This revision is the first in a set of improvements that aim at allowing
more generalized named Linalg op generation from a mathematical
specification.
This revision allows creating a new op and checks that the parser,
printer and verifier are hooked up properly.
This opened up a few design points that will be addressed in the future:
1. A named linalg op has a static region builder instead of an
explicitly parsed region. This is not currently compatible with
assemblyFormat so a custom parser / printer are needed.
2. The convention for structured ops and tensor return values needs to
evolve to allow tensor-land and buffer land specifications to agree
3. ReferenceIndexingMaps and referenceIterators will need to become
static to allow building attributes at parse time.
4. Error messages will be improved once we have 3. and we pretty print
in custom form.
Differential Revision: https://reviews.llvm.org/D78327
Unfortunately FileCheck ignores directives with whitespace between the directive and the colon (`CHECK :` for example), thus most of the directives of this test were ignored.
Differential Revision: https://reviews.llvm.org/D78548
The promotion transformation is promoting all input and output buffers of the transformed op. The user might want to only promote some of these buffers.
Differential Revision: https://reviews.llvm.org/D78498
The function attribute in generic ops is not paying for itself.
A region is the more standardized way of specifying a custom computation.
If needed this region can call a function directly.
This is deemed more natural than managing a dedicated function attribute.
This also simplifies named ops generation by trimming unnecessary complexity.
Differential Revision: https://reviews.llvm.org/D78266
The inversePermutation method returns a null map on failure. Update
uses of this method within Linalg to handle this. In LinalgToLoops the
null return value was used to emit scalar code. Modify that to return
failure, and emit scalar implementation when affine map is "empty",
i.e. 1 dims, 0 symbols and no result exprs.
Differential Revision: https://reviews.llvm.org/D77964
The invertPermutation method does not return a nullptr anymore, but
rather returns an empty map for the scalar case. Update the check in
LinalgToLoops to reflect this.
Also add test case for generating scalar code.
The outer parallel loops of a linalg operation is lowered to
loop.parallel, with the other loops lowered to loop.for. This gets the
lowering to loop.parallel on par with the loop.for lowering. In future
the reduction loop could also be lowered to loop.parallel.
Also add a utility function that returns the loops that are
created.
Differential Revision: https://reviews.llvm.org/D77678
This revision builds a simple "fused pass" consisting of 2 levels of tiling, memory promotion and vectorization using linalg transformations written as composable pattern rewrites.
This revision removes the reliance of Promotion on `linalg.slice` which is meant
for the rank-reducing case.
Differential Revision: https://reviews.llvm.org/D77676
Summary:
This revision adds a tensor_reshape operation that operates on tensors.
In the tensor world the constraints are less stringent and we can allow more
arbitrary dynamic reshapes, as long as they are contractions.
The expansion of a dynamic dimension into multiple dynamic dimensions is under-specified and is punted on for now.
Differential Revision: https://reviews.llvm.org/D77360
Summary:
The RAW fusion happens only if the produecer block dominates the consumer block.
The WAW pattern also works with the precondition. I.e., if a producer can
dominate the consumer, they can fairly fuse together.
Since they are all tilable, we can think the pattern like this way:
Input:
```
linalg_op1 view
tile_loop
subview_2
linalg_op2 subview_2
```
Tile the first Linalg op as same as the second Linalg.
```
tile_loop
subview_1
linalg_op1 subview_1
tile_loop
subview_2
liangl_op2 subview_2
```
Since the first Linalg op is tilable in the same way and the computation are
independently, it's fair to fuse it with the second Linalg op.
```
tile_loop
subview_1
linalg_op1 subview_1
linalg_op2 subview_2
```
In short, this patch includes:
- Handling both RAW and WAW pattern.
- Adding a interface method to get input and output buffers.
- Exposing a method to get a StringRef of a dependency type.
- Fixing existing WAW tests and add one more use case: initialize the buffer
before conv op.
