[Driver] Single-module lowering flow in driver_api.cc #14985
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment.
Generated by tvm-bot |
Lunderberg
force-pushed
the
single_module_during_build
branch
from
e459552 to
e7bfa2d
Compare
Lunderberg
added a commit
to Lunderberg/tvm
that referenced
this pull request
May 30, 2023
The functionality tested in this commit was added across several recent PRs, each of which tested their features in isolation. This PR adds unit tests to validate the end-to-end behavior of TIR subroutine calls. PRs building up to this point: - TVMScript - apache#14889 - apache#14915 - apache#14919 - apache#14941 - Functionality improvements of existing TIR passes - apache#14913 - apache#14914 - apache#14918 - apache#14951 - Changes to the TIR lowering flow - apache#14942 - apache#14985 - Codegen updates - apache#14958 - apache#14901 - Compatibility updates/fixes - apache#14892 - apache#14950 - apache#14943 - apache#14944 - apache#14945 - apache#14952 - apache#14982 - apache#14949
Lunderberg
force-pushed
the
single_module_during_build
branch
from
af1628f to
870ba06
Compare
|
Rebased onto main now that #14944 has landed, marked as ready for review. |
Lunderberg
force-pushed
the
single_module_during_build
branch
from
870ba06 to
5286508
Compare
Lunderberg
force-pushed
the
single_module_during_build
branch
from
5286508 to
3a3329f
Compare
Lunderberg
force-pushed
the
single_module_during_build
branch
from
3a3329f to
4c2dd50
Compare
Lunderberg
added a commit
to Lunderberg/tvm
that referenced
this pull request
Jun 10, 2023
The functionality tested in this commit was added across several recent PRs, each of which tested their features in isolation. This PR adds unit tests to validate the end-to-end behavior of TIR subroutine calls. PRs building up to this point: - TVMScript - apache#14889 - apache#14915 - apache#14919 - apache#14941 - Functionality improvements of existing TIR passes - apache#14913 - apache#14914 - apache#14918 - apache#14951 - Changes to the TIR lowering flow - apache#14942 - apache#14985 - Codegen updates - apache#14958 - apache#14901 - Compatibility updates/fixes - apache#14892 - apache#14950 - apache#14943 - apache#14944 - apache#14945 - apache#14952 - apache#14982 - apache#14949
Lunderberg
force-pushed
the
single_module_during_build
branch
5 times, most recently
from
99e7073 to
03c39ef
Compare
Lunderberg
force-pushed
the
single_module_during_build
branch
from
1d8454e to
cc7d384
Compare
Lunderberg
added a commit
to Lunderberg/tvm
that referenced
this pull request
Jun 16, 2023
The functionality tested in this commit was added across several recent PRs, each of which tested their features in isolation. This PR adds unit tests to validate the end-to-end behavior of TIR subroutine calls. PRs building up to this point: - TVMScript - apache#14889 - apache#14915 - apache#14919 - apache#14941 - Functionality improvements of existing TIR passes - apache#14913 - apache#14914 - apache#14918 - apache#14951 - Changes to the TIR lowering flow - apache#14942 - apache#14985 - Codegen updates - apache#14958 - apache#14901 - Compatibility updates/fixes - apache#14892 - apache#14950 - apache#14943 - apache#14944 - apache#14945 - apache#14952 - apache#14982 - apache#14949
Lunderberg
added a commit
to Lunderberg/tvm
that referenced
this pull request
Jun 16, 2023
The functionality tested in this commit was added across several recent PRs, each of which tested their features in isolation. This PR adds unit tests to validate the end-to-end behavior of TIR subroutine calls. PRs building up to this point: - TVMScript - apache#14889 - apache#14915 - apache#14919 - apache#14941 - Functionality improvements of existing TIR passes - apache#14913 - apache#14914 - apache#14918 - apache#14951 - Changes to the TIR lowering flow - apache#14942 - apache#14985 - Codegen updates - apache#14958 - apache#14901 - Compatibility updates/fixes - apache#14892 - apache#14950 - apache#14943 - apache#14944 - apache#14945 - apache#14952 - apache#14982 - apache#14949
Lunderberg
added a commit
to Lunderberg/tvm
that referenced
this pull request
Jun 21, 2023
The functionality tested in this commit was added across several recent PRs, each of which tested their features in isolation. This PR adds unit tests to validate the end-to-end behavior of TIR subroutine calls. PRs building up to this point: - TVMScript - apache#14889 - apache#14915 - apache#14919 - apache#14941 - Functionality improvements of existing TIR passes - apache#14913 - apache#14914 - apache#14918 - apache#14951 - Changes to the TIR lowering flow - apache#14942 - apache#14985 - Codegen updates - apache#14958 - apache#14901 - Compatibility updates/fixes - apache#14892 - apache#14950 - apache#14943 - apache#14944 - apache#14945 - apache#14952 - apache#14982 - apache#14949
