nvfuser-cu121-torch25 0.2.25.dev20250201__cp312-cp312-manylinux_2_28_x86_64.whl
Sign up to get free protection for your applications and to get access to all the features.
- nvfuser/_C.cpython-312-x86_64-linux-gnu.so +0 -0
- nvfuser/__init__.py +618 -0
- nvfuser/__init__.pyi +4 -0
- nvfuser/contrib/__init__.py +9 -0
- nvfuser/contrib/nn/__init__.py +13 -0
- nvfuser/contrib/nn/normalization.py +725 -0
- nvfuser/include/nvfuser/alias_analysis.h +116 -0
- nvfuser/include/nvfuser/bfs.h +929 -0
- nvfuser/include/nvfuser/codegen.h +26 -0
- nvfuser/include/nvfuser/compute_at.h +28 -0
- nvfuser/include/nvfuser/compute_at_map.h +394 -0
- nvfuser/include/nvfuser/contiguity.h +351 -0
- nvfuser/include/nvfuser/cuda_utils.h +50 -0
- nvfuser/include/nvfuser/debug.h +50 -0
- nvfuser/include/nvfuser/device_lower/analysis/bank_conflict.h +53 -0
- nvfuser/include/nvfuser/device_lower/analysis/circular_buffer.h +109 -0
- nvfuser/include/nvfuser/device_lower/analysis/device_version.h +65 -0
- nvfuser/include/nvfuser/device_lower/analysis/divisible_split.h +28 -0
- nvfuser/include/nvfuser/device_lower/analysis/fused_reduction.h +36 -0
- nvfuser/include/nvfuser/device_lower/analysis/index_compute.h +322 -0
- nvfuser/include/nvfuser/device_lower/analysis/predicate_elimination.h +71 -0
- nvfuser/include/nvfuser/device_lower/analysis/sync_information.h +47 -0
- nvfuser/include/nvfuser/device_lower/analysis/tensor_memory.h +65 -0
- nvfuser/include/nvfuser/device_lower/analysis/thread_predicate.h +158 -0
- nvfuser/include/nvfuser/device_lower/analysis/tma.h +93 -0
- nvfuser/include/nvfuser/device_lower/analysis/trivial_broadcast.h +75 -0
- nvfuser/include/nvfuser/device_lower/id_model_options.h +135 -0
- nvfuser/include/nvfuser/device_lower/lower2device.h +391 -0
- nvfuser/include/nvfuser/device_lower/pass/alias_memory.h +37 -0
- nvfuser/include/nvfuser/device_lower/pass/allocation.h +32 -0
- nvfuser/include/nvfuser/device_lower/pass/circular_buffer.h +191 -0
- nvfuser/include/nvfuser/device_lower/pass/expr_sort.h +17 -0
- nvfuser/include/nvfuser/device_lower/pass/fusion_simplifier.h +21 -0
- nvfuser/include/nvfuser/device_lower/pass/grid_serialization.h +26 -0
- nvfuser/include/nvfuser/device_lower/pass/index.h +200 -0
- nvfuser/include/nvfuser/device_lower/pass/inline_ptx.h +16 -0
- nvfuser/include/nvfuser/device_lower/pass/insert_syncs.h +39 -0
- nvfuser/include/nvfuser/device_lower/pass/instrument.h +24 -0
- nvfuser/include/nvfuser/device_lower/pass/loop_rotation.h +150 -0
- nvfuser/include/nvfuser/device_lower/pass/loops.h +68 -0
- nvfuser/include/nvfuser/device_lower/pass/magic_zero.h +86 -0
- nvfuser/include/nvfuser/device_lower/pass/misaligned_vectorization.h +118 -0
- nvfuser/include/nvfuser/device_lower/pass/predicate.h +23 -0
- nvfuser/include/nvfuser/device_lower/pass/replace_size.h +24 -0
- nvfuser/include/nvfuser/device_lower/pass/scalar_hoist.h +115 -0
- nvfuser/include/nvfuser/device_lower/pass/unroll.h +98 -0
- nvfuser/include/nvfuser/device_lower/pass/vectorize_welford.h +45 -0
- nvfuser/include/nvfuser/device_lower/pass/warp_reduce.h +23 -0