Differential Revision: https://reviews.llvm.org/D76897
Summary:
Performs an N-D pooling operation similarly to the description in the TF
documentation:
https://www.tensorflow.org/api_docs/python/tf/nn/pool
Different from the description, this operation doesn't perform on batch and
channel. It only takes tensors of rank `N`.
```
output[x[0], ..., x[N-1]] =
REDUCE_{z[0], ..., z[N-1]}
input[
x[0] * strides[0] - pad_before[0] + dilation_rate[0]*z[0],
...
x[N-1]*strides[N-1] - pad_before[N-1] + dilation_rate[N-1]*z[N-1]
],
```
The required optional arguments are:
- strides: an i64 array specifying the stride (i.e. step) for window
loops.
- dilations: an i64 array specifying the filter upsampling/input
downsampling rate
- padding: an i64 array of pairs (low, high) specifying the number of
elements to pad along a dimension.
If strides or dilations attributes are missing then the default value is
one for each of the input dimensions. Similarly, padding values are zero
for both low and high in each of the dimensions, if not specified.
Differential Revision: https://reviews.llvm.org/D76414
Existing tiling implementation of Linalg would still work for tiling
the batch dimensions of the convolution op.
Differential Revision: https://reviews.llvm.org/D76637
Summary:
To enable this, two changes are needed:
1) Add an optional attribute `padding` to linalg.conv.
2) Compute if the indices accessing is out of bound in the loops. If so, use the
padding value `0`. Otherwise, use the value derived from load.
In the patch, the padding only works for lowering without other transformations,
e.g., tiling, fusion, etc.
Differential Revision: https://reviews.llvm.org/D75722
Summary:
The direct lowering of vector.broadcast into LLVM has been replaced by
progressive lowering into elementary vector ops. This also required a
small refactoring of a llvm.mlir test that used a direct vector.broadcast
operator (just to define a matmul).
Reviewers: nicolasvasilache, andydavis1, rriddle
Reviewed By: nicolasvasilache
Subscribers: mehdi_amini, rriddle, jpienaar, burmako, shauheen, antiagainst, nicolasvasilache, arpith-jacob, mgester, lucyrfox, liufengdb, Joonsoo, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D76143
The current mechanism for identifying is a bit hacky and extremely adhoc, i.e. we explicit check 1-result, 0-operand, no side-effect, and always foldable and then assume that this is a constant. Adding a trait adds structure to this, and makes checking for a constant much more efficient as we can guarantee that all of these things have already been verified.
Differential Revision: https://reviews.llvm.org/D76020
This revision takes advantage of the empty AffineMap to specify the
0-D edge case. This allows removing a bunch of annoying corner cases
that ended up impacting users of Linalg.
Differential Revision: https://reviews.llvm.org/D75831
This attribute details the segment sizes for operand groups within the operation. This revision add support for automatically populating this attribute in the declarative parser.
Differential Revision: https://reviews.llvm.org/D75315
output has zero rank.
While lowering to loops, no indices should be used in the load/store
operation if the buffer is zero-rank.
Differential Revision: https://reviews.llvm.org/D75391
This patch implements the RFCs proposed here:
https://llvm.discourse.group/t/rfc-modify-ifop-in-loop-dialect-to-yield-values/463https://llvm.discourse.group/t/rfc-adding-operands-and-results-to-loop-for/459/19.
It introduces the following changes:
- All Loop Ops region, except for ReduceOp, terminate with a YieldOp.
- YieldOp can have variadice operands that is used to return values out of IfOp and ForOp regions.
- Change IfOp and ForOp syntax and representation to define values.
- Add unit-tests and update .td documentation.
- YieldOp is a terminator to loop.for/if/parallel
- YieldOp custom parser and printer
Lowering is not supported at the moment, and will be in a follow-up PR.
Thanks.
Reviewed By: bondhugula, nicolasvasilache, rriddle
Differential Revision: https://reviews.llvm.org/D74174
Fixing a bug where using a zero-rank shaped type operand to
linalg.generic ops hit an unrelated assert. This also meant that
lowering the operation to loops was not supported. Adding roundtrip
tests and lowering to loops test for zero-rank shaped type operand
with fixes to make the test pass.