Lunderberg
force-pushed
the
single_module_during_build
branch
from
2c850d6 to
93fde73
Compare
Lunderberg
force-pushed
the
single_module_during_build
branch
2 times, most recently
from
27153ac to
f20f78c
Compare
Lunderberg
added a commit
to Lunderberg/tvm
that referenced
this pull request
Jul 3, 2023
The functionality tested in this commit was added across several recent PRs, each of which tested their features in isolation. This PR adds unit tests to validate the end-to-end behavior of TIR subroutine calls. PRs building up to this point: - TVMScript - apache#14889 - apache#14915 - apache#14919 - apache#14941 - Functionality improvements of existing TIR passes - apache#14913 - apache#14914 - apache#14918 - apache#14951 - Changes to the TIR lowering flow - apache#14942 - apache#14985 - Codegen updates - apache#14958 - apache#14901 - Compatibility updates/fixes - apache#14892 - apache#14950 - apache#14943 - apache#14944 - apache#14945 - apache#14952 - apache#14982 - apache#14949
Lunderberg
force-pushed
the
single_module_during_build
branch
from
f20f78c to
b50ccd5
Compare
Lunderberg
force-pushed
the
single_module_during_build
branch
from
af2964d to
2296be8
Compare
Lunderberg
added a commit
to Lunderberg/tvm
that referenced
this pull request
Jul 6, 2023
The functionality tested in this commit was added across several recent PRs, each of which tested their features in isolation. This PR adds unit tests to validate the end-to-end behavior of TIR subroutine calls. PRs building up to this point: - TVMScript - apache#14889 - apache#14915 - apache#14919 - apache#14941 - Functionality improvements of existing TIR passes - apache#14913 - apache#14914 - apache#14918 - apache#14951 - Changes to the TIR lowering flow - apache#14942 - apache#14985 - Codegen updates - apache#14958 - apache#14901 - Compatibility updates/fixes - apache#14892 - apache#14950 - apache#14943 - apache#14944 - apache#14945 - apache#14952 - apache#14982 - apache#14949
Lunderberg
added a commit
to Lunderberg/tvm
that referenced
this pull request
Jul 7, 2023
The functionality tested in this commit was added across several recent PRs, each of which tested their features in isolation. This PR adds unit tests to validate the end-to-end behavior of TIR subroutine calls. PRs building up to this point: - TVMScript - apache#14889 - apache#14915 - apache#14919 - apache#14941 - Functionality improvements of existing TIR passes - apache#14913 - apache#14914 - apache#14918 - apache#14951 - Changes to the TIR lowering flow - apache#14942 - apache#14985 - Codegen updates - apache#14958 - apache#14901 - Compatibility updates/fixes - apache#14892 - apache#14950 - apache#14943 - apache#14944 - apache#14945 - apache#14952 - apache#14982 - apache#14949
Lunderberg
force-pushed
the
single_module_during_build
branch
from
2b7c0af to
3e493d5
Compare
Lunderberg
added a commit
to Lunderberg/tvm
that referenced
this pull request
Jul 12, 2023
Prior to this commit, if device initialization is required, the AOT main function produced a `call_extern()` that included the device context as input. This commit updates the AOT main function to provide the device context only if the function being called accepts a device context as input. If an extra device context argument is included at the call site, the C codegen would produce a function signature that includes the device context for the caller's compilation unit, but a signature without the device context for the callee's compilation unit. While this can compile and run in some cases, it is undefined behavior for the signature to vary between compilation units, and should be avoided. This was initially discovered while debugging apache#14985, in which changes to the lowering flow resulted in the caller and callee being within the same compilation unit.
tqchen
pushed a commit
that referenced
this pull request
Jul 18, 2023
Prior to this commit, if device initialization is required, the AOT main function produced a `call_extern()` that included the device context as input. This commit updates the AOT main function to provide the device context only if the function being called accepts a device context as input. If an extra device context argument is included at the call site, the C codegen would produce a function signature that includes the device context for the caller's compilation unit, but a signature without the device context for the callee's compilation unit. While this can compile and run in some cases, it is undefined behavior for the signature to vary between compilation units, and should be avoided. This was initially discovered while debugging #14985, in which changes to the lowering flow resulted in the caller and callee being within the same compilation unit.
