Squashed commit

[Meta Schedule][M3c] Schedule Rules, Mutator & Postprocs (apache#485)

[Meta Schedule][M3c] PostOrderApply (apache#486)

Fix Post Order Apply (apache#490)

[MetaSchedule] Relay Integration (apache#489)

[M3c][Meta Schedule] Add Trace Correctness Test for PostOrderApply (apache#492)

Fix replay trace. (apache#493)

[M3c][Meta Schedule] Implement the Replay Func class. (apache#495)

[PR] Test script for meta-schedule task extraction. Interface to load… (apache#494)

[Meta Schedule Refactor] Get child blocks (apache#500)

Read-at && Write-at (apache#497)

[M3c][Meta Schedule] Measure Callbacks (apache#498)

[Bug] Fix Infinite Loop Caused When Calling Methods Not Overrided In PyClass (apache#496)

[MetaSchedule] Sample-Perfect-Tile (apache#501)

[MetaSchedule] TE Workloads (apache#502)

[TensorIR] GetProducer, GetConsumer (apache#506)

[MetaScheduleRefactor] Annotate&Unannotate (apache#505)

[MetaSchedule] Multi-Level-Tiling & Auto-Inline (apache#503)

[Tests] Add unittests for auto-inline and multi-level-tiling (apache#508)

[Meta Schedule] Minor Fixes (apache#507)

[MetaSchedule] Rewrite Cooperative-Fetching / Unbound-Block / Reduction-Block (apache#509)

[MetaSchedule] Rewrite Parallel-Vectorize-Unroll / Verify-GPU / Disallow-Dynamic-Loops (apache#499)

[Meta Schedule] Add Helper Function & Minor Modification (apache#512)

[MetaSchedule] Test for Rewrite Parallel-Vectorize-Unroll  (apache#513)

[Meta Schedule] Feature Extractor & Cost Model (apache#510)

Blockize & Tensorize (apache#514)

Layout Rewriting: Suggest-Index-Map (apache#520)

[MetaSchedule] Parallel-Vectorize-Unroll & Random-Compute-Location (apache#516)

[Meta Schedule] Per-Store-Feature (apache#521)

Add traced schedule for blockize & tensorize (apache#526)

[Meta Schedule] Add XGBoost Model & Random Model (apache#519)

User-Interface: Tune-TIR (apache#525)

User-Interface: Tune-TE (apache#527)

[Minor] More logging on python (apache#528)

Get CUDA tuning working (apache#529)

[MetaSchedule] TensorRT BYOC (apache#518)

[BugFix] LocalBuilder API (apache#531)

[Meta Schedule] Add Cost Model Update Measure Callback (apache#530)

[Bugfix] BuilderInput with default params (apache#532)

[MetaSchedule] Mutator-Tile-Size, Mutate-Parallel, Mutate-Unroll (apache#534)

[Meta Schedule] Evolutionary Search (apache#522)

[BugFix] Remove duplicated definition of MakeMultinomialSampler (apache#535)

[Meta Schedule] Fix some bugs (apache#537)

Initiate Experiments for CPU Performance Alignment with Ansor (apache#538)

[Meta Schedule] Tweak experiment scripts (apache#539)

[Meta Schedule] Initiate experiments on CUDA (apache#540)

[TIR][Schedule] Buffer transform (apache#523)

Auto Tensor Core (apache#524)

Working on Evo Search (apache#542)

[Meta Schedule] Add Replay Tuning Interface (apache#543)

Evolutionary Search on CPU (apache#544)

Misc improvement over the error message (apache#545)

[TIR][Schedule] Software pipelining (apache#533)

[Meta Schedule Refactor] fixing unit tests (apache#547)

[MetaSchedule] Mutator-Compute-Location (apache#548)

Misc Improvement of Evolutionary Search (apache#549)

Hotfix for software pipeline (apache#552)

Misc Improvement (apache#550)

[Cherry-Pick][TensorIR] Primitive "SetScope" (apache#9738) (apache#555)

Rule RFactor (apache#551)

[MemHammer] Rewrite Rules (apache#554)

[MetaSchedule] Schedule Rule: Cross-Thread Reduction (apache#556)

[MetaSchedule] Performance Alignment - NRM and SFM (CUDA) (apache#559)

