We have spv.entry_point_abi for specifying the local workgroup size.
It should be decorated onto input gpu.func ops to drive the SPIR-V
CodeGen to generate the proper SPIR-V module execution mode. Compared
to using command-line options for specifying the configuration, using
attributes also has the benefits that 1) we are now able to use
different local workgroup for different entry points and 2) the
tests contains the configuration directly.
Differential Revision: https://reviews.llvm.org/D74012
This CL refactors EDSCs to layer them better and break unnecessary
dependencies. After this refactoring, the top-level EDSC target only
depends on IR but not on Dialects anymore and each dialect has its
own EDSC directory.
This simplifies the layering and breaks cyclic dependencies.
In particular, the declarative builder + folder are made explicit and
are now confined to Linalg.
As the refactoring occurred, certain classes and abstractions that were not
paying for themselves have been removed.
Differential Revision: https://reviews.llvm.org/D74302
The current standard to llvm conversion pass lowers subview ops only if
dynamic offsets are provided. This commit extends the lowering with a
code path that uses the constant offset of the target memref for the
subview op lowering (see Example 3 of the subview op definition for an
example) if no dynamic offsets are provided.
Differential Revision: https://reviews.llvm.org/D74280
The existing (default) calling convention for memrefs in standard-to-LLVM
conversion was motivated by interfacing with LLVM IR produced from C sources.
In particular, it passes a pointer to the memref descriptor structure when
calling the function. Therefore, the descriptor is allocated on stack before
the call. This convention leads to several problems. PR44644 indicates a
problem with stack exhaustion when calling functions with memref-typed
arguments in a loop. Allocating outside of the loop may lead to concurrent
access problems in case the loop is parallel. When targeting GPUs, the contents
of the stack-allocated memory for the descriptor (passed by pointer) needs to
be explicitly copied to the device. Using an aggregate type makes it impossible
to attach pointer-specific argument attributes pertaining to alignment and
aliasing in the LLVM dialect.
Change the default calling convention for memrefs in standard-to-LLVM
conversion to transform a memref into a list of arguments, each of primitive
type, that are comprised in the memref descriptor. This avoids stack allocation
for ranked memrefs (and thus stack exhaustion and potential concurrent access
problems) and simplifies the device function invocation on GPUs.
Provide an option in the standard-to-LLVM conversion to generate auxiliary
wrapper function with the same interface as the previous calling convention,
compatible with LLVM IR porduced from C sources. These auxiliary functions
pack the individual values into a descriptor structure or unpack it. They also
handle descriptor stack allocation if necessary, serving as an allocation
scope: the memory reserved by `alloca` will be freed on exiting the auxiliary
function.
The effect of this change on MLIR-generated only LLVM IR is minimal. When
interfacing MLIR-generated LLVM IR with C-generated LLVM IR, the integration
only needs to require auxiliary functions and change the function name to call
the wrapper function instead of the original function.
This also opens the door to forwarding aliasing and alignment information from
memrefs to LLVM IR pointers in the standrd-to-LLVM conversion.
The existing lowering of gpu.block_dim added a global variable with
the WorkGroupSize decoration. This raises an error within
Vulkan/SPIR-V validation since Vulkan requires this to have a constant
initializer. This is not yet supported in SPIR-V dialect. Changing the
lowering to return the workgroup size as a constant value instead,
obtained from spv.entry_point_abi attribute gets around the issue for
now. The validation goes through since the workgroup size is specified
using spv.execution_mode operation.
Summary:
The `vector.fma` operation is portable enough across targets that we do not want
to keep it wrapped under `vector.outerproduct` and `llvm.intrin.fmuladd`.
This revision lifts the op into the vector dialect and implements the lowering to LLVM by using two patterns:
1. a pattern that lowers from n-D to (n-1)-D by unrolling when n > 2
2. a pattern that converts from 1-D to the proper LLVM representation
Reviewers: ftynse, stellaraccident, aartbik, dcaballe, jsetoain, tetuante
Reviewed By: aartbik
Subscribers: fhahn, dcaballe, merge_guards_bot, mehdi_amini, rriddle, jpienaar, burmako, shauheen, antiagainst, arpith-jacob, mgester, lucyrfox, aartbik, liufengdb, Joonsoo, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D74075
Summary:
This revision exposes the portable `llvm.fma` intrinsic in LLVMOps and uses it
in lieu of `llvm.fmuladd` when lowering the `vector.outerproduct` op to LLVM.
This guarantees proper `fma` instructions will be emitted if the target ISA
supports it.
`llvm.fmuladd` does not have this guarantee in its semantics, despite evidence
that the proper x86 instructions are emitted.
For more details, see https://llvm.org/docs/LangRef.html#llvm-fmuladd-intrinsic.
Reviewers: ftynse, aartbik, dcaballe, fhahn
Reviewed By: aartbik
Subscribers: mehdi_amini, rriddle, jpienaar, burmako, shauheen, antiagainst, arpith-jacob, mgester, lucyrfox, liufengdb, Joonsoo, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D74219
The recent refactoring of build files broke building with the MIR CUDA
integration enabled. This fixes it by adding some additional
dependencies to mlir-opt.
Differential Revision: https://reviews.llvm.org/D74041
We were using normal dictionary attribute for target environment
specification. It becomes cumbersome with more and more fields.
This commit changes the modelling to a dialect-specific attribute,
where we can have control over its storage and assembly form.
Differential Revision: https://reviews.llvm.org/D73959
Summary:
This patch is a step towards enabling BUILD_SHARED_LIBS=on, which
builds most libraries as DLLs instead of statically linked libraries.
The main effect of this is that incremental build times are greatly
reduced, since usually only one library need be relinked in response
to isolated code changes.
The bulk of this patch is fixing incorrect usage of cmake, where library
dependencies are listed under add_dependencies rather than under
target_link_libraries or under the LINK_LIBS tag. Correct usage should be
like this:
add_dependencies(MLIRfoo MLIRfooIncGen)
target_link_libraries(MLIRfoo MLIRlib1 MLIRlib2)
A separate issue is that in cmake, dependencies between static libraries
are automatically included in dependencies. In the above example, if MLIBlib1
depends on MLIRlib2, then it is sufficient to have only MLIRlib1 in the
target_link_libraries. When compiling with shared libraries, it is necessary
to have both MLIRlib1 and MLIRlib2 specified if MLIRfoo uses symbols from both.
Reviewers: mravishankar, antiagainst, nicolasvasilache, vchuravy, inouehrs, mehdi_amini, jdoerfert
Reviewed By: nicolasvasilache, mehdi_amini
Subscribers: Joonsoo, merge_guards_bot, jholewinski, mgorny, mehdi_amini, rriddle, jpienaar, burmako, shauheen, antiagainst, csigg, arpith-jacob, mgester, lucyrfox, herhut, aartbik, liufengdb, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D73653
The patterns for converting `std.alloc` and `std.dealoc` can be configured to
use `llvm.alloca` instead of calling `malloc` and `free`. This configuration
has been only possible through a command-line flag, despite the presence of a
(misleading) parameter in the pass constructor. Use the parameter instead and
only initalize it from the command line flags if the pass is constructed from
the mlir-opt registration.
Summary:
These hooks were originally introduced to support passes deriving the
StandardToLLVM conversion, in particular converting types from different
dialects to LLVM types in a single-step conversion. They are no longer in use
since the pass and conversion infrastructure has evolved sufficiently to make
defining new passes with exactly the same functionality simple through the use
of populate* functions, conversion targets and type converters. Remove the
hooks. Any users of this hooks can call the dialect conversion infrastructure
directly instead, which is likely to require less LoC than these hooks.
Differential Revision: https://reviews.llvm.org/D73795
Summary:
In the original design, gpu.launch required explicit capture of uses
and passing them as operands to the gpu.launch operation. This was
motivated by infrastructure restrictions rather than design. This
change lifts the requirement and removes the concept of kernel
arguments from gpu.launch. Instead, the kernel outlining
transformation now does the explicit capturing.
This is a breaking change for users of gpu.launch.
Differential Revision: https://reviews.llvm.org/D73769
Summary:
This patch introduces an alternative calling convention for
MemRef function arguments in LLVM dialect. It converts MemRef
function arguments to LLVM bare pointers to the MemRef element
type instead of creating a MemRef descriptor. Bare pointers are
then promoted to a MemRef descriptors at the beginning of the
function. This calling convention is only enabled with a flag.
Reviewers: ftynse, bondhugula, nicolasvasilache, rriddle, mehdi_amini
Reviewed By: ftynse, rriddle, mehdi_amini
Subscribers: Joonsoo, flaub, merge_guards_bot, jholewinski, mehdi_amini, rriddle, jpienaar, burmako, shauheen, antiagainst, csigg, arpith-jacob, mgester, lucyrfox, herhut, aartbik, liufengdb, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D72802
Summary:
Rationale:
When lowering to LLVM for different rank insert (n vs k), the offset
arrays needs to drop one dimension (becomes n-1), but the strides
array needs to be preserved (remains k). With regression test.
