
622 : orig(Var()),
623 var(Var()) {
624 }
625 };
628 bool parallel = false;
629 bool vectorized = false;
632
633 // In order from innermost to outermost. Each group of d is one tiling level.
635
636 // In order from innermost to outermost. Each group of d is one tiling level.
638 vector<int64_t> gpu_thread_extents;
639
642
643 // From outermost in
645
646 std::ostringstream schedule_source;
647 };
648
653 int num_serial_loops() const;
655
657 const NodeMap<bool> &all_inlined) const;
659 const LoopNest *parent) const;
660
661 // Apply the schedule represented by this loop nest to a Halide pipeline.
663 StageMap<std::unique_ptr<StageScheduleState>> &state_map,
664 double num_cores,
665 int depth,
666 const LoopNest *parent,
667 const LoopNest *compute_site,
668 const Target &target,
669 std::vector<StageScheduleState *> &ancestors,
670 const NodeMap<bool> &all_inlined) const;
671
672 double max_idle_lane_wastage(const Target &target,
674
676
678 NodeMap<bool> &inlined_nodes) const;
679
681
684
686 const LoopNest *compute_root_loop_nest = nullptr) const;
687};
688
689struct Filter {
691 bool logging = false;
692
693 explicit Filter(const LoopNest *loop_nest)
696 if (logging) {
697 std::cerr << "\nState filtered: \n";
698 loop_nest->dump();
699 std::cerr << "Reason: ";
700 }
701 }
702
703 template<typename T>
705 if (logging) {
706 std::cerr << std::forward<T>(x);
707 }
708 return *this;
709 }
710
712};
713
714} // namespace Autoscheduler
715} // namespace Internal
716} // namespace Halide
717
718#endif // LOOP_NEST_H
Data structure containing information about the current GPU loop nest hierarchy of blocks,...
Data structures that help track memory access information.
Data structure containing information about GPU threads for a particular location in the loop nest an...
A class representing a reference count to be used with IntrusivePtr.
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 variable, to be used when defining functions.
Definition Var.h:19
int64_t get_active_block_hardware_limit(const Anderson2021Params &params)
bool all(const vector< int > &v)
bool are_valid_thread_extents(const vector< int64_t > &counts)
bool accessed_at_constant_indices(const std::vector< int > &unrolled, const FunctionDAG::Edge *e)
constexpr int64_t get_register_mem_alloc_limit()
Definition LoopNest.h:60
PerfectHashMap< FunctionDAG::Node::Stage, T > StageMap
Definition LoopNest.h:24
PerfectHashMap< FunctionDAG::Node, T > NodeMap
Definition LoopNest.h:21
int64_t get_active_warp_hardware_limit(const Anderson2021Params &params)
bool may_subtile(const Anderson2021Params &params)
int get_unroll_limit(const Target &target)
int64_t get_shared_memory_limit(const Anderson2021Params &params)
std::string stringify(GPU_parallelism label)
This file defines the class FunctionDAG, which is our representation of a Halide pipeline,...
@ Internal
Not visible externally, similar to 'static' linkage in C.
Internal::ConstantInterval cast(Type t, const Internal::ConstantInterval &a)
Cast operators for ConstantIntervals.