- nvfuser/include/nvfuser/device_lower/utils.h +382 -0
- nvfuser/include/nvfuser/device_lower/validation.h +74 -0
- nvfuser/include/nvfuser/disjoint_set.h +556 -0
- nvfuser/include/nvfuser/dispatch.h +334 -0
- nvfuser/include/nvfuser/driver_api.h +49 -0
- nvfuser/include/nvfuser/dynamic_transform.h +316 -0
- nvfuser/include/nvfuser/dynamic_type/C++20/type_traits +37 -0
- nvfuser/include/nvfuser/dynamic_type/dynamic_type.h +969 -0
- nvfuser/include/nvfuser/dynamic_type/error.h +24 -0
- nvfuser/include/nvfuser/dynamic_type/type_traits.h +703 -0
- nvfuser/include/nvfuser/evaluator_common.h +295 -0
- nvfuser/include/nvfuser/exceptions.h +283 -0
- nvfuser/include/nvfuser/expr_evaluator.h +125 -0
- nvfuser/include/nvfuser/expr_simplifier.h +218 -0
- nvfuser/include/nvfuser/flatbuffers/allocator.h +68 -0
- nvfuser/include/nvfuser/flatbuffers/array.h +253 -0
- nvfuser/include/nvfuser/flatbuffers/base.h +486 -0
- nvfuser/include/nvfuser/flatbuffers/buffer.h +154 -0
- nvfuser/include/nvfuser/flatbuffers/buffer_ref.h +53 -0
- nvfuser/include/nvfuser/flatbuffers/code_generator.h +80 -0
- nvfuser/include/nvfuser/flatbuffers/code_generators.h +234 -0
- nvfuser/include/nvfuser/flatbuffers/default_allocator.h +64 -0
- nvfuser/include/nvfuser/flatbuffers/detached_buffer.h +114 -0
- nvfuser/include/nvfuser/flatbuffers/flatbuffer_builder.h +1225 -0
- nvfuser/include/nvfuser/flatbuffers/flatbuffers.h +272 -0
- nvfuser/include/nvfuser/flatbuffers/flatc.h +130 -0
- nvfuser/include/nvfuser/flatbuffers/flex_flat_util.h +36 -0
- nvfuser/include/nvfuser/flatbuffers/flexbuffers.h +1889 -0
- nvfuser/include/nvfuser/flatbuffers/grpc.h +300 -0
- nvfuser/include/nvfuser/flatbuffers/hash.h +127 -0
- nvfuser/include/nvfuser/flatbuffers/idl.h +1359 -0
- nvfuser/include/nvfuser/flatbuffers/minireflect.h +420 -0
- nvfuser/include/nvfuser/flatbuffers/reflection.h +522 -0
- nvfuser/include/nvfuser/flatbuffers/reflection_generated.h +1471 -0
- nvfuser/include/nvfuser/flatbuffers/registry.h +128 -0
- nvfuser/include/nvfuser/flatbuffers/stl_emulation.h +513 -0
- nvfuser/include/nvfuser/flatbuffers/string.h +64 -0
- nvfuser/include/nvfuser/flatbuffers/struct.h +53 -0
- nvfuser/include/nvfuser/flatbuffers/table.h +168 -0
- nvfuser/include/nvfuser/flatbuffers/util.h +731 -0
- nvfuser/include/nvfuser/flatbuffers/vector.h +393 -0
- nvfuser/include/nvfuser/flatbuffers/vector_downward.h +273 -0
- nvfuser/include/nvfuser/flatbuffers/verifier.h +317 -0
- nvfuser/include/nvfuser/fusion.h +511 -0
- nvfuser/include/nvfuser/fusion_guard.h +37 -0
- nvfuser/include/nvfuser/fusion_profiler.h +311 -0
- nvfuser/include/nvfuser/fusion_segmenter.h +751 -0
- nvfuser/include/nvfuser/global_allocator.h +27 -0
- nvfuser/include/nvfuser/grouped_reduction.h +47 -0
- nvfuser/include/nvfuser/host_ir/container.h +60 -0
- nvfuser/include/nvfuser/host_ir/executor.h +152 -0
- nvfuser/include/nvfuser/host_ir/host_ir.h +320 -0
- nvfuser/include/nvfuser/host_ir/lower.h +35 -0
- nvfuser/include/nvfuser/id_model/circular_buffer_indexing.h +56 -0
- nvfuser/include/nvfuser/id_model/contiguity.h +166 -0
- nvfuser/include/nvfuser/id_model/id_model.h +359 -0