Note that this example was actually in the documentation, so
extra important to do it right :-)
Reviewers: nicolasvasilache, andydavis1, ftynse
Reviewed By: nicolasvasilache, ftynse
Subscribers: Joonsoo, merge_guards_bot, mehdi_amini, rriddle, jpienaar, burmako, shauheen, antiagainst, arpith-jacob, mgester, lucyrfox, liufengdb, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D73733
This commit adds a pattern to lower linalg.generic for reduction
to spv.GroupNonUniform* ops. Right now this only supports integer
reduction on 1-D input memref. Shader entry point ABI is queried
to make sure that the input memref's shape matches the local
workgroup's invocation configuration. This makes sure that the
workload fits in one local workgroup so that we can leverage
SPIR-V group non-uniform operations.
linglg.generic is a structured op that preserves the right level
of information. It is easier to recognize reduction at this level
than performing analysis on loops.
This commit also exposes `getElementPtr` in SPIRVLowering.h given
that it's a generally useful utility function.
Differential Revision: https://reviews.llvm.org/D73437
Summary:
The current code assumes that one always maps at least one loop to block
dimensions and at least one loop to thread dimensions. If either is not
the case, a loop would get mapped twice.
Differential Revision: https://reviews.llvm.org/D73685
The refactored MemRefType::get() calls all intend to clone from another
memref type, with some modifications. In fact, some calls dropped memory space
during the cloning. Migrate them to the cloning API so that nothing gets
dropped if they are not explicitly listed.
It's close to NFC but not quite, as it helps with propagating memory spaces in
some places.
Differential Revision: https://reviews.llvm.org/D73296
Summary:
In the scope of the lowering phase from GPU to ROCDL, the intructions for the conversion patterns seems to be wrong.
According to https://github.com/ROCm-Developer-Tools/HIP/blob/master/include/hip/hcc_detail/math_fwd.h the instructions need two underscores in the beginning instead of one.
Reviewers: nicolasvasilache, herhut, rriddle
Reviewed By: herhut, rriddle
Subscribers: merge_guards_bot, mehdi_amini, rriddle, jpienaar, burmako, shauheen, antiagainst, csigg, arpith-jacob, mgester, lucyrfox, herhut, liufengdb, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D73535
Summary:
The 'gpu.terminator' operation is used as the terminator for the
regions of gpu.launch. This is to disambugaute them from the
return operation on 'gpu.func' functions.
This is a breaking change and users of the gpu dialect will need
to adapt their code when producting 'gpu.launch' operations.
Reviewers: nicolasvasilache
Subscribers: mehdi_amini, rriddle, jpienaar, burmako, shauheen, antiagainst, csigg, arpith-jacob, mgester, lucyrfox, liufengdb, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D73620
This is how it should've been and brings it more in line with
std::string_view. There should be no functional change here.
This is mostly mechanical from a custom clang-tidy check, with a lot of
manual fixups. It uncovers a lot of minor inefficiencies.
This doesn't actually modify StringRef yet, I'll do that in a follow-up.
Summary:
This also removes the explicit pattern for loop.terminator to ensure
that the terminator is only erased if the parent op is rewritten.
Reductions are not yet supported.
Reviewers: nicolasvasilache
Subscribers: mehdi_amini, rriddle, jpienaar, burmako, shauheen, antiagainst, arpith-jacob, mgester, lucyrfox, aartbik, liufengdb, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D73348
Summary:
The tanh lowering from Standard dialect to NVVM and ROCDL was not working.
The conversion pattern are inserted in the lowering files.
The test cases for the lowerings were added in the test files.
Reviewers: nicolasvasilache, ftynse, herhut
Reviewed By: ftynse, herhut
Subscribers: merge_guards_bot, ftynse, jholewinski, mehdi_amini, rriddle, jpienaar, burmako, shauheen, antiagainst, csigg, arpith-jacob, mgester, lucyrfox, herhut, liufengdb, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D73471
Summary: The new internal representation of operation results now allows for accessing the result types to be more efficient. Changing the API to ArrayRef is more efficient and removes the need to explicitly materialize vectors in several places.
Differential Revision: https://reviews.llvm.org/D73429
Summary:
Affine minimum computation will be used in tiling transformation. The
implementation is mostly boilerplate as we already lower the minimum in the
upper bound of an affine loop.
Differential Revision: https://reviews.llvm.org/D73488
This commit exposes the func op conversion pattern via a new
`populateBuiltinFuncToSPIRVPatterns` function from the standard
to SPIR-V conversion passs. This is structurally better given
that func op belongs to the builtin dialect. More importantly,
this makes the pattern reusable to other dialect to SPIR-V
dialect conversion as other dialect can well adopt builtin
func op instead of having its own. Besides, it's very common
to use func ops as test wrappers in lit tests, so test passes
will need to handle func ops too.
Differential Revision: https://reviews.llvm.org/D73421
Add lowering for constant operation with ranked tensor type to
spv.constant with spv.array type.
Differential Revision: https://reviews.llvm.org/D73022
Summary:
This is based on the use of code constantly checking for an attribute on
a model and instead represents the distinct operaion with a different
op. Instead, this op can be used to provide better filtering.
Reverts "Revert "[mlir] Create a gpu.module operation for the GPU Dialect.""
This reverts commit ac446302ca4145cdc89f377c0c364c29ee303be5 after
fixing internal Google issues.
This additionally updates ROCDL lowering to use the new gpu.module.
Reviewers: herhut, mravishankar, antiagainst, nicolasvasilache
Subscribers: jholewinski, mgorny, mehdi_amini, jpienaar, burmako, shauheen, csigg, arpith-jacob, mgester, lucyrfox, aartbik, liufengdb, llvm-commits, mravishankar, rriddle, antiagainst, bkramer
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D72921
mlir currently fails to build on Solaris:
/vol/llvm/src/llvm-project/dist/mlir/lib/Conversion/VectorToLoops/ConvertVectorToLoops.cpp:78:20: error: reference to 'index_t' is ambiguous
IndexHandle zero(index_t(0)), one(index_t(1));
^
/usr/include/sys/types.h:103:16: note: candidate found by name lookup is 'index_t'
typedef short index_t;
^
/vol/llvm/src/llvm-project/dist/mlir/include/mlir/EDSC/Builders.h:27:8: note: candidate found by name lookup is 'mlir::edsc::index_t'
struct index_t {
^
and many more.
Given that POSIX reserves all identifiers ending in `_t` 2.2.2 The Name Space <https://pubs.opengroup.org/onlinepubs/9699919799/functions/V2_chap02.html>, it seems
quite unwise to use such identifiers in user code, even more so without a distinguished
prefix.
The following patch fixes this by renaming `index_t` to `index_type`.
cases.
Tested on `amd64-pc-solaris2.11` and `sparcv9-sun-solaris2.11`.
Differential Revision: https://reviews.llvm.org/D72619
When lowering `loop.if` to `spv.selection` we explicitly create
a selection header block before the control flow diverges and a
merge block where control flow subsequently converges.
Differential Revision: https://reviews.llvm.org/D72836
Summary:
This diff moves the conversion pass declaration closer to its definition
and makes the namespacing of passes consistent with the rest of the
infrastructure (i.e. `mlir::linalg::createXXXPass` -> `mlir::createXXXPass`).
Reviewers: ftynse, jpienaar, mehdi_amini
Subscribers: rriddle, burmako, shauheen, antiagainst, arpith-jacob, mgester, lucyrfox, aartbik, liufengdb, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D72766
This commit defines a new SPIR-V dialect attribute for specifying
a SPIR-V target environment. It is a dictionary attribute containing
the SPIR-V version, supported extension list, and allowed capability
list. A SPIRVConversionTarget subclass is created to take in the
target environment and sets proper dynmaically legal ops by querying
the op availability interface of SPIR-V ops to make sure they are
available in the specified target environment. All existing conversions
targeting SPIR-V is changed to use this SPIRVConversionTarget. It
probes whether the input IR has a `spv.target_env` attribute,
otherwise, it uses the default target environment: SPIR-V 1.0 with
Shader capability and no extra extensions.
Differential Revision: https://reviews.llvm.org/D72256
Summary:
This is based on the use of code constantly checking for an attribute on
a model and instead represents the distinct operaion with a different
op. Instead, this op can be used to provide better filtering.
Reviewers: herhut, mravishankar, antiagainst, rriddle
Reviewed By: herhut, antiagainst, rriddle
Subscribers: liufengdb, aartbik, jholewinski, mgorny, mehdi_amini, rriddle, jpienaar, burmako, shauheen, antiagainst, nicolasvasilache, csigg, arpith-jacob, mgester, lucyrfox, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D72336
We were seeing some occasional build failures that would come and go.
It appeared to be this missing dependence.