unsigned __INT64_TYPE__ uint64_t
signed __INT64_TYPE__ int64_t
Filter(const LoopNest *loop_nest)
Definition LoopNest.h:693
std::vector< const LoopNest * > inlined_innermosts
Definition LoopNest.h:224
NodeMap< std::vector< std::pair< const LoopNest *, std::vector< const FunctionDAG::Edge * > > > > producers_to_be_staged
Definition LoopNest.h:641
vector< pair< int, int > > collect_producers(const StageMap< Sites > &sites) const
bool is_gpu_thread(const Target &target) const
Definition LoopNest.h:150
vector< IntrusivePtr< const LoopNest > > compute_in_tiles(const FunctionDAG::Node *f, const LoopNest *parent, const Anderson2021Params &params, const Target &target, const SearchSpaceOptions &search_space_options, int v, bool in_realization, bool in_threads_loop, bool is_pre_pass, vector< int64_t > union_counts=vector< int64_t >()) const
const LoopNest * get_enclosing_block(const LoopNest *parent, const LoopNest *grandparent) const
int num_serial_loops(const FunctionDAG::Node::Stage *stage) const
int64_t points_accessed_per_thread(const Anderson2021Params &params, const Target &target, const GPULoopInfo &gpu_loop_info, const std::vector< const FunctionDAG::Edge * > &edge_chain, const LoadJacobian &jac, const LoopNest *parent, const LoopNest *grandparent, int64_t n, const ScheduleFeatures &feat, const LoadJacobian &serial_jac, bool producer_has_been_scheduled, int producer_innermost_dim, const GPUMemoryType &mem_type, bool verbose) const
bool has_constant_region_required(const FunctionDAG::Node *node) const
std::map< uint64_t, StageMap< ScheduleFeatures > > features
Definition LoopNest.h:144
int get_pure_stage_vectorized_loop_index(const FunctionDAG::Node *node) const
const FunctionDAG::Node * node
Definition LoopNest.h:57
void dump(T &stream, string prefix, const LoopNest *parent) const
int64_t product_of_self_and_descendants(int loop_index) const
bool all_strides_exist(const LoadJacobian &jac, const FunctionDAG::Node *storage_node, const LoopNest &root) const
void recompute_inlined_features(const StageMap< Sites > &sites, StageMap< ScheduleFeatures > *features) const
void inline_func(const FunctionDAG::Node *f)
void generate_vec_dim_serial_tilings(vector< int > &serial_sizes) const
bool region_computed_shrinks(const FunctionDAG::Node *f, const LoopNest *parent) const
void compute_gpu_store_features(const LoadJacobian &jac, int consumer_innermost_dim, const FunctionDAG::Node *node, const Bound &consumer_store_bounds, const GPULoopInfo &gpu_loop_info, const std::vector< int64_t > &inner_serial_loop_extents, const Sites &consumer_site, ScheduleFeatures &feat, const LoopNest *parent, const LoopNest &root, GlobalMemInfo &global_mem_loads, SharedMemInfo &shared_mem_loads, LocalMemInfo &local_mem_loads, bool verbose=false) const
int64_t compute_licm_amortization(const LoopNest *innermost, const LoopNest *parent, const ScheduleFeatures &feat, const LoadJacobian &jac, int producer_dims) const
int64_t product_of_descendants(int loop_index) const
bool requires_dynamic_allocation(const FunctionDAG::Node *f, const Target &target, bool in_threads_loop) const
bool is_gpu_block(const Target &target) const
Definition LoopNest.h:154
bool node_has_dynamic_region_computed(const FunctionDAG::Node *f) const
bool exceeds_serial_extents_limit(const Target &target, const LoopNest *parent, bool in_threads_loop) const
void collect_stages(std::set< const FunctionDAG::Node::Stage * > &stages) const
Bound get_bounds_along_edge_chain(const FunctionDAG::Node *f, const vector< const FunctionDAG::Edge * > &edge_chain) const
int64_t get_total_constant_local_mem_alloc_size() const
const Bound & get_bounds(const FunctionDAG::Node *f) const
std::pair< double, double > compute_local_mem_store_features(const LoadJacobian &jac, int consumer_innermost_dim, const FunctionDAG::Node *node, const Bound &consumer_store_bounds, const LoopNest &root, double serial_loop_extents) const
