mlir.dialects.gpu¶

Submodules¶

mlir.dialects.gpu.passes

Attributes¶

`_ods_ir`
`_ods_ir`
`launch`
`_printf`

Classes¶

`_Dialect`
`AllReduceOp`	The `all_reduce` op reduces the value of every work item across a local
`AllocOp`	The `gpu.alloc` operation allocates a region of memory on the GPU. It is
`BarrierOp`	The `barrier` op synchronizes all work items of a workgroup. It is used
`BinaryOp`	GPU binaries provide a semantic mechanism for storing GPU objects,
`BlockDimOp`	Returns the number of threads in the thread block (aka the block size) along
`BlockIdOp`	Returns the block id, i.e. the index of the current block within the grid
`ClusterBlockIdOp`	Returns the block id within the cluster along the x, y, or z `dimension`.
`ClusterDimBlocksOp`	Returns the number of thread blocks in the cluster along
`ClusterDimOp`	Returns the number of cluster identifiers per grid along
`ClusterIdOp`	Returns the cluster id, i.e. the index of the current cluster within the
`Create2To4SpMatOp`	The `gpu.create_2to4_spmat` operation initializes a sparse matrix in dense
`CreateBsrOp`	The `gpu.create_bsr` operation initializes a sparse matrix in BSR format
`CreateCooAoSOp`	The `gpu.create_coo_aos` operation initializes a sparse matrix in COO format
`CreateCooOp`	The `gpu.create_coo` operation initializes a sparse matrix in COO format
`CreateCscOp`	The `gpu.create_csc` operation initializes a sparse matrix in CSC format
`CreateCsrOp`	The `gpu.create_csr` operation initializes a sparse matrix in CSR format
`CreateDnTensorOp`	The `gpu.create_dn_tensor` operation initializes a dense tensor from
`DeallocOp`	The `gpu.dealloc` operation frees the region of memory referenced by a
`DestroyDnTensorOp`	The `gpu.destroy_dn_tensor` operation releases all resources of a dense
`DestroySpMatOp`	The `gpu.destroy_sp_mat` operation releases all resources of a sparse
`DynamicSharedMemoryOp`	This operation provides a memref pointer to the start of dynamic shared
`GPUFuncOp`	Defines a function that can be executed on a GPU. This supports memory
`GPUModuleOp`	GPU module contains code that is intended to be run on a GPU. A host device
`GlobalIdOp`	Returns the unique global workitem/thread id, i.e., the unique index of the
`GridDimOp`	Returns the number of thread blocks in the grid along the x, y, or z
`HostRegisterOp`	This op maps the provided host buffer into the device address space.
`HostUnregisterOp`	This op unmaps the provided host buffer from the device address space.
`LaneIdOp`	Returns the lane id within the subgroup (warp/wave).
`LaunchFuncOp`	Launch a kernel function on the specified grid of thread blocks.
`LaunchOp`	Launch a kernel on the specified grid of thread blocks. The body of the
`MemcpyOp`	The `gpu.memcpy` operation copies the content of one memref to another.
`MemsetOp`	The `gpu.memset` operation sets the content of memref to a scalar value.
`NumSubgroupsOp`	Returns the number of subgroups within a workgroup.
`PrintfOp`	`gpu.printf` takes a literal format string `format` and an arbitrary number of
`ReturnOp`	A terminator operation for regions that appear in the body of `gpu.func`
`RotateOp`	The "rotate" op moves values across lanes in a subgroup (a.k.a., local
`SDDMMBufferSizeOp`	The `gpu.sddmm_buffer_size` operation returns the buffer size required
`SDDMMOp`	The `gpu.sddmm` operation performs the SDDMM operation on the given sparse and
`SetCsrPointersOp`	The `gpu.set_csr_pointers` assigns the given positions, coordinates,