junrushao
pushed a commit
to junrushao/tvm
that referenced
this pull request
Jul 24, 2023
Prior to this commit, if device initialization is required, the AOT main function produced a `call_extern()` that included the device context as input. This commit updates the AOT main function to provide the device context only if the function being called accepts a device context as input. If an extra device context argument is included at the call site, the C codegen would produce a function signature that includes the device context for the caller's compilation unit, but a signature without the device context for the callee's compilation unit. While this can compile and run in some cases, it is undefined behavior for the signature to vary between compilation units, and should be avoided. This was initially discovered while debugging apache#14985, in which changes to the lowering flow resulted in the caller and callee being within the same compilation unit.
junrushao
pushed a commit
to junrushao/tvm
that referenced
this pull request
Jul 27, 2023
Prior to this commit, if device initialization is required, the AOT main function produced a `call_extern()` that included the device context as input. This commit updates the AOT main function to provide the device context only if the function being called accepts a device context as input. If an extra device context argument is included at the call site, the C codegen would produce a function signature that includes the device context for the caller's compilation unit, but a signature without the device context for the callee's compilation unit. While this can compile and run in some cases, it is undefined behavior for the signature to vary between compilation units, and should be avoided. This was initially discovered while debugging apache#14985, in which changes to the lowering flow resulted in the caller and callee being within the same compilation unit.
junrushao
pushed a commit
to junrushao/tvm
that referenced
this pull request
Jul 30, 2023
Prior to this commit, if device initialization is required, the AOT main function produced a `call_extern()` that included the device context as input. This commit updates the AOT main function to provide the device context only if the function being called accepts a device context as input. If an extra device context argument is included at the call site, the C codegen would produce a function signature that includes the device context for the caller's compilation unit, but a signature without the device context for the callee's compilation unit. While this can compile and run in some cases, it is undefined behavior for the signature to vary between compilation units, and should be avoided. This was initially discovered while debugging apache#14985, in which changes to the lowering flow resulted in the caller and callee being within the same compilation unit.
Lunderberg
force-pushed
the
single_module_during_build
branch
from
3e493d5 to
5424a49
Compare
csullivan
pushed a commit
that referenced
this pull request
Aug 8, 2023
Analogous to #14901, treat GlobalVar callees as internal function calls in CodeGenC. This specific PR doesn't provide new end-to-end functionality, as the target="c" backend isn't compiled. It does lead into allowing subroutines in any target whose codegen derives from CodeGenC, which will depend on the single-module lowering flow in #14985. * [CodeGenC] Added unit tests for desired behavior * [CodeGenC] Handle GlobalVar callee as internal function call * Update CodeGenC subclasses for updated interface - Call `DeclareFunction` for each `PrimFunc`, prior to any `AddFunction` calls - Provide both `GlobalVar` and `PrimFunc` to `AddFunction` calls. * Updated CRT test to expect forward declaration * Provide forward declarations for call_extern in cmsis * Avoid duplicate forward declaration C's automatic pointer cast (e.g. `void*` to `int*`) means that use of the arguments to infer the function signature may be incorrect. If a `call_extern` refers to a function within the same module, only output a single forward declaration based on the PrimFunc's parameters, not based on the CallNode's arguments. * Updated expected ptx cuda * Cast the AOT pools to the arg type * Improved tvm::GetType for tvm_access_ptr and address_of These `Call` instances can return a `PointerType(PrimType(pointee_dtype))` rather than a `PrimType(DataType::Handle())`. * [ARM][Topi] Update micro kernels to use same argument type as caller Previously, the micro kernels for gemm, avg_pool, max_pool, and tensordot relied on C's implicit type conversions for the arguments, when the caller's argument types differ from the signature's parameter types. This works, except when the codegen has auto-generated a forward declaration based on the caller's argument types, such as during AOT, which then causes a conflicting definition. Since the codegen cannot determine the functions names from the `"pragma_import_c"` in order to suppress these forward declarations, this conflict can be more easily resolved by updating the micro kernel signatures. The three types of mismatches are below. - Use of `int` or `long` parameters, whose width may vary by compiler, instead of fixed-width types. - TIR expecting the data array's integer type to also be used as an error code's return type, rather than the micro kernels' `int32_t` error code. - Pointer conversion done during argument conversion. Type conversions are done at the start of each micro kernel, to avoid changing types that are used within the computational sections of each micro kernel. * Updated unit tests with private=True Required for internal functions after PR #15214 * Docstring updates from review