- nvfuser/include/nvfuser/id_model/id_model_index_compute.h +81 -0
- nvfuser/include/nvfuser/id_model/indexing.h +208 -0
- nvfuser/include/nvfuser/id_model/indexing_traversal.h +72 -0
- nvfuser/include/nvfuser/id_model/indexing_utils.h +62 -0
- nvfuser/include/nvfuser/id_model/loop_promotion.h +180 -0
- nvfuser/include/nvfuser/id_model/predicate_indexing.h +104 -0
- nvfuser/include/nvfuser/id_model/schedule.h +54 -0
- nvfuser/include/nvfuser/id_model/to_string.h +87 -0
- nvfuser/include/nvfuser/id_model/transform_replay.h +58 -0
- nvfuser/include/nvfuser/id_model/utils.h +176 -0
- nvfuser/include/nvfuser/id_model/validation_utils.h +55 -0
- nvfuser/include/nvfuser/index_compute.h +651 -0
- nvfuser/include/nvfuser/instrumentation.h +107 -0
- nvfuser/include/nvfuser/ir/all_nodes.h +14 -0
- nvfuser/include/nvfuser/ir/base_nodes.h +687 -0
- nvfuser/include/nvfuser/ir/builder.h +215 -0
- nvfuser/include/nvfuser/ir/builder_passkey.h +29 -0
- nvfuser/include/nvfuser/ir/cloner.h +185 -0
- nvfuser/include/nvfuser/ir/container.h +226 -0
- nvfuser/include/nvfuser/ir/graphviz.h +119 -0
- nvfuser/include/nvfuser/ir/interface_nodes.h +957 -0
- nvfuser/include/nvfuser/ir/internal_base_nodes.h +744 -0
- nvfuser/include/nvfuser/ir/internal_nodes.h +2792 -0
- nvfuser/include/nvfuser/ir/iostream.h +98 -0
- nvfuser/include/nvfuser/ir/printer.h +57 -0
- nvfuser/include/nvfuser/ir/utils.h +801 -0
- nvfuser/include/nvfuser/iter_visitor.h +661 -0
- nvfuser/include/nvfuser/kernel.h +299 -0
- nvfuser/include/nvfuser/kernel_db/kernel_db.h +109 -0
- nvfuser/include/nvfuser/kernel_db/utils.h +37 -0
- nvfuser/include/nvfuser/kernel_ir.h +1457 -0
- nvfuser/include/nvfuser/kernel_ir_dispatch.h +147 -0
- nvfuser/include/nvfuser/linked_hash_map.h +97 -0
- nvfuser/include/nvfuser/logical_domain_map.h +577 -0
- nvfuser/include/nvfuser/macros.h +23 -0
- nvfuser/include/nvfuser/mma_type.h +257 -0
- nvfuser/include/nvfuser/multidevice/c10d_mock.h +175 -0
- nvfuser/include/nvfuser/multidevice/communication.h +232 -0
- nvfuser/include/nvfuser/multidevice/communicator.h +179 -0
- nvfuser/include/nvfuser/multidevice/device_mesh.h +95 -0
- nvfuser/include/nvfuser/multidevice/executor.h +107 -0
- nvfuser/include/nvfuser/multidevice/multidevice.h +18 -0
- nvfuser/include/nvfuser/multidevice/utils.h +187 -0
- nvfuser/include/nvfuser/non_divisible_split.h +86 -0
- nvfuser/include/nvfuser/opaque_type.h +129 -0
- nvfuser/include/nvfuser/ops/alias.h +192 -0
- nvfuser/include/nvfuser/ops/all_ops.h +13 -0
- nvfuser/include/nvfuser/ops/arith.h +712 -0
- nvfuser/include/nvfuser/ops/composite.h +130 -0
- nvfuser/include/nvfuser/ops/indexing.h +55 -0
- nvfuser/include/nvfuser/ops/normalization.h +263 -0
- nvfuser/include/nvfuser/ops/utils.h +127 -0
- nvfuser/include/nvfuser/options.h +313 -0
- nvfuser/include/nvfuser/parallel_dimension_map.h +95 -0
- nvfuser/include/nvfuser/parallel_type_bitmap.h +365 -0
- nvfuser/include/nvfuser/polymorphic_value.h +432 -0
- nvfuser/include/nvfuser/predicate_compute.h +213 -0
- nvfuser/include/nvfuser/python_frontend/distributed_tensor.h +50 -0
