* [TIR.Constant] U1 usecase
Constants are now aggregated into one struct and initialized in default_lib0.c
file
Change-Id: I34d61f8139c8a92c06944fe990ba892a660476fd
Unit test fixed
Change-Id: I436e7b6d6b3064b3f8bbfbb048d4296b63a6b69c
* Refactored
Addressed:
* PoolInfo splitted to WorkspacePoolInfo and ConstantPoolInfo
* workspace_byte_alignment moved to ExecutorCodegenMetadata
* getModuleAlignment -> GetModuleAlignment
* GenerateInternalWorkspaceBuffers refactored
* reverted format change of src/tir/transforms/legalize_packed_calls.cc
* addressed comments for src/tir/usmp/analysis/extract_buffer_info.cc
* removed commented code from include/tvm/tir/usmp/utils.h
Change-Id: I7d1b32884b0e5992e2e00c7838c85e425d9c25fd
* more unit test fixes
Change-Id: I573a05fa1cb4037ae83691f7dff2c2724b1d7700
* More refactoring and unit test fixes
Added ConstantMemoryPools
Change-Id: If1e391c631575980564bca790ba33748c82d907f
* bugfix
Change-Id: Iacc7a9d734a505dfa0d8d32d23ea3f57e6de8582
* refactoring. added constant_alignment
added constant_alignment
unit tests updated
Change-Id: I378193cb9e675e352c61d96ff4e09655090053e1
* unit-test bugix
Change-Id: Ia4411d59c4a376c01326fed366cdb196a432899e
* unit test fix
Change-Id: Ia2077bdeb1d2c6c9827eeef90ab410ae31b8c4a4
* Added support for c++ runtime
* refactored
* renamed pools and consts
renamed pools and consts to workspace_pools and constant_pools
* addressed upstream comments
* addressed upstream comments-2
* addressed upstream comments-3
* [TFLite] Support quantized EQUAL op in TFLite frontend
Support EQUAL quantization operation conversion as part of issue #9187
* [TFLite] Support quantized EQUAL op in TFLite frontend
Update elementwise quantized test for EQUAL op
Change-Id: I3897d1ac07051ebfc10356ad45397117b592f878
* changed x86/concat to use lists of ints instead of te.tensor.Tensor for loop extents and array offsets
* typos fixed
* removed unused import
* fixed micro model test
* fixed micro model test
The blob-embedding code creates a new LLVM module for which is needs more
information than just the target triple. The `_get_target_triple` function
in LLVMModule returned the triple with additional options appended to the
string. Instead of piggy-backing those extra options on top of the triple,
replace `_get_target_triple` with `_get_target_string`, which will return
the entire target string.
* [UPSTREAM][HEXAGON] Slice ops added - add, subtract, multiply
* Change to v68
* Change transform_numpy function call
* Do not disbale pylint errors and fix them
* Fix variable names
* Move the test file to topi
* Resolve conflict
* Modify init
Previously `cpu-flush` option existed as a boolean or integer argument, which is a bit counter-intuitive because for argparse, any non-empty string such as `False` will be parsed to `True` when using as a boolean and integer a little bit vague here IMHO. This PR used a function from `distutils` to directly parse input string to boolean, which makes the usage more stragiht-forward like `--cpu-flush True` or `--cpu-flush False`. Meanwhile it still supports usage of `0/1` and made sure the argument is always required.
* [TVMC] Fix tvmc run when using rpc
As described in #11707, the RPC mechanism does not support
objects of type Map which breaks the use of tvmc run when using
RPC after #9889. This commit intends to workaround this issue by
providing a fallback to the old implementation when RPC is being
used. Further, a test has been provided to help prevent this
regression in the future.
Change-Id: I70c1863d00098270e27c08ba834a3587e9132d69
* fix lint
Change-Id: I958cf4e19988d047bdd2e02f6475b9f70afe80c8
This PR added a utility function `GetAutoTensorizeMappingInfo` to propose mapping from workload block iters to the iters in the tensor intrin. An example usage is conv2d, where the computation block has more iters than the matmul tensor intrin.
* [Runtime][PipleineExecutor] Added Interface to Track Number of Global Inputs
Added a feature to PipelineExecutor to track number of Global Inputs.
* Fixed CI Error
* Fixed remaining CI Error
This PR includes the distributed measurement of tuning candidates using builder and async runner, as well as some auxiliary functions. It enables multiple builders and multiple runners with a tracker connecting in between. The hierarchy of files in the database can be further compacted to make the database more concise.
Previously only static shape computation is allowed in our tuning
system. However, one special case is overlooked: the reduction iter vars
could still have dynamic iteration domains which depend on other data
parallel vars. This PR rules out this case by carefully checking all the
loop extents during task extraction.
Related issue: https://github.com/apache/tvm/issues/11746.
* emptycommit 2nd try
* add operator and test
* example output
* lint with black
* register param index
* remove assert as it is a warning in torch
* fix algo bug
Co-authored-by: yuanfz <42092999+FZYUAN-1@users.noreply.github.com>
Minor modification to scoped timers to cover 99% of all the time cost during MS tuning. Allow `ApplyHistoryBest` and `TaskExtraction` time to be counted during tune_relay.
Follow-up of #11622, per discussion with @Kathryn-cat
- [x] Allow using a string `"default"` in `TuneContext` to quickly specify a set of target-specific
rules
- [x] Enhance detection of `ScheduleFn` in `TuneContext` to make it easier for users to quickly try
out template-driven scheduling on TIR.
Next PR:
- Add `TuneContext.tune` to allow directly tuning without task scheduler.