Differential Revision: https://reviews.llvm.org/D72419
Summary:
This diff implements the progressive lowering of insert_strided_slice.
Two cases appear:
1. when the source and dest vectors have different ranks, extract the dest
subvector at the proper offset and reduce to case 2.
2. when they have the same rank N:
a. if the source and dest type are the same, the insertion is trivial:
just forward the source
b. otherwise, iterate over all N-1 D subvectors and create an
extract/insert_strided_slice/insert replacement, reducing the problem
to vecotrs of the same N-1 rank.
This combines properly with the other conversion patterns to lower all the way to LLVM.
Reviewers: ftynse, rriddle, AlexEichenberger, andydavis1, tetuante, nicolasvasilache
Reviewed By: andydavis1
Subscribers: merge_guards_bot, mehdi_amini, jpienaar, burmako, shauheen, antiagainst, arpith-jacob, mgester, lucyrfox, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D72317
Summary:
This diff implements the progressive lowering of strided_slice to either:
1. extractelement + insertelement for the 1-D case
2. extract + optional strided_slice + insert for the n-D case.
This combines properly with the other conversion patterns to lower all the way to LLVM.
Appropriate tests are added.
Reviewers: ftynse, rriddle, AlexEichenberger, andydavis1, tetuante
Reviewed By: andydavis1
Subscribers: merge_guards_bot, mehdi_amini, jpienaar, burmako, shauheen, antiagainst, arpith-jacob, mgester, lucyrfox, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D72310
Summary:
This diff adds lowering of the linalg.reshape op to LLVM.
A new descriptor is created with fields initialized as follows:
1. allocatedPTr, alignedPtr and offset are copied from the source descriptor
2. sizes are copied from the static destination shape
3. strides are copied from the static strides collected with `getStridesAndOffset`
Only the static case in which the target view conforms to strided memref
semantics is supported. Other cases are left for future work and will be added on
a per-need basis.
Reviewers: ftynse, mravishankar
Subscribers: mehdi_amini, rriddle, jpienaar, burmako, shauheen, antiagainst, arpith-jacob, mgester, lucyrfox, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D72316
Summary:
This diff adds support to allow `linalg.generic` and
`linalg.indexed_generic` to take tensor input and output
arguments.
The subset of output tensor operand types must appear
verbatim in the result types after an arrow. The parser,
printer and verifier are extended to accomodate this
behavior.
The Linalg operations now support variadic ranked tensor
return values. This extension exhibited issues with the
current handling of NativeCall in RewriterGen.cpp. As a
consequence, an explicit cast to `SmallVector<Value, 4>`
is added in the proper place to support the new behavior
(better suggestions are welcome).
Relevant cleanups and name uniformization are applied.
Relevant invalid and roundtrip test are added.
Reviewers: mehdi_amini, rriddle, jpienaar, antiagainst, ftynse
Subscribers: burmako, shauheen, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D72022
for (const auto &x : llvm::zip(..., ...))
->
for (auto x : llvm::zip(..., ...))
The return type of zip() is a wrapper that wraps a tuple of references.
> warning: loop variable 'p' is always a copy because the range of type 'detail::zippy<detail::zip_shortest, ArrayRef<long> &, ArrayRef<long> &>' does not return a reference [-Wrange-loop-analysis]
The conversion from std.and/std.or to spv.LogicalAnd/spv.LogicalOr is
only valid for boolean (i1) types. Modify BinaryOpPattern in
StandardToSPIRV.td to allow limiting the type of the operands for
which the pattern is applied.
Differential Revision: https://reviews.llvm.org/D71881
This change refactors pass options to be more similar to how statistics are modeled. More specifically, the options are specified directly on the pass instead of in a separate options class. (Note that the behavior and specification for pass pipelines remains the same.) This brings about several benefits:
* The specification of options is much simpler
* The round-trip format of a pass can be generated automatically
* This gives a somewhat deeper integration with "configuring" a pass, which we could potentially expose to users in the future.
PiperOrigin-RevId: 286953824
This means that in-place, or root, updates need to use explicit calls to `startRootUpdate`, `finalizeRootUpdate`, and `cancelRootUpdate`. The major benefit of this change is that it enables in-place updates in DialectConversion, which simplifies the FuncOp pattern for example. The major downside to this is that the cases that *may* modify an operation in-place will need an explicit cancel on the failure branches(assuming that they started an update before attempting the transformation).
PiperOrigin-RevId: 286933674
This is an initial step to refactoring the representation of OpResult as proposed in: https://groups.google.com/a/tensorflow.org/g/mlir/c/XXzzKhqqF_0/m/v6bKb08WCgAJ
This change will make it much simpler to incrementally transition all of the existing code to use value-typed semantics.
PiperOrigin-RevId: 286844725
Rename the 'shlis' operation in the standard dialect to 'shift_left'. Add tests
for this operation (these have been missing so far) and add a lowering to the
'shl' operation in the LLVM dialect.
Add also 'shift_right_signed' (lowered to LLVM's 'ashr') and 'shift_right_unsigned'
(lowered to 'lshr').
The original plan was to name these operations 'shift.left', 'shift.right.signed'
and 'shift.right.unsigned'. This works if the operations are prefixed with 'std.'
in MLIR assembly. Unfortunately during import the short form is ambigous with
operations from a hypothetical 'shift' dialect. The best solution seems to omit
dots in standard operations for now.
Closestensorflow/mlir#226
PiperOrigin-RevId: 286803388
This will allow us to lower most of gpu.all_reduce (when all_reduce
doesn't exist in the target dialect) within the GPU dialect, and only do
target-specific lowering for the shuffle op.
PiperOrigin-RevId: 286548256
Added test cases for the newly added LLVM operations and lowering features.
Closestensorflow/mlir#300
COPYBARA_INTEGRATE_REVIEW=https://github.com/tensorflow/mlir/pull/300 from dfki-jugr:std_to_llvm da6168bbc1a369ae2e99ad3881fdddd82f075dd4
PiperOrigin-RevId: 286231169
* Fixes use of anonymous namespace for static methods.
* Uses explicit qualifiers(mlir::) instead of wrapping the definition with the namespace.
PiperOrigin-RevId: 286222654
Introduce affine.prefetch: op to prefetch using a multi-dimensional
subscript on a memref; similar to affine.load but has no effect on
semantics, but only on performance.
Provide lowering through std.prefetch, llvm.prefetch and map to llvm's
prefetch instrinsic. All attributes reflected through the lowering -
locality hint, rw, and instr/data cache.
affine.prefetch %0[%i, %j + 5], false, 3, true : memref<400x400xi32>
Signed-off-by: Uday Bondhugula <uday@polymagelabs.com>
Closestensorflow/mlir#225
COPYBARA_INTEGRATE_REVIEW=https://github.com/tensorflow/mlir/pull/225 from bondhugula:prefetch 4c3b4e93bc64d9a5719504e6d6e1657818a2ead0
PiperOrigin-RevId: 286212997
When memory attributions are present in `gpu.func`, require that they are of
memref type and live in memoryspaces 3 and 5 for workgroup and private memory
attributions, respectively. Adapt the conversion from the GPU dialect to the
NVVM dialect to drop the private memory space from attributions as NVVM is able
to model them as local `llvm.alloca`s in the default memory space.
PiperOrigin-RevId: 286161763
The lowering of MemRef types to the LLVM dialect is connected to the underlying
runtime representation of structured memory buffers. It has changed several
times in the past and reached the current state of a LLVM structured-typed
descriptor containing two pointers and all sizes. In several reported use
cases, a different, often simpler, lowering scheme is required. For example,
lowering statically-shaped memrefs to bare LLVM pointers to simplify aliasing
annotation. Split the pattern population functions into those include
memref-related operations and the remaining ones. Users are expected to extend
TypeConverter::convertType to handle the memref types differently.
PiperOrigin-RevId: 286030610
This function has become redundant with MemRefDescriptor::getElementType and is
no longer necessary. Use the MemRefDescriptor pervasively to concentrate
descriptor-related logic in one place and drop the utility function.
PiperOrigin-RevId: 286024168
This updates the lowering pipelines from the GPU dialect to lower-level
dialects (NVVM, SPIRV) to use the recently introduced gpu.func operation
instead of a standard function annotated with an attribute. In particular, the
kernel outlining is updated to produce gpu.func instead of std.func and the
individual conversions are updated to consume gpu.funcs and disallow standard
funcs after legalization, if necessary. The attribute "gpu.kernel" is preserved
in the generic syntax, but can also be used with the custom syntax on
gpu.funcs. The special kind of function for GPU allows one to use additional
features such as memory attribution.
PiperOrigin-RevId: 285822272
The conversion from the Loops dialect to the Standard dialect, also known as
loop-to-cfg lowering, has been historically a function pass. It can be required
on non-Standard function Ops, in particular the recently introduced GPU
functions. Make the conversion an operation pass instead of a function pass.