[MetaSchedule] Perf Alignment - NRM on CUDA (apache#560)

[TIR] Reorder the block iters of the blocks generated by RFactor (apache#561)

Removing 2 unit tests for software pipelining (apache#562)

[MemHammer] Lower Pass + Unittests (apache#557)

Perf Align: Remove Auto-inline before Multi-level-tiling (apache#564)

Fix Sketch Generation Unittests (apache#565)

speed up VerifyGpuCode (apache#568)

[Performance Align] fixing codegen problems (apache#569)

[Meta schedule] improve search space (apache#1)

Hot fix for bound predicate (apache#3)

[Meta Schedule] Update Tune Relay (apache#4)

[Performance Align] fixing codegen problems (apache#5)

[PerfAlign] NRM & SFM on Raspi Aligned (apache#6)

[BugFix] Apply bound predicate directly to loops when possible (apache#12)

[BugFix] Fix CrossThreadReduction on CUDA (apache#13)

[MetaSchedule] Enable BertTuning with MetaScheduler (apache#11)

[Minor][MemHammer] Minor tweaks in code review (apache#14)

[Meta Schedule] Add customizable search space to PostOrderApply. (apache#16)

Fix cooperative fetching (apache#17)

Fixes for codegen (apache#18)

[Hotfix] A unittest (apache#19)

Fix for GRP sketch gen (apache#21)

Add threadIdx filtering in Multi-Level-Tiling and Verify-GPU-Code (apache#20)

[BugFix][TIR] Fix cross-thread reduction when single reduction loop with predicate (apache#10016) (apache#22)

[MemHammer][Refactor] Code Review (apache#15)

[Meta Schedule] Add Winograd Test for Customizable Search Space (apache#24)

Co-authored-by: Siyuan Feng <Hzfengsy@sjtu.edu.cn>
Co-authored-by: Bohan Hou <32121147+spectrometerHBH@users.noreply.github.com>
Co-authored-by: Hongyi Jin <3231950289@qq.com>
Co-authored-by: Ruihang Lai <lairuihangdongdong@qq.com>
Co-authored-by: Junru Shao <junrushao1994@gmail.com>
Co-authored-by: Wuwei Lin <wuwei@apache.org>
Co-authored-by: Sunghyun Park <49998730+sunggg@users.noreply.github.com>
Co-authored-by: Xiyou Zhou <xiyou@octoml.ai>

include/tvm/meta_schedule/database.h

@@ -237,6 +237,7 @@ class PyDatabaseNode : public DatabaseNode {
     // PackedFuncs are all not visited, because the reflection system doesn't take care of them,
     // so it cannot be accessible on the python side. If there is such need from the future,
     // we can then add corresponding accessor methods to help access on python.
+    //
     // `f_has_workload` is not visited
     // `f_commit_workload` is not visited
     // `f_commit_tuning_record` is not visited

include/tvm/meta_schedule/tune_context.h

@@ -53,7 +53,7 @@ class TuneContextNode : public runtime::Object {
   /*! \brief The probability of using certain mutator. */
   Map<Mutator, FloatImm> mutator_probs;
   /*! \brief The name of the tuning task. */
-  Optional<String> task_name;
+  String task_name;
   /*! \brief The random state. */
   support::LinearCongruentialEngine::TRandState rand_state;
   /*! \brief The number of threads to be used. */

include/tvm/tir/schedule/schedule.h

@@ -500,28 +500,28 @@ class ScheduleNode : public runtime::Object {
   /******** Schedule: Annotation ********/
   /*!
    * \brief Annotate a loop with a key value pair
-   * \param loop_rv The loop to be annotated
+   * \param loop The loop to be annotated
    * \param ann_key The annotation key
    * \param ann_val The annotation value, a string or a ExprRV
    */
   virtual void Annotate(const LoopRV& loop_rv, const String& ann_key, const ObjectRef& ann_val) = 0;
   /*!
    * \brief Annotate a block with a key value pair
-   * \param block_rv The block to be annotated
+   * \param loop The block to be annotated
    * \param ann_key The annotation key
    * \param ann_val The annotation value, a string or a ExprRV
    */
   virtual void Annotate(const BlockRV& block_rv, const String& ann_key,
                         const ObjectRef& ann_val) = 0;
   /*!
    * \brief Unannotate a loop's annotation with key ann_key
-   * \param loop_rv The loop to be unannotated
+   * \param loop The loop to be unannotated
    * \param ann_key The annotation key
    */
   virtual void Unannotate(const LoopRV& loop_rv, const String& ann_key) = 0;
   /*!
    * \brief Unannotate a block's annotation with key ann_key
-   * \param block_rv The block to be unannotated
+   * \param loop The block to be unannotated
    * \param ann_key The annotation key
    */
   virtual void Unannotate(const BlockRV& block_rv, const String& ann_key) = 0;