`SetDefaultDeviceOp`	Operation that sets the current default GPU, using a zero-based index
`ShuffleOp`	The "shuffle" op moves values across lanes in a subgroup (a.k.a., local
`SpGEMMCopyOp`	The `gpu.spgemm_copy` operation copies the sparse matrix result of
`SpGEMMCreateDescrOp`	The `gpu.spgemm_create_descr` creates a descriptor for the SpGEMM operation.
`SpGEMMDestroyDescrOp`	The `gpu.spgemm_destroy_descr` destroys the SpGEMM operation descriptor.
`SpGEMMWorkEstimationOrComputeOp`	The `gpu.spgemm_work_estimation_or_compute` is used to call
`SpMMBufferSizeOp`	The `gpu.spmm_buffer_size` operation returns the buffer size required
`SpMMOp`	The `gpu.spmm` operation performs the SpMM operation on the given sparse and
`SpMVBufferSizeOp`	The `gpu.spmv_buffer_size` operation returns the buffer size required
`SpMVOp`	The `gpu.spmv` operation performs the SpMV operation on the given sparse matrix,
`SpMatGetSizeOp`	The `gpu.spmat_get_size` operation retrieves the number of rows, number of
`SubgroupBroadcastOp`	Broadcasts a value from one lane to all active lanes in a subgroup. The
`SubgroupIdOp`	Returns the subgroup id, i.e., the index of the current subgroup within the
`SubgroupMmaComputeOp`	The `gpu.subgroup_mma_compute` operation performs a matrix-multiply accumulate (mma)
`SubgroupMmaConstantMatrixOp`	The `gpu.subgroup_mma_constant_matrix` creates a `!gpu.mma_matrix` with
`SubgroupMmaElementwiseOp`	The `gpu.subgroup_mma_elementwise` takes `!gpu.mma_matrix` inputs and
`SubgroupMmaExtractThreadLocalOp`	The `gpu.subgroup_mma_extract_thread_local` operation extracts a value from `!gpu.mma_matrix`
`SubgroupMmaInsertThreadLocalOp`	The `gpu.subgroup_mma_insert_thread_local` operation inserts a value to `!gpu.mma_matrix`
`SubgroupMmaLoadMatrixOp`	The `gpu.subgroup_mma_load_matrix` operation loads a matrix collectively
`SubgroupMmaStoreMatrixOp`	The `gpu.subgroup_mma_store_matrix` operation stores a matrix collectively
`SubgroupReduceOp`	The `subgroup_reduce` op reduces the values of lanes (work items) across a
`SubgroupSizeOp`	Returns the number of threads within a subgroup.
`TerminatorOp`	A terminator operation for regions that appear in the body of `gpu.launch`
`ThreadIdOp`	Returns the thread id, i.e. the index of the current thread within the block
`WaitOp`	This op synchronizes the host or the device with a list of dependent ops.
`WarpExecuteOnLane0Op`	`warp_execute_on_lane_0` is an operation used to bridge the gap between
`YieldOp`	`gpu.yield` is a special terminator operation for blocks inside regions
`_Dialect`
`AddressSpace`	GPU address space
`AllReduceOperation`	built-in reduction operations supported by gpu.allreduce.
`BroadcastType`	a lane to broadcast from
`CompilationTarget`	GPU compilation format
`Dimension`	a dimension, either 'x', 'y', or 'z'
`Prune2To4SpMatFlag`	pruning strategy for 2:4 sparse matrix
`ShuffleMode`	Indexing modes supported by gpu.shuffle.
`SpGEMMWorkEstimationOrComputeKind`	choose whether spgemm_work_estimation_or_compute does work estimation or compute
`TransposeMode`	transpose mode of sparse matrix supported by sparse tensor ops
`MMAElementwiseOp`	elementwise operation to apply to mma matrix
`MappingId`	Mapping ids for loop mapping
`Processor`	processor for loop mapping
`GPUFuncOp`	Defines a function that can be executed on a GPU. This supports memory
`LaunchFuncOp`	Launch a kernel function on the specified grid of thread blocks.
`LaunchOp`	Launch a kernel on the specified grid of thread blocks. The body of the