- nvfuser/include/nvfuser/python_frontend/fusion_cache.h +298 -0
- nvfuser/include/nvfuser/python_frontend/fusion_definition.h +372 -0
- nvfuser/include/nvfuser/python_frontend/fusion_record.h +3124 -0
- nvfuser/include/nvfuser/python_frontend/fusion_state.h +143 -0
- nvfuser/include/nvfuser/python_frontend/python_bindings.h +27 -0
- nvfuser/include/nvfuser/python_frontend/segmentation.h +246 -0
- nvfuser/include/nvfuser/python_frontend/translation.h +20 -0
- nvfuser/include/nvfuser/python_frontend/translation_utils.h +308 -0
- nvfuser/include/nvfuser/scheduler/all_schedulers.h +17 -0
- nvfuser/include/nvfuser/scheduler/ampere_multi_matmul.h +206 -0
- nvfuser/include/nvfuser/scheduler/cache_policy_refiner.h +19 -0
- nvfuser/include/nvfuser/scheduler/compile_time_info.h +322 -0
- nvfuser/include/nvfuser/scheduler/debug_utils.h +68 -0
- nvfuser/include/nvfuser/scheduler/expr_eval_sched.h +45 -0
- nvfuser/include/nvfuser/scheduler/heuristic.h +113 -0
- nvfuser/include/nvfuser/scheduler/hopper_multi_matmul.h +204 -0
- nvfuser/include/nvfuser/scheduler/mark_aliases.h +19 -0
- nvfuser/include/nvfuser/scheduler/matmul.h +40 -0
- nvfuser/include/nvfuser/scheduler/matmul_heuristic.h +293 -0
- nvfuser/include/nvfuser/scheduler/matmul_heuristic_plugin.h +65 -0
- nvfuser/include/nvfuser/scheduler/matmul_heuristic_plugin_api.h +99 -0
- nvfuser/include/nvfuser/scheduler/matmul_utils.h +54 -0
- nvfuser/include/nvfuser/scheduler/mma_utils.h +500 -0
- nvfuser/include/nvfuser/scheduler/multi_matmul.h +74 -0
- nvfuser/include/nvfuser/scheduler/no_op.h +48 -0
- nvfuser/include/nvfuser/scheduler/normalization_inner.h +49 -0
- nvfuser/include/nvfuser/scheduler/normalization_inner_outer.h +51 -0
- nvfuser/include/nvfuser/scheduler/normalization_outer.h +48 -0
- nvfuser/include/nvfuser/scheduler/normalization_utils.h +379 -0
- nvfuser/include/nvfuser/scheduler/pointwise.h +183 -0
- nvfuser/include/nvfuser/scheduler/pointwise_heuristic.h +118 -0
- nvfuser/include/nvfuser/scheduler/pointwise_utils.h +24 -0
- nvfuser/include/nvfuser/scheduler/reduction.h +43 -0
- nvfuser/include/nvfuser/scheduler/reduction_heuristic.h +339 -0
- nvfuser/include/nvfuser/scheduler/reduction_utils.h +159 -0
- nvfuser/include/nvfuser/scheduler/registry.h +97 -0
- nvfuser/include/nvfuser/scheduler/registry_utils.h +111 -0
- nvfuser/include/nvfuser/scheduler/resize.h +41 -0
- nvfuser/include/nvfuser/scheduler/resize_heuristic.h +67 -0
- nvfuser/include/nvfuser/scheduler/runtime_info.h +166 -0
- nvfuser/include/nvfuser/scheduler/scheduler_types.h +80 -0
- nvfuser/include/nvfuser/scheduler/transpose.h +114 -0
- nvfuser/include/nvfuser/scheduler/transpose_heuristic.h +164 -0
- nvfuser/include/nvfuser/scheduler/utils.h +771 -0
- nvfuser/include/nvfuser/scheduler/vectorize_helper.h +349 -0
- nvfuser/include/nvfuser/serde/factory.h +55 -0
- nvfuser/include/nvfuser/serde/fusion_cache_generated.h +4319 -0
- nvfuser/include/nvfuser/serde/fusion_record.h +124 -0
- nvfuser/include/nvfuser/serde/polymorphic_value.h +52 -0
- nvfuser/include/nvfuser/serde/utils.h +34 -0
- nvfuser/include/nvfuser/struct.inl +127 -0