include/tvm/tir/stmt.h

@@ -1442,6 +1442,93 @@ constexpr const char* nested_software_pipeline_stage = "nested_software_pipeline
  */
 constexpr const char* nested_software_pipeline_order = "nested_software_pipeline_order";
 
+/*!
+ * \brief Mark that the block need to add predicate for block var bounds during lowering
+ */
+constexpr const char* require_block_var_bound_predicate = "require_bound_predicate";
+
+/*!
+ * \brief Mark that the loop should be further skip and bound to environment threads to enable
+ * cooperative fetching.
+ */
+constexpr const char* meta_schedule_cooperative_fetch = "meta_schedule.cooperative_fetch";
+
+/*!
+ * \brief Mark that the block should be further rewritten using tensorization.
+ */
+constexpr const char* meta_schedule_auto_tensorize = "meta_schedule.auto_tensorize";
+
+/*! \brief Mark that tensor core is enabled in the PrimExpr */
+constexpr const char* meta_schedule_tensor_core_enabled = "meta_schedule.tensor_core_enabled";
+
+/*! \brief The allowed range of thread extent in thread bindings */
+constexpr const char* meta_schedule_thread_extent_low_inclusive =
+    "meta_schedule.thread_extent_low_inclusive";
+
+/*! \brief The allowed range of thread extent in thread bindings */
+constexpr const char* meta_schedule_thread_extent_high_inclusive =
+    "meta_schedule.thread_extent_high_inclusive";
+
+/*!
+ * \brief Mark a block as generated by cache_read or cache_write block.
+ * 0 means cache_read; 1 means cache_write.
+ * \sa meta_schedule_cache_type_read
+ * \sa meta_schedule_cache_type_write
+ */
+constexpr const char* meta_schedule_cache_type = "meta_schedule.cache_type";
+
+/*! \sa meta_schedule_cache_type */
+constexpr const int meta_schedule_cache_type_read = 0;
+
+/*! \sa meta_schedule_cache_type */
+constexpr const int meta_schedule_cache_type_write = 1;
+
+/*! \brief Mark the tiling structure of blocks that are applied by rule Multi-Level-Tiling */
+constexpr const char* meta_schedule_tiling_structure = "meta_schedule.tiling_structure";
+
+/*! \brief Mark the block whose producer needs to be applied by rule Random-Compute-Location */
+constexpr const char* meta_schedule_random_compute_producer =
+    "meta_schedule.random_compute_producer";
+
+/*! \brief Mark auto-parallel setting on the block. */
+constexpr const char* meta_schedule_parallel = "meta_schedule.parallel";
+
+/*! \brief Mark auto-vectorize setting on the block. */
+constexpr const char* meta_schedule_vectorize = "meta_schedule.vectorize";
+
+/*! \brief Mark auto-unroll setting on the block. */
+constexpr const char* meta_schedule_unroll_explicit = "meta_schedule.unroll_explicit";
+
+/*! \brief Mark auto-unroll setting on the block. */
+constexpr const char* meta_schedule_unroll_implicit = "meta_schedule.unroll_implicit";
+
+/*! \brief Pragma: auto-unroll, max_step */
+constexpr const char* pragma_auto_unroll_max_step = "pragma_auto_unroll_max_step";
+
+/*! \brief Pragma: unroll explicit */
+constexpr const char* pragma_unroll_explicit = "pragma_unroll_explicit";
+
+/*! \brief Mark the scope of the software pipeline */
+constexpr const char* software_pipeline_scope = "software_pipeline_scope";
+
+/*! \brief Mark the stage of a statement in the software pipeline */
+constexpr const char* software_pipeline_stage = "software_pipeline_stage";
+
+/*! \brief Mark the order of a statement in the software pipeline */
+constexpr const char* software_pipeline_order = "software_pipeline_order";
+
+/*! \brief Mark the stage of the result of the software pipeline lowering. This is used to specify
+ * the behavior of nested software pipelines. Should be a 3-tuple consisting of the stage of the
+ * prologue, the body, and the epilogue of the software pipeline.
+ */
+constexpr const char* nested_software_pipeline_stage = "nested_software_pipeline_stage";
+
+/*! \brief Mark the stage of the result of the software pipeline lowering. This is used to specify
+ * the behavior of nested software pipelines. Should be a 3-tuple consisting of the stage of the
+ * prologue, the body, and the epilogue of the software pipeline.
+ */
+constexpr const char* nested_software_pipeline_order = "nested_software_pipeline_order";
+
 /*!
  * \brief Check if attr_key is a pragma key extension
  * \param attr_key The attr key to be compared