Co-Authored-By: Kathryn (Jinqi) Chen <65606304+Kathryn-cat@users.noreply.github.com>
Some python tests are failing in the wheel. This PR skips them if the environment variable `WHEEL_TEST` is set.
This PR is related to https://github.com/tlc-pack/tlcpack/pull/115.
* Implement avg_pool2d slice op
* Address review comments and fix the STIR schedule
* Fix formatting issues
* Address pylint errors
* Additional formatting issues
* more pylint fixes
* Changed arch version to v68 for now
* Changing arch version back to v69
* Move the test to tests/python/contrib/test_hexagon/topi
The tvmc command will throw a error while the passed path of model is not exist, But for PaddlePaddle model, it contains 2 file model_name.pdmodel and model_name.pdiparams, we only pass the prefix like inference_model/model_name.
This pr is same with https://github.com/apache/tvm/pull/11108
Since the origin PR didn't update for a long time, I send this new PR
This PR enables task filtering in Apply-History-Best, which is used in
Relay/Relax integration. Previously, even though a task is ruled out
during task extraction, it still shows up in Relay compilation due to
the lack of filtering on `Apply-History-Best`. However, TE-to-TIR
conversion `te.CreatePrimFunc` doesn't support all cases with hybrid
operators involved, which leads to post-tuning failure affecting
multiple models.
* - add support for host-driven AoT execution on zephyr;
- add initial version of reference counting to prevent python code from inadvertently freeing tensors during garbage collection;
- add support for numerical indices to host-drive AoT get_input();
- add two initial tests for host-driven AoT execution on zephyr;
- rename existing zephyr AoT exec. test;
* address PR feedback
* increase stack size to accommodate qemu_riscv64 stack usage
This PR adds `te.extern_primfunc` which provides the interface around TE ExternOp that allows a TVMScript defined schedulable TIR PrimFunc to be inlined into a TE compute graph. The result is that TIR can be used for compute definitions in Relay OpStrategies and, paired with meta-scheduler support in relay as introduced in #10578, these compute definitions can be scheduled and tuned as demonstrated in the attached tests.
Prior to this, compute definitions were limited to those definable in TE only. As a consequence of this patch and ongoing improvements to TVMScript meta-programming (#11097), TOPI can be extended to include compute and scheduling functions targeting schedulable TIR uniformly.
* [BYOC] Make CUTLASS BYOC integration 'Collage friendly'
(See https://discuss.tvm.apache.org/t/byoc-supporting-cutlass-byoc-with-collage/12796/6 for
context, which in turn is part of Collage (https://github.com/apache/tvm-rfcs/blob/main/rfcs/0062-collage.md).
Currently CUTLASS has four entry points:
- The usual 'partition_for_cutlass' partitioning function, using the
standard pattern table and pass machinery (see cutlass/build.py).
- A 'tune_cutlass_kernels' function which augments CUTLASS partition
functions with the results of building and running test kernels (see cutlass/build.py).
- A 'relay.ext.cutlass' external codegen function which inspects the
turning results and generates a CSourceModule for each partitions
(see cutlass/codegen.cc).
- A 'build_cutlass_kernels_vm' function which runs 'export_library' with
all the nvcc compiler options needed to build all the CSourceModules
(see cutlass/bild.py).
For Collage we'd like CUTLASS to have only two entry points: 'partition_for_cutlass',
and 'relay.ext.cutlass' or equivalent. This makes the CUTLASS external codegen integration
composable with other integrations, which in turn helps Collage avoid having to understand any
external codegen APIs other than the global pattern table and the custom compilation function/pass.
Collage also tends to end up requiring multiple partitions for the same backend since it is
more aggressive at mixing-and-matching smaller sub-graphs between backends. Thus we'd also like
to make sure all tuning, generated code and compilation overhead is shared between all such CUTLASS
partitions.
So, in this PR:
- We add all the CUTLASS-specific tuning and compilation options as new Target
attributes for the 'external codegen' "cutlass" TargetKind (cutlass/target.cc).
The user now has one place to provide those settings, and we've already done the
legwork to plumb the target instance.
- We replace 'relay.ext.cutlass' with a 'RelayToTIR' custom pass hook
'CompileForCutlass' (see cutlass/codegen.cc). This pass obviously can see all
the CUTLASS partitions in the IRModule, so we can now share tuning results
between them all and can be sure to generate a single CSourceModule. The pass can
also invoke the compiler to yield a StaticModule, which we've also already done the
legwork to support. In this way all CUTLASS-specific steps are handled at once.
- For convenience we supply 'finalize_modules' and 'finalize_modules_vm' which
invoke nvcc for final linking (using export_library as usual). However, there's now
nothing CUTLASS specific in those helpers other than their overriding of the 'compiler' to
be nvcc.
- test_cutlass.py is updated to use the new API.
Though this is a breaking change for existing users of the CUTLASS integration the
change is pretty minor, as shown in test_cutlass.py.
* - Masa's comments
* - Remove unnecessary save.
In order to build a dataset for improving the cost model for MetaSchedule, I added several files
including importing models to TVM, extracting tuning tasks, and sampling measure candidates.
Meanwhile, I exposed some methods in C++ to the Python side to assist the process.
* unifies all MKLDNN/DNNL_CODEGEN to DNNL
* translate -lib=mkldnn to -libs=dnnl in target
* type check added before
* rebase and update conv2d from mkldnn to dnnl
* Reuse hexagon launcher in test session
* separate random name generation
* revert get_aot_executor
* Fix launcher for simulator case
* add stop server for simulator
Dmitriy Smirnov