- nvfuser/include/nvfuser/swizzle.h +54 -0
- nvfuser/include/nvfuser/sys_utils.h +40 -0
- nvfuser/include/nvfuser/tensor_metadata.h +118 -0
- nvfuser/include/nvfuser/tma.h +124 -0
- nvfuser/include/nvfuser/transform_iter.h +522 -0
- nvfuser/include/nvfuser/transform_replay.h +297 -0
- nvfuser/include/nvfuser/transform_rfactor.h +33 -0
- nvfuser/include/nvfuser/transform_view.h +136 -0
- nvfuser/include/nvfuser/type.h +1125 -0
- nvfuser/include/nvfuser/type_promotion.h +61 -0
- nvfuser/include/nvfuser/utils.h +619 -0
- nvfuser/include/nvfuser/val_graph.h +446 -0
- nvfuser/include/nvfuser/val_graph_visitor.h +259 -0
- nvfuser/include/nvfuser/validator_utils.h +92 -0
- nvfuser/include/nvfuser/vectorization_info.h +31 -0
- nvfuser/include/nvfuser/visibility.h +21 -0
- nvfuser/lib/libnvfuser_codegen.so +0 -0
- nvfuser/nvfuser_version.py +69 -0
- nvfuser/pytorch_utils.py +184 -0
- nvfuser/share/cmake/nvfuser/NvfuserConfig-release.cmake +20 -0
- nvfuser/share/cmake/nvfuser/NvfuserConfig.cmake +106 -0
- nvfuser/utils.py +18 -0
- nvfuser/version.py +1 -0
- nvfuser_cu121_torch25-0.2.25.dev20250201.dist-info/LICENSE +976 -0
- nvfuser_cu121_torch25-0.2.25.dev20250201.dist-info/METADATA +16 -0
- nvfuser_cu121_torch25-0.2.25.dev20250201.dist-info/RECORD +242 -0
- nvfuser_cu121_torch25-0.2.25.dev20250201.dist-info/WHEEL +5 -0
- nvfuser_cu121_torch25-0.2.25.dev20250201.dist-info/top_level.txt +1 -0
- nvfuser_cu121_torch25.libs/libnvToolsExt-847d78f2.so.1.0.0 +0 -0
@@ -0,0 +1,116 @@
|
|
1
|
+
// clang-format off
|
2
|
+
/*
|
3
|
+
* SPDX-FileCopyrightText: Copyright (c) 2023-present NVIDIA CORPORATION & AFFILIATES.
|
4
|
+
* All rights reserved.
|
5
|
+
* SPDX-License-Identifier: BSD-3-Clause
|
6
|
+
*/
|
7
|
+
// clang-format on
|
8
|
+
#pragma once
|
9
|
+
|
10
|
+
#include <string>
|
11
|
+
#include <unordered_map>
|
12
|
+
#include <vector>
|
13
|
+
|
14
|
+
#include <fusion.h>
|
15
|
+
#include <ir/interface_nodes.h>
|
16
|
+
|
17
|
+
namespace nvfuser {
|
18
|
+
|
19
|
+
struct Layout {
|
20
|
+
std::vector<IterDomain*> allocation_domain;
|
21
|
+
std::vector<std::optional<bool>> contiguity;
|
22
|
+
|
23
|
+
// The size of `allocation_domain` and therefore the size of `contiguity`.
|
24
|
+
int64_t size() const;
|
25
|
+
|
26
|
+
std::string toString(int indent_size = 0) const;
|
27
|
+
|
28
|
+
// Returns whether this layout is compliant with `required`. This is
|
29
|
+
// uni-directional. For example, `contiguity=[t,t]` is compliant with
|
30
|
+
// `contiguity=[f,f]` but not vice versa. As a special case,
|
31
|
+
// an empty `required.allocation` indicates no requirements, i.e., the method
|
32
|
+
// always returns true.
|
33
|
+
bool isCompliantWith(const Layout& required) const;
|
34
|
+
};
|
35
|
+
|
36
|
+
// Holds aliases found in a fusion. The expected user flow is
|
37
|
+
//
|
38
|
+
// ```
|
39
|
+
// AliasAnalysisResult analysis;
|
40
|
+
// analysis.add(...);
|
41
|
+
// ...
|
42
|
+
// analysis.add(...);
|
43
|
+
// analysis.finalize(...);
|
44
|
+
//
|
45
|
+
// // The user can now call const methods to retrieve information.