include/tvm/tir/transform.h

@@ -383,6 +383,20 @@ TVM_DLL Pass LowerInitBlock();
  */
 TVM_DLL Pass PlanAndUpdateBufferAllocationLocation();
 
+/*!
+ * \brief Narrow the extents of some loops by checking whether some constraints in the block iter
+ * bound predicates can be directly applied on the loops.
+ * \return The pass.
+ */
+TVM_DLL Pass ApplyBlockBoundPredicate();
+
+/*!
+ * \brief Narrow the extents of some loops by checking whether some constraints in the block iter
+ * bound predicates can be directly applied on the loops.
+ * \return The pass.
+ */
+TVM_DLL Pass ApplyBlockBoundPredicate();
+
 /*!
  * \brief Substitute all the block vars with the PrimExprs they are bound to, indicated by the
  *        corresponding iter_values in BlockRealize, for opaque blocks by removing all

python/tvm/auto_scheduler/search_task.py

@@ -543,7 +543,8 @@ def print_best(self, log_file, print_mode="schedule"):
         code: str
             The best schedule code in python API or CUDA source code
         """
-        inp, _ = load_best_record(log_file, self.workload_key)
+        inp, res = load_best_record(log_file, self.workload_key)
+        print("Best codes (ms):", [float(c) * 1000.0 for c in res.costs])
         if inp is None:
             raise RuntimeError(
                 "Cannot find any valid schedule for %s in file %s" % (self.workload_key, log_file)

python/tvm/auto_scheduler/workload_registry.py

@@ -194,7 +194,10 @@ def workload_key_to_tensors(workload_key):
     assert callable(value)
 
     args = deserialize_args(workload[1:])
-    return value(*args)
+    result = value(*args)
+    if isinstance(result, tuple):
+        result = list(result)
+    return result
 
 
 def serialize_workload_registry_entry(workload_key):

python/tvm/meta_schedule/builder/local_builder.py

@@ -22,13 +22,28 @@
 
 from tvm._ffi import register_func
 from tvm.ir import IRModule
-from tvm.runtime import Module, NDArray, load_param_dict, save_param_dict
+from tvm.runtime import NDArray
+from tvm.runtime import Module, load_param_dict, save_param_dict
 from tvm.target import Target
 
 from ...contrib.popen_pool import MapResult, PopenPoolExecutor, StatusKind
 from ..utils import cpu_count, get_global_func_with_default_on_worker
 from .builder import BuilderInput, BuilderResult, PyBuilder
 
+logger = logging.getLogger(__name__)
+
+
+def _serialize_params(params: Optional[Dict[str, NDArray]]) -> Optional[bytearray]:
+    if params is None:
+        return None
+    return save_param_dict(params)
+
+
+def _deserialize_params(params: Optional[bytearray]) -> Optional[Dict[str, NDArray]]:
+    if params is None:
+        return None
+    return load_param_dict(params)
+
 
 logger = logging.getLogger(__name__)  # pylint: disable=invalid-name
 
@@ -127,7 +142,6 @@ def __init__(
             The initializer to be used for the worker processes.
         """
         super().__init__()
-
         if max_workers is None:
             max_workers = cpu_count(logical=True)
         logger.info("LocalBuilder: max_workers = %d", max_workers)

python/tvm/meta_schedule/cost_model/cost_model.py

@@ -15,17 +15,19 @@
 # specific language governing permissions and limitations
 # under the License.
 """Meta Schedule CostModel."""
-import ctypes
+
 from typing import List
+import ctypes
+
+import numpy as np
 