PiperOrigin-RevId: 285814560
Similar to insert/extract vector instructions but
(1) work on 1-D vectors only
(2) allow for a dynamic index
%c3 = constant 3 : index
%0 = vector.insertelement %arg0, %arg1[%c : index] : vector<4xf32>
%1 = vector.extractelement %arg0[%c3 : index] : vector<4xf32>
PiperOrigin-RevId: 285792205
During the conversion from the standard dialect to the LLVM dialect,
memref-typed arguments are promoted from registers to memory and passed into
functions by pointer. This had been introduced into the lowering to work around
the abesnce of calling convention modeling in MLIR to enable better
interoperability with LLVM IR generated from C, and has been exerciced for
several months. Make this promotion the default calling covention when
converting to the LLVM dialect. This adds the documentation, simplifies the
code and makes the conversion consistent across function operations and
function types used in other places, e.g. in high-order functions or
attributes, which would not follow the same rule previously.
PiperOrigin-RevId: 285751280
This will be evolved into a simple programming model for custom ops and custom layers in followup CLs.
This CL also deletes the obsolete tablegen's reference-impl.td that was using EDSCs.
PiperOrigin-RevId: 285459545
The clamp value determines the returned predicate. Previously, the clamp value was fixed to 31 and the predicate was therefore always true. This is incorrect for partial warp reductions, but went unnoticed because the returned values happened to be zero (but it could be anything).
PiperOrigin-RevId: 285343160
This type is not used anymore now that Linalg view and subview have graduated to std and that alignment is supported on alloc.
PiperOrigin-RevId: 285213424
Both work for the current use case, but the latter allows implementing
prefix sums and is a little easier to understand for partial warps.
PiperOrigin-RevId: 285145287
These come from a non-standard extenion that is not available on Github, so it
only clutters the documentation source with {.mlir} or {.ebnf} tags.
PiperOrigin-RevId: 284733003
The existing GPU to SPIR-V lowering created a spv.module for every
function with gpu.kernel attribute. A better approach is to lower the
module that the function lives in (which has the attribute
gpu.kernel_module) to a spv.module operation. This better captures the
host-device separation modeled by GPU dialect and simplifies the
lowering as well.
PiperOrigin-RevId: 284574688
This allows for users to provide operand_range and result_range in builder.create<> calls, instead of requiring an explicit copy into a separate data structure like SmallVector/std::vector.
PiperOrigin-RevId: 284360710
This class represents a generic abstraction over the different ways to represent a range of Values: ArrayRef<Value *>, operand_range, result_range. This class will allow for removing the many instances of explicit SmallVector<Value *, N> construction. It has the same memory cost as ArrayRef, and only suffers cost from indexing(if+elsing the different underlying representations).
This change only updates a few of the existing usages, with more to be changed in followups; e.g. 'build' API.
PiperOrigin-RevId: 284307996
Since these operations lower to [insert|extract][element|value] at LLVM
dialect level, neither element nor value would correctly reflect the meaning.
PiperOrigin-RevId: 284240727
Accept the address space of the global as a builder argument when constructing
an LLVM::GlobalOp instance. This decreases the reliance of LLVM::GlobalOp users
on the internal name of the attribute used for this purpose. Update several
uses of the address space in GPU to NVVM conversion.
PiperOrigin-RevId: 284233254
GPU functions use memory attributions, a combination of Op attributes and
region arguments, to specify function-wide buffers placed in workgroup or
private memory spaces. Introduce a lowering pattern for GPU functions to be
converted to LLVM functions taking into account memory attributions. Workgroup
attributions get transformed into module-level globals with unique names
derived from function names. Private attributions get converted into
llvm.allocas inside the function body. In both cases, we inject at the
beginning of the function the IR that obtains the raw pointer to the data and
populates a MemRef descriptor based on the MemRef type of buffer, making
attributions compose with the rest of the MemRef lowering and transparent for
use with std.load and std.store. While using raw pointers instead of
descriptors might have been more efficient, it is better implemented as a
canonicalization or a separate transformation so that non-attribution memrefs
could also benefit from it.
PiperOrigin-RevId: 284208396
SPIR-V/Vulkan spec requires the workgroups size to be specified with
the spv.ExecutionMode operation. This was hard-wired to be set to a
particular value. It is now changed to be configurable by clients of
the pass or of the patterns that implement the lowering from GPU to
SPIRV.
PiperOrigin-RevId: 284017482
In the future, a more configurable malloc and free interface should be used and exposed via
extra parameters to the `createLowerToLLVMPass`. Until requirements are gathered, a simple CL flag allows generating code that runs successfully on hardware that cannot use the stdlib.
PiperOrigin-RevId: 283833424
This CL refactors some of the MLIR vector dependencies to allow decoupling VectorOps, vector analysis, vector transformations and vector conversions from each other.
This makes the system more modular and allows extracting VectorToVector into VectorTransforms that do not depend on vector conversions.
This refactoring exhibited a bunch of cyclic library dependencies that have been cleaned up.
PiperOrigin-RevId: 283660308
Not all StandardOps can be lowered to SPIR-V. For example, subview op
implementation requires use of pointer bitcasts which is not valid
according to SPIR-V spec (or at least is ambiguous about it). Such ops
need to be removed/transformed before lowering to SPIR-V. The
SPIRVLegalizationPass is added a place where such legalizations can be
added. Current implementation folds the subview ops with load/stores
so that the lowering itself does not have to convert a subview op.
PiperOrigin-RevId: 283642981
The SPIR-V lowering used nested !spv.arrays to represented
multi-dimensional arrays, with the hope that in-conjunction with the
layout annotations, the shape and layout of memref can be represented
directly. It is unclear though how portable this representation will
end up being. It will rely on driver compilers implementing complex
index computations faithfully. A more portable approach is to use
linearized arrays to represent memrefs and explicitly instantiate all
the index computation in SPIR-V. This gives added benefit that we can
further optimize the generated code in MLIR before generating the
SPIR-V binary.
PiperOrigin-RevId: 283571167
As described in the documentation, ViewOp is expected to take an optional
dynamic offset followed by a list of dynamic sizes. However, the ViewOp parser
did not include a check for the offset being a single value and accepeted a
list of values instead.
Furthermore, several tests have been exercising the wrong syntax of a ViewOp,
passing multiple values to the dyanmic stride list, which was not caught by the
parser. The trailing values could have been erronously interpreted as dynamic
sizes. This is likely due to resyntaxing of the ViewOp, with the previous
syntax taking the list of sizes before the offset. Update the tests to use the
syntax with the offset preceding the sizes.
Worse, the conversion of ViewOp to the LLVM dialect assumed the wrong order of
operands with offset in the trailing position, and erronously relied on the
permissive parsing that interpreted trailing dynamic offset values as leading
dynamic sizes. Fix the lowering to use the correct order of operands.
PiperOrigin-RevId: 283532506
are constant (i.e., there are no size and stride operands).
We recently added canonicalization that rewrites constant size and stride operands to
SubViewOp into static information in the type, so these patterns now occur during code
generation.
PiperOrigin-RevId: 283524688
A recent commit introduced the Linkage attribute to the LLVM dialect and used
it in the Global Op. Also use it in LLVMFuncOp. As per LLVM Language Reference,
if the linkage attribute is omitted, the function is assumed to have external
linkage.
PiperOrigin-RevId: 283493299
LLVM IR supports linkage on global objects such as global variables and
functions. Introduce the Linkage attribute into the LLVM dialect, backed by an
integer storage. Use this attribute on LLVM::GlobalOp and make it mandatory.
Implement parsing/printing of the attribute and conversion to LLVM IR.
See tensorflow/mlir#277.
PiperOrigin-RevId: 283309328
These changes to SPIR-V lowering while adding support for lowering
SUbViewOp, but are not directly related.
- Change the lowering of MemRefType to
!spv.ptr<!spv.struct<!spv.array<...>[offset]>, ..>
This is consistent with the Vulkan spec.
- To enable testing a simple pattern of lowering functions is added to
ConvertStandardToSPIRVPass. This is just used to convert the type of
the arguments of the function. The added function lowering itself is
not meant to be the way functions are eventually lowered into SPIR-V
dialect.
PiperOrigin-RevId: 282589644
This CL uses the recently added op to finish the implementation of Vector -> Vector unrolling by replacing the "fake join op" by a series of InsertStridedSliceOp.
Test is updated accordingly
PiperOrigin-RevId: 282451126
A mismatch in the function declaration and function definition,
prevented the implementation of the createGPUToSPIRVLoweringPass from
being exposed.
PiperOrigin-RevId: 282419815
To simplify the lowering into SPIR-V, while still respecting the ABI
requirements of SPIR-V/Vulkan, split the process into two
1) While lowering a function to SPIR-V (when the function is an entry
point function), allow specifying attributes on arguments and
function itself that describe the ABI of the function.