Functions¶

`_ods_equally_sized_accessor`(elements, n_simple, ...)	Returns a starting position and a number of elements per variadic group
`_ods_get_default_loc_context`([location])	Returns a context in which the defaulted location is created. If the location
`_get_op_results_or_values`(...)	Returns the given sequence of values or the results of the given op.
`_ods_segmented_accessor`(elements, raw_segments, idx)	Returns a slice of elements corresponding to the idx-th segment.
`all_reduce`(→ _ods_ir)
`alloc`(→ Union[_ods_ir, _ods_ir, AllocOp])
`barrier`(→ BarrierOp)
`binary`(→ BinaryOp)
`block_dim`(→ _ods_ir)
`block_id`(→ _ods_ir)
`cluster_block_id`(→ _ods_ir)
`cluster_dim_blocks`(→ _ods_ir)
`cluster_dim`(→ _ods_ir)
`cluster_id`(→ _ods_ir)
`create_2to4_spmat`(→ Union[_ods_ir, _ods_ir, ...)
`create_bsr`(→ Union[_ods_ir, _ods_ir, CreateBsrOp])
`create_coo_aos`(→ Union[_ods_ir, _ods_ir, CreateCooAoSOp])
`create_coo`(→ Union[_ods_ir, _ods_ir, CreateCooOp])
`create_csc`(→ Union[_ods_ir, _ods_ir, CreateCscOp])
`create_csr`(→ Union[_ods_ir, _ods_ir, CreateCsrOp])
`create_dn_tensor`(→ Union[_ods_ir, _ods_ir, ...)
`dealloc`(→ Union[_ods_ir, _ods_ir, DeallocOp])
`destroy_dn_tensor`(→ Union[_ods_ir, _ods_ir, ...)
`destroy_sp_mat`(→ Union[_ods_ir, _ods_ir, DestroySpMatOp])
`dynamic_shared_memory`(→ _ods_ir)
`func`(→ GPUFuncOp)
`module`(→ GPUModuleOp)
`global_id`(→ _ods_ir)
`grid_dim`(→ _ods_ir)
`host_register`(→ HostRegisterOp)
`host_unregister`(→ HostUnregisterOp)
`lane_id`(→ _ods_ir)
`launch_func`(→ Union[mlir.ir.Value, ...)
`launch`
`memcpy`(→ Union[_ods_ir, _ods_ir, MemcpyOp])
`memset`(→ Union[_ods_ir, _ods_ir, MemsetOp])
`num_subgroups`(→ _ods_ir)
`printf`(format, *args[, loc, ip])
`return_`(→ ReturnOp)
`rotate`(→ _ods_ir)
`sddmm_buffer_size`(→ Union[_ods_ir, _ods_ir, ...)
`sddmm`(→ Union[_ods_ir, _ods_ir, SDDMMOp])
`set_csr_pointers`(→ Union[_ods_ir, _ods_ir, ...)
`set_default_device`(→ SetDefaultDeviceOp)
`shuffle`(→ _ods_ir)
`spgemm_copy`(→ Union[_ods_ir, _ods_ir, SpGEMMCopyOp])
`spgemm_create_descr`(→ Union[_ods_ir, _ods_ir, ...)
`spgemm_destroy_descr`(→ Union[_ods_ir, _ods_ir, ...)
`spgemm_work_estimation_or_compute`(→ Union[_ods_ir, ...)
`spmm_buffer_size`(→ Union[_ods_ir, _ods_ir, ...)
`spmm`(→ Union[_ods_ir, _ods_ir, SpMMOp])
`spmv_buffer_size`(→ Union[_ods_ir, _ods_ir, ...)
`spmv`(→ Union[_ods_ir, _ods_ir, SpMVOp])
`spmat_get_size`(→ Union[_ods_ir, _ods_ir, SpMatGetSizeOp])
`subgroup_broadcast`(→ _ods_ir)
`subgroup_id`(→ _ods_ir)
`subgroup_mma_compute`(→ _ods_ir)
`subgroup_mma_constant_matrix`(→ _ods_ir)
`subgroup_mma_elementwise`(→ _ods_ir)
`subgroup_mma_extract_thread_local`(→ _ods_ir)
`subgroup_mma_insert_thread_local`(→ _ods_ir)
`subgroup_mma_load_matrix`(→ _ods_ir)
`subgroup_mma_store_matrix`(→ SubgroupMmaStoreMatrixOp)
`subgroup_reduce`(→ _ods_ir)
`subgroup_size`(→ _ods_ir)
`terminator`(→ TerminatorOp)
`thread_id`(→ _ods_ir)
`wait`(→ Union[mlir.ir.Value, List[mlir.ir.Value], ...)
`warp_execute_on_lane_0`(→ Union[_ods_ir, _ods_ir, ...)
`yield_`(→ YieldOp)
`register_attribute_builder`(kind[, replace])
`_gpu_addressspaceenum`(x, context)
`_gpu_allreduceoperation`(x, context)
`_gpu_broadcasttype`(x, context)
`_gpu_compilationtargetenum`(x, context)
`_gpu_dimension`(x, context)
`_gpu_prune2to4spmatflag`(x, context)
`_gpu_shufflemode`(x, context)
`_gpu_spgemmworkestimationorcomputekind`(x, context)
`_gpu_transposemode`(x, context)
`_mmaelementwise`(x, context)
`_mappingidenum`(x, context)
`_processorenum`(x, context)
`_gpu_addressspaceattr`(x, context)
`_gpu_allreduceoperationattr`(x, context)
`_gpu_broadcasttypeattr`(x, context)
`_gpu_dimensionattr`(x, context)
`_gpu_prune2to4spmatflagattr`(x, context)
`_gpu_shufflemodeattr`(x, context)
`_gpu_spgemmworkestimationorcomputekindattr`(x, context)
`_gpu_transposemodeattr`(x, context)
`_mmaelementwiseattr`(x, context)
`gpu_async_token`()
`_convert_literal_to_constant`(→ mlir.ir.Value)
`launch_func`(→ Union[mlir.ir.Value, ...)
`wait`(→ Union[mlir.ir.Value, List[mlir.ir.Value], ...)
`launch_`(grid_size, block_size[, async_dependencies, ...])
`printf`(format, *args[, loc, ip])