-import numpy as np  # type: ignore
 from tvm._ffi import register_object
 from tvm.runtime import Object
 
 from .. import _ffi_api
 from ..runner import RunnerResult
-from ..search_strategy import MeasureCandidate
 from ..tune_context import TuneContext
+from ..search_strategy import MeasureCandidate
 from ..utils import _get_hex_address, check_override
 
 

python/tvm/meta_schedule/cost_model/metric.py

@@ -15,10 +15,11 @@
 # specific language governing permissions and limitations
 # under the License.
 """Cost model metrics for meta schedule"""
-import numpy as np  # type: ignore
+from typing import List
+import numpy as np
 
 
-def max_curve(trial_scores: np.ndarray) -> np.ndarray:
+def max_curve(trial_scores: np.ndarray) -> List[float]:
     """f(n) = max([s[i] fo i < n])
 
     Parameters
@@ -28,8 +29,8 @@ def max_curve(trial_scores: np.ndarray) -> np.ndarray:
 
     Returns
     -------
-    curve : np.ndarray
-        A vector, the max-curve function values
+    curve : List[float]
+        function values
     """
     ret = np.empty(len(trial_scores))
     keep = -1e9

python/tvm/meta_schedule/cost_model/random_model.py

@@ -17,14 +17,14 @@
 """
 Random cost model
 """
-from typing import List, Optional, Tuple, Union
+from typing import List, Union, Tuple, Optional
 
-import numpy as np  # type: ignore
+import numpy as np
 
-from ..cost_model import PyCostModel
 from ..runner import RunnerResult
-from ..search_strategy import MeasureCandidate
 from ..tune_context import TuneContext
+from ..search_strategy import MeasureCandidate
+from ..cost_model import PyCostModel
 
 
 class RandomModel(PyCostModel):
@@ -70,7 +70,7 @@ def load(self, path: str) -> None:
         path : str
             The file path.
         """
-        self.random_state = tuple(np.load(path, allow_pickle=True))  # type: ignore
+        self.random_state = tuple(np.load(path, allow_pickle=True))
 
     def save(self, path: str) -> None:
         """Save the cost model to given file location.
@@ -116,7 +116,7 @@ def predict(self, context: TuneContext, candidates: List[MeasureCandidate]) -> n
             The predicted running results.
         """
         np.random.set_state(self.random_state)
-        # TODO(@zxybazh): Use numpy's RandState object:
+        # todo(@zxybazh): Use numpy's RandState object:
         # https://numpy.org/doc/1.16/reference/generated/numpy.random.RandomState.html#numpy.random.RandomState
         result = np.random.rand(len(candidates)) * self.max_range
         self.random_state = np.random.get_state()

python/tvm/meta_schedule/feature_extractor/random_feature_extractor.py

@@ -17,7 +17,7 @@
 """Random Feature Extractor."""
 from typing import List, Union, Tuple
 
-import numpy as np  # type: ignore
+import numpy as np
 from tvm.runtime.ndarray import NDArray, array
 
 from ..tune_context import TuneContext

python/tvm/meta_schedule/runner/local_runner.py

@@ -33,7 +33,7 @@
     run_evaluator_common,
 )
 
-logger = logging.getLogger(__name__)  # pylint: disable=invalid-name
+logger = logging.getLogger(__name__)
 
 
 class LocalRunnerFuture(RunnerFuture):

python/tvm/meta_schedule/space_generator/post_order_apply.py

@@ -32,5 +32,5 @@ class PostOrderApply(SpaceGenerator):
     def __init__(self):
         """Constructor"""
         self.__init_handle_by_constructor__(
-            _ffi_api.SpaceGeneratorPostOrderApply,  # type: ignore # pylint: disable=no-member
+            _ffi_api.SpaceGeneratorPostOrderApply,  # pylint: disable=no-member
         )