2) Add a pass that materializes the ABI described in the function.
Two attributes are needed.
1) Attribute on arguments of the entry point function that describe
the descriptor_set, binding, storage class, etc, of the
spv.globalVariable this argument will be replaced by
2) Attribute on function that specifies workgroup size, etc. (for now
only workgroup size).
Add the pass -spirv-lower-abi-attrs to materialize the ABI described
by the attributes.
This change makes the SPIRVBasicTypeConverter class unnecessary and is
removed, further simplifying the SPIR-V lowering path.
PiperOrigin-RevId: 282387587
Support for including a file multiple times was added in tablegen, removing the need for these extra guards. This is because we already insert c/c++ style header guards within each of the specific .td files.
PiperOrigin-RevId: 282076728
Change vector op names from VectorFooOp to Vector_FooOp and from
vector::VectorFooOp to vector::FooOp.
Closestensorflow/mlir#257
COPYBARA_INTEGRATE_REVIEW=https://github.com/tensorflow/mlir/pull/257 from Kayjukh:master dfc3a0e04114885aaec8740d5951d6984d6e1577
PiperOrigin-RevId: 281967461
The current SubViewOp specification allows for either all offsets,
shape and stride to be dynamic or all of them to be static. There are
opportunities for more fine-grained canonicalization based on which of
these are static. For example, if the sizes are static, the result
memref is of static shape. The specification of SubViewOp is modified
to allow on or more of offsets, shapes and strides to be statically
specified. The verification is updated to ensure that the result type
of the subview op is consistent with which of these are static and
which are dynamic.
PiperOrigin-RevId: 281560457
This CL uses the pattern rewrite infrastructure to implement a simple VectorOps -> VectorOps legalization strategy to unroll coarse-grained vector operations into finer grained ones.
The transformation is written using local pattern rewrites to allow composition with other rewrites. It proceeds by iteratively introducing fake cast ops and cleaning canonicalizing or lowering them away where appropriate.
This is an example of writing transformations as compositions of local pattern rewrites that should enable us to make them significantly more declarative.
PiperOrigin-RevId: 281555100
The command-line flag name `lower-to-llvm` for the pass performing dialect
conversion from the Standard dialect to the LLVM dialect is misleading and
inconsistent with most of the conversion passses. It leads the user to believe
that there are no restrictions on what can be converted, while in fact only a
subset of the Standard dialect can be converted (with operations from other
dialects converted by separate passes). Use `convert-std-to-llvm` that better
reflects what the pass does and is consistent with most other conversions.
PiperOrigin-RevId: 281238797
The assertion was introduced in the early days of dialect conversion
infrastructure when we had the matching function separate from the rewriting
function. The infrastructure evolved to have a common matchAndRewrite function
and the separate matching function was dropped without chaning the rewriting
that became matchAndRewrite. This has led to assertion being triggered. Return
a matchFailure instead of failing an assertion on unsupported types.
Closestensorflow/mlir#230
PiperOrigin-RevId: 281113741
This improves consistency and will concretely avoid collisions between VectorExtractElementOp and ExtractElementOp when they are included in the same transforms / rewrites.
PiperOrigin-RevId: 281101588
This is step 1/n in refactoring infrastructure along the Vector dialect to make it ready for retargetability and composable progressive lowering.
PiperOrigin-RevId: 280529784
Refactoring the conversion from StandardOps/GPU dialect to SPIR-V
dialect:
1) Move the SPIRVTypeConversion and SPIRVOpLowering class into SPIR-V
dialect.
2) Add header files that expose functions to add patterns for the
dialects to SPIR-V lowering, as well as a pass that does the
dialect to SPIR-V lowering.
3) Make SPIRVOpLowering derive from OpLowering class.
PiperOrigin-RevId: 280486871
Previous commits removed all uses of LLVMTypeConverter::k*PosInMemRefDescriptor
outside of the MemRefDescriptor class. These numbers are an implementation
detail and can be hidden under a layer of more semantic APIs.
PiperOrigin-RevId: 280442444
Following up on the consolidation of MemRef descriptor conversion, update
Vector-to-LLVM conversion to use the helper class that abstracts away the
implementation details of the MemRef descriptor. This also makes the types of
the attributes in emitted llvm.insert/extractelement operations consistently
i64 instead of a mix of index and i64.
PiperOrigin-RevId: 280441451
This CL moves VectorOps to Tablegen and cleans up the implementation.
This is almost NFC but 2 changes occur:
1. an interface change occurs in the padding value specification in vector_transfer_read:
the value becomes non-optional. As a shortcut we currently use %f0 for all paddings.
This should become an OpInterface for vectorization in the future.
2. the return type of vector.type_cast is trivial and simplified to `memref<vector<...>>`
Relevant roundtrip and invalid tests that used to sit in core are moved to the vector dialect.
The op documentation is moved to the .td file.
PiperOrigin-RevId: 280430869
Following up on the consolidation of MemRef descriptor conversion, update
Linalg-to-LLVM conversion to use the helper class that abstracts away the
implementation details of the MemRef descriptor. This required MemRefDescriptor
to become publicly visible. Since this conversion is heavily EDSC-based,
introduce locally an additional wrapper that uses builder and location pointed
to by the EDSC context while emitting descriptor manipulation operations.
PiperOrigin-RevId: 280429228
Memref descriptor is becoming increasingly complex. Memrefs are manipulated by
multiple standard instructions, each of which has a non-trivial lowering to the
LLVM dialect. This leads to verbose code that manipulates the descriptors
exposing the internals of insert/extractelement opreations. Implement a wrapper
class that contains a memref descriptor and provides semantically named methods
that build the primitive IR operations instead.
PiperOrigin-RevId: 280371225
This refactors the implementation of block signature(type) conversion to not insert fake cast operations to perform the type conversion, but to instead create a new block containing the proper signature. This has the benefit of enabling the use of pre-computed analyses that rely on mapping values. It also leads to a much cleaner implementation overall. The major user facing change is that applySignatureConversion will now replace the entry block of the region, meaning that blocks generally shouldn't be cached over calls to applySignatureConversion.
PiperOrigin-RevId: 280226936
Lowering of CmpIOp, DivISOp, RemISOp, SubIOp and SelectOp to SPIR-V
dialect enables the lowering of operations generated by AffineExpr ->
StandardOps conversion into the SPIR-V dialect.
PiperOrigin-RevId: 280039204
Existing check that sets FuncOp to be dynamically legal was just
checking that the types of the argument are SPIR-V compatible. Since
the current conversion from GPU to SPIR-V does not handle lowering
non-kernel functions, change the legality check to verify that the
FuncOp has the gpu.kernel attribute and has void(void) return type.
PiperOrigin-RevId: 280032782
loop::ForOp can be lowered to the structured control flow represented
by spirv::LoopOp by making the continue block of the spirv::LoopOp the
loop latch and the merge block the exit block. The resulting
spirv::LoopOp has a single back edge from the continue to header
block, and a single exit from header to merge.
PiperOrigin-RevId: 280015614
This CL adds an extra pointer to the memref descriptor to allow specifying alignment.
In a previous implementation, we used 2 types: `linalg.buffer` and `view` where the buffer type was the unit of allocation/deallocation/alignment and `view` was the unit of indexing.
After multiple discussions it was decided to use a single type, which conflates both, so the memref descriptor now needs to carry both pointers.
This is consistent with the [RFC-Proposed Changes to MemRef and Tensor MLIR Types](https://groups.google.com/a/tensorflow.org/forum/#!searchin/mlir/std.view%7Csort:date/mlir/-wKHANzDNTg/4K6nUAp8AAAJ).
PiperOrigin-RevId: 279959463
Now that a view op has graduated to the std dialect, we can update Linalg to use it and remove ops that have become obsolete. As a byproduct, the linalg buffer and associated ops can also disappear.
PiperOrigin-RevId: 279073591
This CL ports the lowering of linalg.view to the newly introduced std.view.
Differences in implementation relate to std.view having slightly different semantics:
1. a static or dynamic offset can be specified.
2. the size of the (contiguous) shape is passed instead of a range.
3. static size and stride information is extracted from the memref type rather than the range.
Besides these differences, lowering behaves the same.
A future CL will update Linalg to use this unified infrastructure.
PiperOrigin-RevId: 278948853
The current lowering of loops to GPU only supports lowering of loop
nests where the loops mapped to workgroups and workitems are perfectly
nested. Here a new lowering is added to handle lowering of imperfectly
nested loop body with the following properties
1) The loops partitioned to workgroups are perfectly nested.
2) The loop body of the inner most loop partitioned to workgroups can
contain one or more loop nests that are to be partitioned across
workitems. Each individual loops nests partitioned to workitems should
also be perfectly nested.
3) The number of workgroups and workitems are not deduced from the
loop bounds but are passed in by the caller of the lowering as values.
4) For statements within the perfectly nested loop nest partitioned
across workgroups that are not loops, it is valid to have all threads
execute that statement. This is NOT verified.