int vectorized_load_access_size(const LoadJacobian &jac, const FunctionDAG::Node *accessed, bool accessed_has_been_scheduled, int innermost_dim, const GPUMemoryType &mem_type, bool verbose=false) const
GPUMemoryType get_gpu_memory_type(bool in_block, bool in_thread, bool is_inlined=false) const
void compute_warp_and_block_occupancy(const Anderson2021Params &params, ScheduleFeatures &feat, const GPULoopInfo &gpu_loop_info) const
vector< int64_t > get_union_thread_counts(const FunctionDAG::Node *f) const
void collect_nodes_that_should_be_inlined(const NodeMap< bool > &nodes_to_freeze, NodeMap< bool > &inlined_nodes) const
void apply(LoopLevel here, StageMap< std::unique_ptr< StageScheduleState > > &state_map, double num_cores, int depth, const LoopNest *parent, const LoopNest *compute_site, const Target &target, std::vector< StageScheduleState * > &ancestors, const NodeMap< bool > &all_inlined) const
std::map< uint64_t, StageMap< StageMap< FeatureIntermediates > > > feature_intermediates
Definition LoopNest.h:143
int get_actual_vector_dim(const Bound &store_bounds) const
void compute_shared_mem_occupancy(const Anderson2021Params &params, const Target &target, int64_t total_shared_mem_alloc_size, ScheduleFeatures &feat) const
Strides compute_strides(const LoadJacobian &jac, int innermost_storage_dim, const FunctionDAG::Node *storage_node, const Bound &store_bounds, const ThreadInfo *thread_info, bool verbose=false) const
int get_vectorized_loop_index_from_pure_stage(const LoopNest &root) const
bool computes(const FunctionDAG::Node *f) const
double storage_stride(const LoadJacobian &jac, int innermost_storage_dim, const FunctionDAG::Node *storage_node, const Bound &store_bounds, const LoopNest &root) const
std::vector< IntrusivePtr< const LoopNest > > children
Definition LoopNest.h:42
void set_working_set_at_task_feature(int64_t working_set, StageMap< ScheduleFeatures > *features) const
Definition LoopNest.h:269
bool promote_allocs_to_registers(const Target &target, StageMap< Sites > &sites) const
std::pair< int64_t, int64_t > get_block_and_serial_extents(const LoopNest *block) const
const Bound & set_bounds(const FunctionDAG::Node *f, BoundContents *b) const
Definition LoopNest.h:482
bool has_dynamic_allocation_inside_thread(bool in_thread_loop) const
void memoize_points_computed_minimum(StageMap< ScheduleFeatures > &memoized_features, const StageMap< ScheduleFeatures > *features) const
bool has_constant_region_computed(const FunctionDAG::Node *node) const
double compute_local_mem_stride(double stride, double bytes) const
MemInfoType< T > compute_mem_store_info(const LoadJacobian &jac, int consumer_innermost_dim, const FunctionDAG::Node *node, const Bound &consumer_store_bounds, const ThreadInfo *thread_info, double serial_loop_extents, bool verbose) const
bool add_gpu_thread_tilings(const FunctionDAG::Node *f, const Anderson2021Params &params, const Target &target, int v, vector< IntrusivePtr< const LoopNest > > &result, const vector< int64_t > &max_size)
IntrusivePtr< const LoopNest > parallelize_in_tiles(const vector< int64_t > &tiling, const LoopNest *parent, const Anderson2021Params &params, const Target &target, bool inner_tiling, bool adjust_tiling, bool move_all_rvars_inward=true, const vector< int > &rvars_to_move_inward={}) const
void memoize_features(StageMap< ScheduleFeatures > &memoized_features, const StageMap< ScheduleFeatures > *features) const
void collect_all_inlined(NodeMap< bool > &all_inlined) const
std::vector< int > unrolled_loops(const Target &target, const LoopNest *parent, const LoopNest *grandparent) const
bool is_gpu_serial(const Target &target) const
Definition LoopNest.h:146
double max_idle_lane_wastage(const Target &target, GPULoopInfo gpu_loop_info) const
static void hash_combine(uint64_t &h, uint64_t next)
Definition LoopNest.h:191
const LoopNest * find_pure_stage_loop_nest(const FunctionDAG::Node *node) const