|
46
|
+
// ```
|
47
|
+
class AliasAnalysisResult {
|
48
|
+
public:
|
49
|
+
AliasAnalysisResult() = default;
|
50
|
+
AliasAnalysisResult(const AliasAnalysisResult&) = delete;
|
51
|
+
AliasAnalysisResult& operator=(const AliasAnalysisResult&) = delete;
|
52
|
+
AliasAnalysisResult(AliasAnalysisResult&&) = default;
|
53
|
+
AliasAnalysisResult& operator=(AliasAnalysisResult&&) = default;
|
54
|
+
|
55
|
+
// Marks `source` as the immediate aliasing source of `alias` and sets the
|
56
|
+
// preferred layout.
|
57
|
+
void add(const TensorView* alias, TensorView* source, Layout&& layout);
|
58
|
+
|
59
|
+
// Computes transitive aliases and caches them in `alias_to_root_`.
|
60
|
+
// See `findAliases` for the meaning of
|
61
|
+
// `can_override_empty_allocation_domain`.
|
62
|
+
void finalize(bool can_override_empty_allocation_domain);
|
63
|
+
|
64
|
+
// Returns the preferred layout. If `alias` is not in `alias_to_source_`,
|
65
|
+
// returns the `TensorView`'s initial layout.
|
66
|
+
Layout preferredLayout(const Val* alias) const;
|
67
|
+
|
68
|
+
std::string toString(int indent_size = 0) const;
|
69
|
+
|
70
|
+
// Returns the mapped value in `alias_to_root_` or null.
|
71
|
+
TensorView* getRoot(const TensorView* alias) const;
|
72
|
+
|
73
|
+
private:
|
74
|
+
// Maps an alias (e.g. the output of a `ViewOp`) to its direct source (e.g.
|
75
|
+
// the input of the same `ViewOp`). Also stores the preferred output layout
|
76
|
+
// for the alias. Consider path compression, a common optimization used in
|
77
|
+
// disjoint-set data structure, so it's easy to figure out the root of an
|
78
|
+
// alias.
|
79
|
+
std::unordered_map<const TensorView*, std::pair<TensorView*, Layout>>
|
80
|
+
alias_to_source_;
|
81
|
+
|
82
|
+
// Maps an alias to its "highest ancestor" according to `alias_to_source_`,
|
83
|
+
// if its preferred layout is compliant with its actual layout.
|
84
|
+
std::unordered_map<const TensorView*, TensorView*> alias_to_root_;
|
85
|
+
};
|
86
|
+
|
87
|
+
// Finds aliases of the fusion inputs. The analysis should be conservative --
|
88
|
+
// when the analysis says B is an alias of input A and that B's preferred layout
|
89
|
+
// is compliant with the required layout, `ExpressionEvaluator::evaluate(B)`
|
90
|
+
// should produce an `at::Tensor` that's an alias of the `at::Tensor` bound to
|
91
|
+
// A.
|
92
|
+
//
|
93
|
+
// [Note on overriding empty allocation domains]
|
94
|
+
//
|
95
|
+
// We can override an empty allocation domain to any layout before segmentation
|
96
|
+
// but not after. For example,
|
97
|
+
// ```
|
98
|
+
// auto in = makeContigConcreteTensor({2, 3});
|
99
|
+
// auto slice_out = segment_set(slice(in, {0, 0}, {2, 2}));
|
100
|
+
// auto add_out = add(slice_out, slice_out)
|
101
|
+
// ```
|
102
|
+
// will be split at `slice_out` into two segments. `slice_out`'s contiguity
|
103
|
+
// needs to be [f,t] so it can be made an alias. If we were to let a scheduler
|
104
|
+
// (thus after segmentation) change its contiguity, we would have to change the
|
105
|
+
// input contiguity of the second segment as well. This is possible but hard to
|
106
|
+
// implement given the current infrastructure.
|
107
|
+
//
|
108
|
+
// Therefore, I chose to run alias analysis both before segmentation and in
|
109
|
+
// schedulers. The former, used by MarkAliasesPreparePass, updates layouts to
|
110
|
+
// enable aliases; the latter, used by NoOpScheduler, calls
|
111
|
+
// Fusion::aliasOutputToInput to mark aliases.
|
112
|
+
AliasAnalysisResult findAliases(
|
113
|
+
Fusion* fusion,
|
114
|
+
bool can_override_empty_allocation_domain = true);
|
115
|
+
|
116
|
+
} // namespace nvfuser
|