PiperOrigin-RevId: 277958868
Upstream LLVM gained support for #ifndef with https://reviews.llvm.org/D61888
This is changed mechanically via the following command:
find . -name "*.td" -exec sed -i -e ':a' -e 'N' -e '$!ba' -e 's/#ifdef \([A-Z_]*\)\n#else/#ifndef \1/g' {} \;
PiperOrigin-RevId: 277789427
nvvm.shfl.sync.bfly optionally returns a predicate whether source lane was active. Support for this was added to clang in https://reviews.llvm.org/D68892.
Add an optional 'pred' unit attribute to the instruction to return this predicate. Specify this attribute in the partial warp reduction so we don't need to manually compute the predicate.
PiperOrigin-RevId: 275616564
A VectorTypeCastOp can only be used to lower between statically sized contiguous memrefs of scalar and matching vector type. The sizes and strides are thus fully static and easy to determine.
A relevant test is added.
This is a step towards solving tensorflow/mlir#189.
PiperOrigin-RevId: 275538981
Makes the spv.module generated by the GPU to SPIR-V conversion SPIR-V
spec compliant (validated using spirv-val from Vulkan tools).
1) Separate out the VulkanLayoutUtils from
DecorateSPIRVCompositeTypeLayoutPass to make it reusable within the
Type converter in SPIR-V lowering infrastructure. This is used to
compute the layout of the !spv.struct used in global variable type
description.
2) Set the capabilities of the spv.module to Shader (needed for use of
Logical Memory Model, and the extensions to
SPV_KHR_storage_buffer_storage_class for use of Storage Buffer)
PiperOrigin-RevId: 275081486
In addition to specifying the type of accumulation through the 'op' attribute, the accumulation can now also be specified as arbitrary code region.
Adds a gpu.yield op to specify the result of the accumulation.
Also support more types (integers) and accumulations (mul).
PiperOrigin-RevId: 275065447
The current SignatureConversion framework (part of DialectConversion)
allows remapping input arguments to a function from 1->0, 1->1 or
1->many arguments during conversion. Another case is where the
argument itself is dropped, but it's use are remapped to another
Value*.
An example of this is: The Vulkan/SPIR-V spec requires entry functions
to be of type void(void). The GPU -> SPIR-V conversion implemented
this without having the DialectConversion framework track the
remapping that lead to some undefined behavior. The changes here
addresses that.
PiperOrigin-RevId: 275059656
This CL adds a missing lowering for splat of multi-dimensional vectors.
Additional support is also added to the runtime utils library to allow printing memrefs with such vectors.
PiperOrigin-RevId: 274794723
Originally, the lowering of `alloc` operations has been computing the number of
bytes to allocate when lowering based on the properties of MLIR type. This does
not take into account type legalization that happens when compiling LLVM IR
down to target assembly. This legalization can widen the type, potentially
leading to out-of-bounds accesses to `alloc`ed data due to mismatches between
address computation that takes the widening into account and allocation that
does not. Use the LLVM IR's equivalent of `sizeof` to compute the number of
bytes to be allocated:
%0 = getelementptr %type* null, %indexType 0
%1 = ptrtoint %type* %0 to %indexType
adapted from
http://nondot.org/sabre/LLVMNotes/SizeOf-OffsetOf-VariableSizedStructs.txt
PiperOrigin-RevId: 274159900
- dropping what looks like outdated code post some of the previous
updates
Signed-off-by: Uday Bondhugula <uday@polymagelabs.com>
Closestensorflow/mlir#179
COPYBARA_INTEGRATE_REVIEW=https://github.com/tensorflow/mlir/pull/179 from bondhugula:llfix 2a72ea441fe1b3924802273ffbe9870afeb90f91
PiperOrigin-RevId: 274158273
In Standard to LLVM dialect conversion, the binary op conversion pattern
implicitly assumed some operands were of LLVM IR dialect type. This is not
necessarily true, for example if the Ops that produce those operands did not
match the existing convresion patterns. Check if all operands are of LLVM IR
dialect type and if not, fail to patch the binary op pattern.
Closestensorflow/mlir#168
PiperOrigin-RevId: 274063207
The lowering is specified as a pattern and is done only if the result
is a SPIR-V scalar type or vector type.
Handling ConstantOp with index return type needs special handling
since SPIR-V dialect does not have index types. Based on the bitwidth
of the attribute value, either i32 or i64 is chosen.
Other constant lowerings are left as a TODO.
PiperOrigin-RevId: 274056805
This function-like operation allows one to define functions that have wrapped
LLVM IR function type, in particular variadic functions. The operation was
added in parallel to the existing lowering flow, this commit only switches the
flow to use it.
Using a custom function type makes the LLVM IR dialect type system more
consistent and avoids complex conversion rules for functions that previously
had to use the built-in function type instead of a wrapped LLVM IR dialect type
and perform conversions during the analysis.
PiperOrigin-RevId: 273910855
The lowering infrastructure needs to be enhanced to lower into a
spv.Module that is consistent with the SPIR-V spec. The following
changes are needed
1) The Vulkan/SPIR-V validation rules dictates entry functions to have
signature of void(void). This requires changes to the function
signature conversion infrastructure within the dialect conversion
framework. When an argument is dropped from the original function
signature, a function can be specified that when invoked will return
the value to use as a replacement for the argument from the original
function.
2) Some changes to the type converter to make the converted type
consistent with the Vulkan/SPIR-V validation rules,
a) Add support for converting dynamically shaped tensors to
spv.rtarray type.
b) Make the global variable of type !spv.ptr<!spv.struct<...>>
3) Generate the entry point operation for the kernel functions and
automatically compute all the interface variables needed
PiperOrigin-RevId: 273784229
Originally, we were attaching attributes containing CUBIN blobs to the kernel
function called by `gpu.launch_func`. This kernel is now contained in a nested
module that is used as a compilation unit. Attach compiled CUBIN blobs to the
module rather than to the function since we were compiling the module. This
also avoids duplication of the attribute on multiple kernels within the same
module.
PiperOrigin-RevId: 273497303
Originally, the CUBIN getter function was introduced as a mechanism to
circumvent the absence of globals in the LLVM dialect. It would allocate memory
and populate it with the CUBIN data. LLVM dialect now supports globals and they
are already used to store CUBIN data, making the getter function a trivial
address computation of a global. Emit the address computation directly at the
place of `gpu.launch_func` instead of putting it in a function and calling it.
This simplifies the conversion flow and prepares it for using the
DialectConversion infrastructure.
PiperOrigin-RevId: 273496221
Now that the accessor function is a trivial getter of the global variable, it
makes less sense to have the getter generation as a separate pass. Move the
getter generation into the lowering of `gpu.launch_func` to CUDA calls. This
change is mostly code motion, but the process can be simplified further by
generating the addressof inplace instead of using a call. This is will be done
in a follow-up.
PiperOrigin-RevId: 273492517
The kernel function called by gpu.launch_func is now placed into an isolated
nested module during the outlining stage to simplify separate compilation.
Until recently, modules did not have names and could not be referenced. This
limitation was circumvented by introducing a stub kernel at the same name at
the same nesting level as the module containing the actual kernel. This
relation is only effective in one direction: from actual kernel function to its
launch_func "caller".
Leverage the recently introduced symbol name attributes on modules to refer to
a specific nested module from `gpu.launch_func`. This removes the implicit
connection between the identically named stub and kernel functions. It also
enables support for `gpu.launch_func`s to call different kernels located in the
same module.
PiperOrigin-RevId: 273491891
The GPUIndexIntrinsicOpLowering template is currently used by the code in both the GPUToNVVM and GPUToROCDL dirs.
Moving it to a common location to remove code duplication.
Closestensorflow/mlir#163
COPYBARA_INTEGRATE_REVIEW=https://github.com/tensorflow/mlir/pull/163 from deven-amd:deven-refactor-gpu-index-ops-lowering b8dc2a5f5353df196039b6ff2ad42106028693ed
PiperOrigin-RevId: 272863297
This makes the name of the conversion pass more consistent with the naming
scheme, since it actually converts from the Loop dialect to the Standard
dialect rather than working with arbitrary control flow operations.
PiperOrigin-RevId: 272612112
This also adds coverage with a missing test, which uncovered a bug in the conditional for testing whether an offset is dynamic or not.
PiperOrigin-RevId: 272505798
This is a follow-up to the PRtensorflow/mlir#146 which introduced the ROCDL Dialect. This PR introduces a pass to lower GPU Dialect to the ROCDL Dialect. As with the previous PR, this one builds on the work done by @whchung, and addresses most of the review comments in the original PR.
