A new first-party modeling for LLVM IR types in the LLVM dialect has been
developed in parallel to the existing modeling based on wrapping LLVM `Type *`
instances. It resolves the long-standing problem of modeling identified
structure types, including recursive structures, and enables future removal of
LLVMContext and related locking mechanisms from LLVMDialect.
This commit only switches the modeling by (a) renaming LLVMTypeNew to LLVMType,
(b) removing the old implementaiton of LLVMType, and (c) updating the tests. It
is intentionally minimal. Separate commits will remove the infrastructure built
for the transition and update API uses where appropriate.
Depends On D85020
Reviewed By: rriddle
Differential Revision: https://reviews.llvm.org/D85021
Allow for dynamic indices in the `dim` operation.
Rather than an attribute, the index is now an operand of type `index`.
This allows to apply the operation to dynamically ranked tensors.
The correct lowering of dynamic indices remains to be implemented.
Differential Revision: https://reviews.llvm.org/D81551
Having the input dumped on failure seems like a better
default: I debugged FileCheck tests for a while without knowing
about this option, which really helps to understand failures.
Remove `-dump-input-on-failure` and the environment variable
FILECHECK_DUMP_INPUT_ON_FAILURE which are now obsolete.
Differential Revision: https://reviews.llvm.org/D81422
This allows verifying op-indepent attributes (e.g., attributes that do not require the op to have been created) before constructing an operation. These include checking whether required attributes are defined or constraints on attributes (such as I32 attribute). This is not perfect (e.g., if one had a disjunctive constraint where one part relied on the op and the other doesn't, then this would not try and extract the op independent from the op dependent).
The next step is to move these out to a trait that could be verified earlier than in the generated method. The first use case is for inferring the return type while constructing the op. At that point you don't have an Operation yet and that ends up in one having to duplicate the same checks, e.g., verify that attribute A is defined before querying A in shape function which requires that duplication. Instead this allows one to invoke a method to verify all the traits and, if this is checked first during verification, then all other traits could use attributes knowing they have been verified.
It is a little bit funny to have these on the adaptor, but I see the adaptor as a place to collect information about the op before the op is constructed (e.g., avoiding stringly typed accessors, verifying what is possible to verify before the op is constructed) while being cheap to use even with constructed op (so layer of indirection between the op constructed/being constructed). And from that point of view it made sense to me.
Differential Revision: https://reviews.llvm.org/D80842
Add SubgroupId, SubgroupSize and NumSubgroups to GPU dialect ops and add the
lowering of those ops to SPIRV.
Differential Revision: https://reviews.llvm.org/D81042
All ops of the SCF dialect now use the `scf.` prefix instead of `loop.`. This
is a part of dialect renaming.
Differential Revision: https://reviews.llvm.org/D79844
Summary:
Previously operations like std.load created methods for obtaining their
effects but did not inherit from the SideEffect interfaces when their
parameters were decorated with the information. The resulting situation
was that passes had no information on the SideEffects of std.load/store
and had to treat them more cautiously. This adds the inheritance
information when creating the methods.
As a side effect, many tests are modified, as they were using std.load
for testing and this oepration would be folded away as part of pattern
rewriting. Tests are modified to use store or to reutn the result of the
std.load.
Reviewers: mravishankar, antiagainst, nicolasvasilache, herhut, aartbik, ftynse!
Subscribers: mehdi_amini, rriddle, jpienaar, shauheen, antiagainst, nicolasvasilache, csigg, arpith-jacob, mgester, lucyrfox, liufengdb, Joonsoo, bader, grosul1, frgossen, Kayjukh, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D78802
Ensure that `gpu.func` is only used within the dedicated `gpu.module`.
Implement the constraint to the GPU dialect and adopt test cases.
Differential Revision: https://reviews.llvm.org/D78541
Summary:
Use a nested symbol to identify the kernel to be invoked by a `LaunchFuncOp` in the GPU dialect.
This replaces the two attributes that were used to identify the kernel module and the kernel within seperately.
Differential Revision: https://reviews.llvm.org/D78551
Summary:
Use the shortcu `kernel` for the `gpu.kernel` attribute of `gpu.func`.
The parser supports this and test cases are easier to read.