void dump(std::ostream &os, string prefix, const LoopNest *parent) const
void get_stage_sizes(const FunctionDAG::Node *f, vector< vector< int64_t > > &stage_sizes, vector< vector< int > > &pure_dims, vector< int > &vectorized_indices) const
uint64_t compute_hash_of_producers_stored_at_root(const StageMap< Sites > &sites) const
const FunctionDAG::Node::Stage * stage
Definition LoopNest.h:60
bool other_stage_has_same_producer(const FunctionDAG::Node *producer) const
void get_stages_computed_in_each_compute_root_loop(StageMap< StageMap< bool > > &descendants, const LoopNest *compute_root_loop_nest=nullptr) const
void compute_features(const FunctionDAG &dag, const Anderson2021Params &params, const Target &target, const StageMap< Sites > &sites, int64_t instances, int64_t parallelism, const LoopNest *parent, const LoopNest *grandparent, const LoopNest &root, GPULoopInfo gpu_loop_info, bool use_memoized_features, const StageMap< int64_t > &total_shared_mem_alloc_sizes, int64_t *working_set, int64_t *working_set_local_constant, int64_t *working_set_local_dynamic, StageMap< ScheduleFeatures > *features, Statistics &stats, bool verbose=false) const
void get_sites(const Target &target, StageMap< Sites > &sites, StageMap< int64_t > &shared_mem_alloc_sizes, const LoopNest *task=nullptr, const LoopNest *parent=nullptr, const LoopNest *current_thread_loop=nullptr) const
bool calls(const FunctionDAG::Node *f) const
bool producer_computed_here_or_further_in(const FunctionDAG::Node *producer) const
void copy_from_including_features(const LoopNest &n)
bool can_vectorize_access_for_innermost_dim(const LoadJacobian &jac, const FunctionDAG::Node *accessed, int innermost_dim, int loop_index) const
void update_producers_to_be_staged(StageScheduleState &state, const NodeMap< bool > &all_inlined) const
bool can_vectorize_store_access(const LoadJacobian &jac, const FunctionDAG::Node *accessed, bool accessed_has_been_scheduled, int innermost_dim, int loop_index, const GPUMemoryType &mem_type) const
vector< IntrusivePtr< const LoopNest > > children
Definition LoopNest.h:93
void compute_mem_load_features(const LoadJacobian &jac, int producer_innermost_dim, const FunctionDAG::Node *node, const Bound &producer_store_bounds, bool producer_has_been_scheduled, const ThreadInfo *thread_info, MemInfoType< T > &mem_info, double serial_loop_extents, bool verbose=false) const
void get_allocs_that_can_be_promoted_to_registers(const Target &target, StageMap< Sites > &sites, NodeMap< bool > &can_be_promoted_to_registers, const LoopNest *grandparent, const LoopNest *parent) const
void compute_warp_features(ScheduleFeatures &features, const GPULoopInfo &gpu_loop_info) const
void structural_hash(uint64_t &h, int depth) const
void compute_num_mem_accesses_per_block(const LoadJacobian &jac, const FunctionDAG::Node *node, const Bound &store_bounds, const ThreadInfo *thread_info, int innermost_dim, double num_requests_per_warp, MemInfoType< T > &mem_info, bool verbose=false) const
std::pair< const LoopNest *, const LoopNest * > find_innermost_and_parent() const
std::set< const FunctionDAG::Node * > store_at
Definition LoopNest.h:49
std::pair< int64_t, bool > compute_alloc_size_of_node_here(const FunctionDAG::Node *f) const
bool compute_here(const FunctionDAG::Node *f, bool tileable, int v, bool in_threads_loop, const Anderson2021Params &params, const Target &target)
int64_t get_total_local_mem_alloc_size(bool constant_allocs_only=false, bool in_threads_loop=false) const
void compute_working_set_from_features(int64_t *working_set, const StageMap< ScheduleFeatures > *features) const
int vectorized_access_size(size_t loop_index, bool verbose=false) const
Intrusive shared pointers have a reference count (a RefCount object) stored in the class itself.
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.
A class that can represent Vars or RVars.
Definition Func.h:29