Closestensorflow/mlir#154
COPYBARA_INTEGRATE_REVIEW=https://github.com/tensorflow/mlir/pull/154 from deven-amd:deven-lower-gpu-to-rocdl 809893e08236da5ab6a38e3459692fa04247773d
PiperOrigin-RevId: 272390729
A recent ABI compatibility change affected the conversion from standard
CallOp/CallIndirectOp to LLVM::CallOp by changing its signature. In order to
analyze the signature, the code was looking up the callee symbol in the module.
This is incorrect since, during the conversion, the module may contain both the
original and the converted function op that have the same symbol name. There is
no strict guarantee on which of the two symbols will be found by the lookup.
The conversion was not failing because the type legalizer converts the LLVM
types to themselves making the original and the converted function signatures
ultimately produce the same type.
Instead of looking up the function signature to get the list of result types,
use the types of the CallOp/CallIndirectOp results which must match those of
the function in valid IR. These types are guaranteed to be the original,
unconverted types when converting the operation. Furthermore, this avoids the
need to perform a lookup of a symbol name in the module which may be expensive.
Finally, propagate attributes as-is from the original op to the converted op
since they share the attribute name for the callee of direct calls and the rest
of attributes are not affected by the conversion. This removes the need for
additional contorsions between direct and indirect calls to extract the name of
the optional callee attribute only to insert it back. This also prevents the
conversion from unintentionally dropping the other attributes of the op.
PiperOrigin-RevId: 272218871
This CL finishes the implementation of the Linalg + Affine type unification of the [strided memref RFC](https://groups.google.com/a/tensorflow.org/forum/#!topic/mlir/MaL8m2nXuio).
As a consequence, the !linalg.view type, linalg::DimOp, linalg::LoadOp and linalg::StoreOp can now disappear and Linalg can use standard types everywhere.
PiperOrigin-RevId: 272187165
Perform second reduce only with first warp. This requires an additional __sync_threads(), but doesn't need special handling when the last warp is small. This simplifies support for block sizes that are not multiple of 32.
Supporting partial warp reduce will be done in a separate CL.
PiperOrigin-RevId: 272168917
This CL finishes the implementation of the lowering part of the [strided memref RFC](https://groups.google.com/a/tensorflow.org/forum/#!topic/mlir/MaL8m2nXuio).
Strided memrefs correspond conceptually to the following templated C++ struct:
```
template <typename Elem, size_t Rank>
struct {
Elem *ptr;
int64_t offset;
int64_t sizes[Rank];
int64_t strides[Rank];
};
```
The linearization procedure for address calculation for strided memrefs is the same as for linalg views:
`base_offset + SUM_i index_i * stride_i`.
The following CL will unify Linalg and Standard by removing !linalg.view in favor of strided memrefs.
PiperOrigin-RevId: 272033399
The strided MemRef RFC discusses a normalized descriptor and interaction with library calls (https://groups.google.com/a/tensorflow.org/forum/#!topic/mlir/MaL8m2nXuio).
Lowering of nested LLVM structs as value types does not play nicely with externally compiled C/C++ functions due to ABI issues.
Solving the ABI problem generally is a very complex problem and most likely involves taking
a dependence on clang that we do not want atm.
A simple workaround is to pass pointers to memref descriptors at function boundaries, which this CL implement.
PiperOrigin-RevId: 271591708
The reduction operation is currently fixed to "add", and the scope is fixed to "workgroup".
The implementation is currently limited to sizes that are multiple 32 (warp size) and no larger than 1024.
PiperOrigin-RevId: 271290265
- introduce splat op in standard dialect (currently for int/float/index input
type, output type can be vector or statically shaped tensor)
- implement LLVM lowering (when result type is 1-d vector)
- add constant folding hook for it
- while on Ops.cpp, fix some stale names
Signed-off-by: Uday Bondhugula <uday@polymagelabs.com>
Closestensorflow/mlir#141
COPYBARA_INTEGRATE_REVIEW=https://github.com/tensorflow/mlir/pull/141 from bondhugula:splat 48976a6aa0a75be6d91187db6418de989e03eb51
PiperOrigin-RevId: 270965304
The RFC for unifying Linalg and Affine compilation passes into an end-to-end flow with a predictable ABI and linkage to external function calls raised the question of why we have variable sized descriptors for memrefs depending on whether they have static or dynamic dimensions (https://groups.google.com/a/tensorflow.org/forum/#!topic/mlir/MaL8m2nXuio).
This CL standardizes the ABI on the rank of the memrefs.
The LLVM struct for a memref becomes equivalent to:
```
template <typename Elem, size_t Rank>
struct {
Elem *ptr;
int64_t sizes[Rank];
};
```
PiperOrigin-RevId: 270947276
Roll forward of commit 5684a12.
When outlining GPU kernels, put the kernel function inside a nested module. Then use a nested pipeline to generate the cubins, independently per kernel. In a final pass, move the cubins back to the parent module.
PiperOrigin-RevId: 270639748
This adds sign- and zero-extension and truncation of integer types to the
standard dialects. This allows to perform integer type conversions without
having to go to the LLVM dialect and introduce custom type casts (between
standard and LLVM integer types).
Closestensorflow/mlir#134
COPYBARA_INTEGRATE_REVIEW=https://github.com/tensorflow/mlir/pull/134 from ombre5733:sext-zext-trunc-in-std c7657bc84c0ca66b304e53ec03797e09152e4d31
PiperOrigin-RevId: 270479722
When outlining GPU kernels, put the kernel function inside a nested module. Then use a nested pipeline to generate the cubins, independently per kernel. In a final pass, move the cubins back to the parent module.
PiperOrigin-RevId: 269987720
The helper functions makePositionAttr() and positionAttr() were originally
introduced in the lowering-to-LLVM-dialect pass to construct integer array
attributes that are used for static positions in extract/insertelement.
Constructing an integer array attribute being fairly common, a utility function
Builder::getI64ArrayAttr was later introduced into the Builder API. Drop
makePositionAttr and similar homegrown functions and use that API instead.
PiperOrigin-RevId: 269295836
Add support for specifying extended instructions sets. The operations
in SPIR-V dialect are named as 'spv.<extension-name>.<op-name>'. Use
this mechanism to define a 'Exp' operation from GLSL(450)
instructions.
Later CLs will add support for (de)serialization of these operations,
and update the dialect generation scripts to auto-generate the
specification using the spec directly.
Additional changes:
Add a Type Constraint to OpBase.td to check for vector of specified
lengths. This is used to check that the vector type used in SPIR-V
dialect are of lengths 2, 3 or 4.
Update SPIRVBase.td to use this Type constraints for vectors.
PiperOrigin-RevId: 269234377
SPIR-V recently publishes v1.5, which brings a bunch of symbols
into core. So the suffix "KHR"/"EXT"/etc. is removed from the
symbols. We use a script to pull information from the spec
directly.
Also changed conversion and tests to use GLSL450 instead of
VulkanKHR memory model. GLSL450 is still the main memory model
supported by Vulkan shaders and it does not require extra
capability to enable.
PiperOrigin-RevId: 268992661
This will allow us to use MLIR's folding infrastructure to deduplicate
SPIR-V constants.
This CL also changed isValidSPIRVType in SPIRVDialect to a static method.
PiperOrigin-RevId: 266984403
- the list of passes run by mlir-cpu-runner included -lower-affine and
-lower-to-llvm but was missing -lower-to-cfg (because -lower-affine at
some point used to lower straight to CFG); add -lower-to-cfg in
between. IR with affine ops can now be run by mlir-cpu-runner.
- update -lower-to-cfg to be consistent with other passes (create*Pass methods
were changed to return unique ptrs, but -lower-to-cfg appears to have been
missed).
- mlir-cpu-runner was unable to parse custom form of affine op's - fix
link options
- drop unnecessary run options from test/mlir-cpu-runner/simple.mlir
(none of the test cases had loops)
- -convert-to-llvmir was changed to -lower-to-llvm at some point, but the
create pass method name wasn't updated (this pass converts/lowers to LLVM
dialect as opposed to LLVM IR). Fix this.
(If we prefer "convert", the cmd-line options could be changed to
"-convert-to-llvm/cfg" then.)
Signed-off-by: Uday Bondhugula <uday@polymagelabs.com>
Closestensorflow/mlir#115
PiperOrigin-RevId: 266666909
This change refactors and cleans up the implementation of the operation walk methods. After this refactoring is that the explicit template parameter for the operation type is no longer needed for the explicit op walks. For example:
op->walk<AffineForOp>([](AffineForOp op) { ... });
is now accomplished via:
op->walk([](AffineForOp op) { ... });
PiperOrigin-RevId: 266209552
Instead of lowering the program in two steps (Standard->LLVM followed
by GPU->NVVM), leading to invalid IR inbetween, the runner now uses
one pattern based rewrite step to go directly from Standard+GPU to
LLVM+NVVM.