Lunderberg
force-pushed
the
single_module_during_build
branch
from
5424a49 to
85b1d5e
Compare
Lunderberg
added a commit
to Lunderberg/tvm
that referenced
this pull request
Aug 8, 2023
The functionality tested in this commit was added across several recent PRs, each of which tested their features in isolation. This PR adds unit tests to validate the end-to-end behavior of TIR subroutine calls. PRs building up to this point: - TVMScript - apache#14889 - apache#14915 - apache#14919 - apache#14941 - Functionality improvements of existing TIR passes - apache#14913 - apache#14914 - apache#14918 - apache#14951 - Changes to the TIR lowering flow - apache#14942 - apache#14985 - Codegen updates - apache#14958 - apache#14901 - Compatibility updates/fixes - apache#14892 - apache#14950 - apache#14943 - apache#14944 - apache#14945 - apache#14952 - apache#14982 - apache#14949
3 tasks
Lunderberg
force-pushed
the
single_module_during_build
branch
from
85b1d5e to
1124239
Compare
|
With #16184 landed, this PR should (hopefully) be able to land without requiring all low-level codegen to directly support function calls. For codegens that do not yet support function calls, device-side kernels can have subroutines inlined to avoid a recurrence of #16033. Currently, this PR is rebased on top of main, but does not use of |
Lunderberg
force-pushed
the
single_module_during_build
branch
from
1124239 to
38e3ac9
Compare
Lunderberg
force-pushed
the
single_module_during_build
branch
from
38e3ac9 to
485bd3e
Compare
Lunderberg
force-pushed
the
single_module_during_build
branch
from
485bd3e to
136a459
Compare
Lunderberg
force-pushed
the
single_module_during_build
branch
from
b3ff7c4 to
04ccfef
Compare
Prior to this commit, a build that used multiple targets needed to provide `tvm::build` with a `Map<Target, IRModule>` specifying which target should be used to compile each `IRModule`. As a result, lowering passes could not introduce new targets based on a PrimFunc's content (e.g. a `with T.target()` frame to delegate out to another device), nor simplify based on cross-device subroutines (e.g. simplify a host-side conditional based on the known output of a device-side internal subroutine). This commit makes the `tvm::attr::kTarget` attribute (`"target"`) be the single source of truth for where a `PrimFunc` will be executed. Other existing methods for specifying the target (the `target` parameter for `tvm.build`, the keys in a `Map<Target,IRModule>`, the parameter to the pass `tir::transform::BindTarget`) are still accepted as inputs, and may provide a default value for `tvm::attr::kTarget` if the attribute is missing, but may not overwrite the target attribute. This is part of a series of commits to simplify the handling of multi-target builds.
Lunderberg
force-pushed
the
single_module_during_build
branch
from
04ccfef to
e7bebaf
Compare
|
The failing unit tests here are blocked by #15757. The |
If a flattened buffer is produced for use in `BufferLoad` and `BufferStore` statements, generate a `DeclBuffer`. This is a subset of the changes made in apache#14778, broken out for ease of testing and review.
When producing a flattened buffer for use in `BufferLoad` and `BufferStore` nodes, generate a `DeclBuffer` for the flattened buffer. This is a subset of the changes made in apache#14778, broken out for ease of testing and review.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Prior to this commit, a build that used multiple targets needed to provide
tvm::buildwith aMap<Target, IRModule>specifying which target should be used to compile eachIRModule. As a result, lowering passes could not introduce new targets based on a PrimFunc's content (e.g. awith T.target()frame to delegate out to another device), nor simplify based on cross-device subroutines (e.g. simplify a host-side conditional based on the known output of a device-side internal subroutine).This commit makes the
tvm::attr::kTargetattribute ("target") be the single source of truth for where aPrimFuncwill be executed. Other existing methods for specifying the target (thetargetparameter fortvm.build, the keys in aMap<Target,IRModule>, the parameter to the passtir::transform::BindTarget) are still accepted as inputs, and may provide a default value fortvm::attr::kTargetif the attribute is missing, but may not overwrite the target attribute.This is part of a series of commits to simplify the handling of multi-target builds.