Differential Revision: https://reviews.llvm.org/D74638
Summary:
Linalg's promotion pass was only supporting f32 buffers due to how the
zero value was build for the `fill` operation.
Moreover, `promoteSubViewOperands` was returning a vector with one entry
per float subview while omitting integer subviews. For a program
with only integer subviews the return vector would be of size 0.
However, `promoteSubViewsOperands` would try to access a non zero
number of entries of this vector, resulting in a sefgault.
Reviewers: nicolasvasilache, ftynse
Reviewed By: ftynse
Subscribers: mehdi_amini, rriddle, jpienaar, burmako, shauheen, antiagainst, nicolasvasilache, arpith-jacob, mgester, lucyrfox, liufengdb, Joonsoo, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D74532
Summary:
After D72555 has been landed, `linalg.indexed_generic` also accepts ranked
tensor as input and output. Add a test for it.
Differential Revision: https://reviews.llvm.org/D74267
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.
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 initial implementation of the fusion operation exposes a method to
fuse a consumer with its producer, when
- both the producer and consumer operate on tensors
- the producer has only a single result value
- the producer has only "parallel" iterator types
A new interface method hasTensorSemantics is added to verify that an
operation has all operands and results of type RankedTensorType.
Differential Revision: https://reviews.llvm.org/D74172
LinalgDependenceGraph was not updated after successful producer-consumer
fusion for linalg ops. In this patch it is fixed by reconstructing
LinalgDependenceGraph on every iteration. This is very ineffective and
should be improved by updating LDGraph only when it is necessary.
Summary:
After the `subview` operation was migrated from Linalg to Standard, it changed
semantics and does not guarantee the absence of out-of-bounds accesses through
the created view anymore. Compute the size of the subview to make sure it
always fits within the view (subviews in last iterations of the loops may be
smaller than those in other iterations).
Differential Revision: https://reviews.llvm.org/D73614
Summary:
Canonicalization and folding patterns in StandardOps may interfere with the needs
of Linalg. This revision introduces specific foldings for dynamic memrefs that can
be proven to be static.
Very concretely:
Determines whether it is possible to fold it away in the parent Linalg op:
```mlir
%1 = memref_cast %0 : memref<8x16xf32> to memref<?x?xf32>
%2 = linalg.slice %1 ... : memref<?x?xf32> ...
// or
%1 = memref_cast %0 : memref<8x16xf32, affine_map<(i, j)->(16 * i + j)>>
to memref<?x?xf32>
linalg.generic(%1 ...) : memref<?x?xf32> ...
```
into
```mlir
%2 = linalg.slice %0 ... : memref<8x16xf32> ...
// or
linalg.generic(%0 ... : memref<8x16xf32, affine_map<(i, j)->(16 * i + j)>>
```
Reviewers: ftynse, aartbik, jsetoain, tetuante, asaadaldien
Subscribers: mehdi_amini, rriddle, jpienaar, burmako, shauheen, antiagainst, arpith-jacob, mgester, lucyrfox, liufengdb, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D73565
Summary:
This diff adds a transformation patter to rewrite linalg.fill as broadcasting a scaler into a vector.
It uses the same preconditioning as matmul (memory is contiguous).
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/D73391
Summary:
This is a simple extension to allow vectorization to work not only on GenericLinalgOp
but more generally across named ops too.
For now, this still only vectorizes matmul-like ops but is a step towards more
generic vectorization of Linalg ops.
Reviewers: ftynse
Subscribers: mehdi_amini, rriddle, jpienaar, burmako, shauheen, antiagainst, arpith-jacob, mgester, lucyrfox, aartbik, liufengdb, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D72942
Summary:
This diff fixes issues with the semantics of linalg.generic on tensors that appeared when converting directly from HLO to linalg.generic.
The changes are self-contained within MLIR and can be captured and tested independently of XLA.
The linalg.generic and indexed_generic are updated to:
To allow progressive lowering from the value world (a.k.a tensor values) to
the buffer world (a.k.a memref values), a linalg.generic op accepts
mixing input and output ranked tensor values with input and output memrefs.