PiperOrigin-RevId: 265861934
To support a conversion of a simple load-compute-store kernel from GPU
dialect to SPIR-V dialect, the conversion of operations like
"gpu.block_dim", "gpu.thread_id" which allow threads to get the launch
conversion is needed. In SPIR-V these are specified as global
variables with builin attributes. This CL adds support to specify
builtin variables in SPIR-V conversion framework. This is used to
convert the relevant operations from GPU dialect to SPIR-V dialect.
Also add support for conversion of load/store operation in Standard
dialect to SPIR-V dialect.
To simplify the conversion add a method to build a spv.AccessChain
operation that automatically determines the return type based on the
base pointer type and the indices provided.
PiperOrigin-RevId: 265718525
This conversion has been using a stack-allocated array of i8 to store the
null-terminated kernel name in order to pass it to the CUDA wrappers expecting
a C string because the LLVM dialect was missing support for globals. Now that
the suport is introduced, use a global instead.
Refactor global string construction from GenerateCubinAccessors into a common
utility function living in the LLVM namespace.
PiperOrigin-RevId: 264382489
This CL allows binary operations on n-D vector types to be lowered to LLVMIR by performing an (n-1)-D extractvalue, 1-D vector operation and an (n-1)-D insertvalue.
PiperOrigin-RevId: 264339118
The linalg.view type used to be lowered to a struct containing a data pointer, offset, sizes/strides information. This was problematic when passing to external functions due to ABI, struct padding and alignment issues.
The linalg.view type is now lowered to LLVMIR as a *pointer* to a struct containing the data pointer, offset and sizes/strides. This simplifies the interfacing with external library functions and makes it trivial to add new functions without creating a shim that would go from a value type struct to a pointer type.
The consequences are that:
1. lowering explicitly uses llvm.alloca in lieu of llvm.undef and performs the proper llvm.load/llvm.store where relevant.
2. the shim creation function `getLLVMLibraryCallDefinition` disappears.
3. views are passed by pointer, scalars are passed by value. In the future, other structs will be passed by pointer (on a per-need basis).
PiperOrigin-RevId: 264183671
Switch to C++14 standard method as llvm::make_unique has been removed (
https://reviews.llvm.org/D66259). Also mark some targets as c++14 to ease next
integrates.
PiperOrigin-RevId: 263953918
FuncOps in MLIR use explicit capture. So global variables defined in
module scope need to have a symbol name and this should be used to
refer to the variable within the function. This deviates from SPIR-V
spec, which assigns an SSA value to variables at all scopes that can
be used to refer to the variable, which requires SPIR-V functions to
allow implicit capture. To handle this add a new op,
spirv::GlobalVariableOp that can be used to define module scope
variables.
Since instructions need an SSA value, an new spirv::AddressOfOp is
added to convert a symbol reference to an SSA value for use with other
instructions.
This also means the spirv::EntryPointOp instruction needs to change to
allow initializers to be specified using symbol reference instead of
SSA value
The current spirv::VariableOp which returns an SSA value (as defined
by SPIR-V spec) can still be used to define function-scope variables.
PiperOrigin-RevId: 263951109
This CL adds an optional third argument to the vector.outerproduct instruction.
When such a third argument is specified, it is added to the result of the outerproduct and is lowered to FMA intrinsic when the lowering supports it.
In the future, we can add an attribute on the `vector.outerproduct` instruction to modify the operations for which to emit code (e.g. "+/*", "max/+", "min/+", "log/exp" ...).
This CL additionally performs minor cleanups in the vector lowering and adds tests to improve coverage.
This has been independently verified to result in proper fma instructions for haswell as follows.
Input:
```
func @outerproduct_add(%arg0: vector<17xf32>, %arg1: vector<8xf32>, %arg2: vector<17x8xf32>) -> vector<17x8xf32> {
%2 = vector.outerproduct %arg0, %arg1, %arg2 : vector<17xf32>, vector<8xf32>
return %2 : vector<17x8xf32>
}
}
```
Command:
```
mlir-opt vector-to-llvm.mlir -vector-lower-to-llvm-dialect --disable-pass-threading | mlir-opt -lower-to-cfg -lower-to-llvm | mlir-translate --mlir-to-llvmir | opt -O3 | llc -O3 -march=x86-64 -mcpu=haswell -mattr=fma,avx2
```
Output:
```
outerproduct_add: # @outerproduct_add
# %bb.0:
...
vmovaps 112(%rbp), %ymm8
vbroadcastss %xmm0, %ymm0
...
vbroadcastss 64(%rbp), %ymm15
vfmadd213ps 144(%rbp), %ymm8, %ymm0 # ymm0 = (ymm8 * ymm0) + mem
...
vfmadd213ps 400(%rbp), %ymm8, %ymm9 # ymm9 = (ymm8 * ymm9) + mem
...
```
PiperOrigin-RevId: 263743359
Modify the Type converters to have a SPIRVBasicTypeConverter which
only handles conversion from standard types to SPIRV types. Rename
SPIRVEntryFnConverter to SPIRVTypeConverter. This contains the
SPIRVBasicTypeConverter within it.
Remove SPIRVFnLowering class and have separate utility methods to
lower a function as entry function or a non-entry function. The
current setup could end with diamond inheritence that is not very
friendly to use. For example, you could define the following Op
conversion methods that lower from a dialect "Foo" which resuls in
diamond inheritance.
template<typename OpTy>
class FooDialect : public SPIRVOpLowering<OpTy> {...};
class FooFnLowering : public FooDialect, SPIRVFnLowering {...};
PiperOrigin-RevId: 263597101
The GenerateCubinAccessors was generating functions that fill
dynamically-allocated memory with the binary constant of a CUBIN attached as a
stirng attribute to the GPU kernel. This approach was taken to circumvent the
missing support for global constants in the LLVM dialect (and MLIR in general).
Global constants were recently added to the LLVM dialect. Change the
GenerateCubinAccessors pass to emit a global constant array of characters and a
function that returns a pointer to the first character in the array.
PiperOrigin-RevId: 263092052
Since raw pointers are always passed around for IR construct without
implying any ownership transfer, it can be error prone to have implicit
ownership transferred the same way.
For example this code can seem harmless:
Pass *pass = ....
pm.addPass(pass);
pm.addPass(pass);
pm.run(module);
PiperOrigin-RevId: 263053082
This CL is step 3/n towards building a simple, programmable and portable vector abstraction in MLIR that can go all the way down to generating assembly vector code via LLVM's opt and llc tools.
This CL adds support for converting MLIR n-D vector types to (n-1)-D arrays of 1-D LLVM vectors and a conversion VectorToLLVM that lowers the `vector.extractelement` and `vector.outerproduct` instructions to the proper mix of `llvm.vectorshuffle`, `llvm.extractelement` and `llvm.mulf`.
This has been independently verified to produce proper avx2 code.
Input:
```
func @vec_1d(%arg0: vector<4xf32>, %arg1: vector<8xf32>) -> vector<8xf32> {
%2 = vector.outerproduct %arg0, %arg1 : vector<4xf32>, vector<8xf32>
%3 = vector.extractelement %2[0 : i32]: vector<4x8xf32>
return %3 : vector<8xf32>
}
```
Command:
```
mlir-opt vector-to-llvm.mlir -vector-lower-to-llvm-dialect --disable-pass-threading | mlir-opt -lower-to-cfg -lower-to-llvm | mlir-translate --mlir-to-llvmir | opt -O3 | llc -O3 -march=x86-64 -mcpu=haswell -mattr=fma,avx2
```
Output:
```
vec_1d: # @vec_1d
# %bb.0:
vbroadcastss %xmm0, %ymm0
vmulps %ymm1, %ymm0, %ymm0
retq
```
PiperOrigin-RevId: 262895929
This will allow for reusing the same pattern list, which may be costly to continually reconstruct, on multiple invocations.
PiperOrigin-RevId: 262664599
This adds support for fcmp to the LLVM dialect and adds any necessary lowerings, as well as support for EDSCs.
Closestensorflow/mlir#69
PiperOrigin-RevId: 262475255
This allows for proper forward declaration, as opposed to leaking the internal implementation via a using directive. This also allows for all pattern building to go through 'insert' methods on the OwningRewritePatternList, replacing uses of 'push_back' and 'RewriteListBuilder'.
PiperOrigin-RevId: 261816316
The includes related to the LLVM dialect are not used in this file and
introduce an implicit dependencies between the two libraries which isn't
reflected in the CMakeLists.txt, causing non-deterministic build failures.
PiperOrigin-RevId: 261576935
This CL adds an initial implementation for translation of kernel
function in GPU Dialect (used with a gpu.launch_kernel) op to a
spv.Module. The original function is translated into an entry
function.
Most of the heavy lifting is done by adding TypeConversion and other
utility functions/classes that provide most of the functionality to
translate from Standard Dialect to SPIR-V Dialect. These are intended
to be reusable in implementation of different dialect conversion
pipelines.
Note : Some of the files for have been renamed to be consistent with
the norm used by the other Conversion frameworks.
PiperOrigin-RevId: 260759165
Lei Zhang