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
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 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
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 follows up on the recent restructuring that moved the dialects under
lib/Dialect and inter-dialect conversions to lib/Conversion. Originally, the
tests for both the LLVMIR dialect itself and the conversion from Standard to
LLVMIR dialect lived under test/LLVMIR. This no longer reflects the code
structure. Move the tests to either test/Dialect/LLVMIR or
test/Conversion/StandardToLLVM depending on the features they exercise.
PiperOrigin-RevId: 267159219
Some of the operations in the LLVM dialect are required to model the LLVM IR in
MLIR, for example "constant" operations are needed to declare a constant value
since MLIR, unlike LLVM, does not support immediate values as operands. To
avoid confusion with actual LLVM operations, we prefix such axuiliary
operations with "mlir.".
PiperOrigin-RevId: 266942838
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
LLVM intrinsics have an open name space and their names can potentially overlap
with names of LLVM instructions (LLVM intrinsics are functions, not
instructions). In MLIR, LLVM intrinsics are modeled as operations, so it needs
to make sure their names cannot clash with the instructions. Use the "intr."
prefix for intrinsics in the LLVM dialect.
PiperOrigin-RevId: 264372173
Change the prining/parsing of spv.globalVariable to print the type of
the variable after the ':' to be consistent with MLIR convention.
The spv._address_of should print the variable type after the ':'. It was
mistakenly printing the address of the return value. Add a (missing)
test that should have caught that.
Also move spv.globalVariable and spv._address_of tests to
structure-ops.mlir.
PiperOrigin-RevId: 264204686
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
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
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 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
With the introduction of the Loop dialect, uses of the `linalg.for` operation can now be subsumed 1-to-1 by `loop.for`.
This CL performs the replacement and tests are updated accordingly.
PiperOrigin-RevId: 258322565
These ops should not belong to the std dialect.
This CL extracts them in their own dialect and updates the corresponding conversions and tests.
PiperOrigin-RevId: 258123853
This CL splits the lowering of affine to LLVM into 2 parts:
1. affine -> std
2. std -> LLVM
The conversions mostly consists of splitting concerns between the affine and non-affine worlds from existing conversions.
Short-circuiting of affine `if` conditions was never tested or exercised and is removed in the process, it can be reintroduced later if needed.
LoopParametricTiling.cpp is updated to reflect the newly added ForOp::build.
PiperOrigin-RevId: 257794436
Change the AsmPrinter to number values breadth-first so that values in adjacent regions can have the same name. This allows for ModuleOp to contain operations that produce results. This also standardizes the special name of region entry arguments to "arg[0-9+]" now that Functions are also operations.
PiperOrigin-RevId: 257225069
Extend the utility that converts affine loop nests to support other types of
loops by abstracting away common behavior through templates. This also
slightly simplifies the existing Affine to GPU conversion by always passing in
the loop step as an additional kernel argument even though it is a known
constant. If it is used, it will be propagated into the loop body by the
existing canonicalization pattern and can be further constant-folded, otherwise
it will be dropped by canonicalization.
This prepares for the common loop abstraction that will be used for converting
to GPU kernels, which is conceptually close to Linalg loops, while maintaining
the existing conversion operational.
PiperOrigin-RevId: 257172216
This tool allows to execute MLIR IR snippets written in the GPU dialect
on a CUDA capable GPU. For this to work, a working CUDA install is required
and the build has to be configured with MLIR_CUDA_RUNNER_ENABLED set to 1.
PiperOrigin-RevId: 256551415
annotations.
Getters are required as there are currently no global constants in MLIR and this
is an easy way to unblock CUDA execution while waiting for those.
PiperOrigin-RevId: 255169002
The actual transformation from PTX source to a CUDA binary is now factored out,
enabling compiling and testing the transformations independently of a CUDA
runtime.
MLIR has still to be built with NVPTX target support for the conversions to be
built and tested.
PiperOrigin-RevId: 255167139
The current syntax separates the name and value with ':', but ':' is already overloaded by several other things(e.g. trailing types). This makes the syntax difficult to parse in some situtations:
Old:
"foo: 10 : i32"
New:
"foo = 10 : i32"
PiperOrigin-RevId: 255097928
GPU dialect operations (launch and launch_func) use `index` type for thread and
block index values inside the kernel, for compatibility with affine loops.
NVVM dialect operations, following the NVVM intrinsics, use `!llvm.i32` type,
which does not necessarily have the same bit width as the lowered `index` type.
Optionally sign-extend (indices are signed) or truncate the result of the NVVM
dialect operation to the bit width of the lowered `index` type before passing
it to other operations. This behavior is consistent with `std.index_cast`. We
cannot use the latter since we are targeting LLVM dialect types directly,
rather than standard integer types.
PiperOrigin-RevId: 254980868
This converts entire loops into threads/blocks. No check on the size of the
block or grid, or on the validity of parallelization is performed, it is under
the responsibility of the caller to strip-mine the loops and to perform the
dependence analysis before calling the conversion.
PiperOrigin-RevId: 253189268
Mahesh Ravishankar