Differential Revision: https://reviews.llvm.org/D78542
Summary:
Fix a broken test case in the `invalid.mlir` lit test case.
`expect` was missing its `e`.
Differential Revision: https://reviews.llvm.org/D78540
Summary: This patch add tests when lowering multiple `gpu.all_reduce` operations in the same kernel. This was previously failing.
Differential Revision: https://reviews.llvm.org/D75930
Summary:
This patch add some builtin operation for the gpu.all_reduce ops.
- for Integer only: `and`, `or`, `xor`
- for Float and Integer: `min`, `max`
This is useful for higher level dialect like OpenACC or OpenMP that can lower to the GPU dialect.
Differential Revision: https://reviews.llvm.org/D75766
Summary:
This patch add some builtin operation for the gpu.all_reduce ops.
- for Integer only: `and`, `or`, `xor`
- for Float and Integer: `min`, `max`
This is useful for higher level dialect like OpenACC or OpenMP that can lower to the GPU dialect.
Differential Revision: https://reviews.llvm.org/D75766
The current setup of the GPU dialect is to model both the host and
device side codegen. For cases (like IREE) the host side modeling
might not directly fit its use case, but device-side codegen is still
valuable. First step in accessing just the device-side functionality
of the GPU dialect is to allow just creating a gpu.func operation from
a gpu.launch operation. In addition this change also "inlines"
operations into the gpu.func op at time of creation instead of this
being a later step.
Differential Revision: https://reviews.llvm.org/D75287
Summary:
The mapper assigns annotations to loop.parallel operations that
are compatible with the loop to gpu mapping pass. The outermost
loop uses the grid dimensions, followed by block dimensions. All
remaining loops are mapped to sequential loops.
Differential Revision: https://reviews.llvm.org/D74963
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:
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:
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
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
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
Introduce a set of function that promote a memref argument of a `gpu.func` to
workgroup memory using memory attribution. The promotion boils down to
additional loops performing the copy from the original argument to the
attributed memory in the beginning of the function, and back at the end of the
function using all available threads. The loop bounds are specified so as to
adapt to any size of the workgroup. These utilities are intended to compose
with other existing utilities (loop coalescing and tiling) in cases where the
distribution of work across threads is uneven, e.g. copying a 2D memref with
only the threads along the "x" dimension. Similarly, specialization of the
kernel to specific launch sizes should be implemented as a separate pass
combining constant propagation and canonicalization.
Introduce a simple attribute-driven pass to test the promotion transformation
since we don't have a heuristic at the moment.
Differential revision: https://reviews.llvm.org/D71904
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
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
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
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
Introduce a new function-like operation to the GPU dialect to provide a
placeholder for the execution semantic description and to add support for GPU
memory hierarchy. This aligns with the overall goal of the dialect to expose
the common abstraction layer for GPU devices, in particular by providing an
MLIR unit of semantics (i.e. an operation) for memory modeling.
This proposal has been discussed in the mailing list:
https://groups.google.com/a/tensorflow.org/d/msg/mlir/RfXNP7Hklsc/MBNN7KhjAgAJ
As decided, the "convergence" aspect of the execution model will be factored
out into a new discussion and therefore is not included in this commit. This
commit only introduces the operation but does not hook it up with the remaining
flow. The intention is to develop the new flow while keeping the old flow
operational and do the switch in a simple, separately reversible commit.
PiperOrigin-RevId: 282357599
Due to legacy reasons, a newline character followed by two spaces was always
inserted before the attributes of the function Op in pretty form. This breaks
formatting when functions are nested in some other operations. Don't print the
newline and just put the attributes on the same line, which is also more
consistent with module Op. Line breaking aware of indentation can be introduced
separately into the parser if deemed useful.
PiperOrigin-RevId: 281721793
This code should be exercised using the existing kernel outlining unit test, but
let me know if I should add a dedicated unit test using a fake call instruction
as well.
PiperOrigin-RevId: 279436321
This allows for them to be used on other non-function, or even other function-like, operations. The algorithms are already generic, so this is simply changing the derived pass type. The majority of this change is just ensuring that the nesting of these passes remains the same, as the pass manager won't auto-nest them anymore.
PiperOrigin-RevId: 276573038
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 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 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
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
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
Alex Zinenko