```
%1 = linalg.generic #trait_attribute %A, %B {other-attributes} :
tensor<?x?xf32>,
memref<?x?xf32, stride_specification>
-> (tensor<?x?xf32>)
```
In this case, the number of outputs (args_out) must match the sum of (1) the
number of output buffer operands and (2) the number of tensor return values.
The semantics is that the linalg.indexed_generic op produces (i.e.
allocates and fills) its return values.
Tensor values must be legalized by a buffer allocation pass before most
transformations can be applied. Such legalization moves tensor return values
into output buffer operands and updates the region argument accordingly.
Transformations that create control-flow around linalg.indexed_generic
operations are not expected to mix with tensors because SSA values do not
escape naturally. Still, transformations and rewrites that take advantage of
tensor SSA values are expected to be useful and will be added in the near
future.
Subscribers: bmahjour, mehdi_amini, rriddle, jpienaar, burmako, shauheen, antiagainst, arpith-jacob, mgester, lucyrfox, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D72555
Summary: The current syntax for AffineMapAttr and IntegerSetAttr conflict with function types, making it currently impossible to round-trip function types(and e.g. FuncOp) in the IR. This revision changes the syntax for the attributes by wrapping them in a keyword. AffineMapAttr is wrapped with `affine_map<>` and IntegerSetAttr is wrapped with `affine_set<>`.
Reviewed By: nicolasvasilache, ftynse
Differential Revision: https://reviews.llvm.org/D72429
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 a new operation to linalg to allow reshaping of an
existing view into a new view in the same buffer at the same offset.
More specifically:
The `linalg.reshape` op produces a new view whose sizes are a reassociation
of the original `view`. Depending on whether or not the reassociated
MemRefType is contiguous, the resulting memref may require explicit alloc
and copies.
A reassociation is defined as a continous grouping of dimensions and is
represented with a affine map array attribute. In the future, non-continous
groupings may be allowed (i.e. permutations, reindexings etc).
For now, it is assumed that either:
1. a reassociation produces and consumes contiguous MemRefType or,
2. the reshape op will be folded into its consumers (by changing the shape
of the computations).
All other cases are undefined behavior and a reshape op may not lower to
LLVM if it cannot be proven statically that it does not require alloc+copy.
A reshape may either collapse or expand dimensions, depending on the
relationship between source and target memref ranks. The verification rule
is that the reassociation maps are applied to the memref with the larger
rank to obtain the memref with the smaller rank. In the case of a dimension
expansion, the reassociation maps can be interpreted as inverse maps.
Examples:
```mlir
// Dimension collapse (i, j) -> i' and k -> k'
%1 = linalg.reshape %0 [(i, j, k) -> (i, j),
(i, j, k) -> (k)] :
memref<?x?x?xf32, stride_spec> into memref<?x?xf32, stride_spec_2>
```
```mlir
// Dimension expansion i -> (i', j') and (k) -> (k')
%1 = linalg.reshape %0 [(i, j, k) -> (i, j),
(i, j, k) -> (k)] :
memref<?x?xf32, stride_spec> into memref<?x?x?xf32, stride_spec_2>
```
The relevant invalid and roundtripping tests are added.
Reviewers: AlexEichenberger, ftynse, rriddle, asaadaldien, yangjunpro
Subscribers: kiszk, merge_guards_bot, mehdi_amini, jpienaar, burmako, shauheen, antiagainst, arpith-jacob, mgester, lucyrfox, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D72168
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
This PR targest issue tensorflow/mlir#295. It exposes the already existing
subiew promotion pass as a declarative pattern
Change-Id: If901ebef9fb53fcd0b12ecc536f6b174ce320b92
Closestensorflow/mlir#315
COPYBARA_INTEGRATE_REVIEW=https://github.com/tensorflow/mlir/pull/315 from tetuante:issue295 8e5f268b6d85f31015c33505329dbd7a4db97ac5
PiperOrigin-RevId: 285801463
Nicolas Vasilache