Package Contents¶

mlir.dialects.gpu._ods_equally_sized_accessor(elements, n_simple, n_variadic, n_preceding_simple, n_preceding_variadic)¶

Returns a starting position and a number of elements per variadic group assuming equally-sized groups and the given numbers of preceding groups.

elements: a sequential container. n_simple: the number of non-variadic groups in the container. n_variadic: the number of variadic groups in the container. n_preceding_simple: the number of non-variadic groups preceding the current group. n_preceding_variadic: the number of variadic groups preceding the current group.

mlir.dialects.gpu._ods_get_default_loc_context(location=None)¶: Returns a context in which the defaulted location is created. If the location is None, takes the current location from the stack.

mlir.dialects.gpu._get_op_results_or_values(arg: mlir._mlir_libs._mlir.ir.OpView | mlir._mlir_libs._mlir.ir.Operation | Sequence[mlir._mlir_libs._mlir.ir.OpView | mlir._mlir_libs._mlir.ir.Operation | mlir._mlir_libs._mlir.ir.Value]) → Sequence[mlir._mlir_libs._mlir.ir.OpView | mlir._mlir_libs._mlir.ir.Operation | mlir._mlir_libs._mlir.ir.Value] | mlir._mlir_libs._mlir.ir.OpResultList¶

Returns the given sequence of values or the results of the given op.

This is useful to implement op constructors so that they can take other ops as lists of arguments instead of requiring the caller to extract results for every op.

mlir.dialects.gpu._ods_segmented_accessor(elements, raw_segments, idx)¶

Returns a slice of elements corresponding to the idx-th segment.

elements: a sliceable container (operands or results). raw_segments: an mlir.ir.Attribute, of DenseI32Array subclass containing sizes of the segments. idx: index of the segment.

mlir.dialects.gpu._ods_ir¶

class mlir.dialects.gpu._Dialect(descriptor: object)¶

Bases: _ods_ir

DIALECT_NAMESPACE = 'gpu'¶

class mlir.dialects.gpu.AllReduceOp(value, *, op=None, uniform=None, results=None, loc=None, ip=None)¶

Bases: _ods_ir

The all_reduce op reduces the value of every work item across a local workgroup. The result is equal for all work items of a workgroup.

For example, both

%1 = gpu.all_reduce add %0 {} : (f32) -> (f32)
%2 = gpu.all_reduce %0 {
^bb(%lhs : f32, %rhs : f32):
  %sum = arith.addf %lhs, %rhs : f32
  "gpu.yield"(%sum) : (f32) -> ()
} : (f32) -> (f32)

compute the sum of each work item’s %0 value. The first version specifies the accumulation as operation, whereas the second version specifies the accumulation as code region. The reduction operation must be one of:

Integer types: add, mul, minui, minsi, maxui, maxsi, and,

or, xor * Floating point types: add, mul, minnumf, maxnumf, minimumf, maximumf

If uniform flag is set either none or all work items of a workgroup need to execute this op in convergence.

OPERATION_NAME = 'gpu.all_reduce'¶

_ODS_REGIONS = (1, True)¶

value() → _ods_ir¶

op() → _ods_ir | None¶

uniform() → bool¶

result() → _ods_ir¶: Shortcut to get an op result if it has only one (throws an error otherwise).

body() → _ods_ir¶

mlir.dialects.gpu.all_reduce(value, *, op=None, uniform=None, results=None, loc=None, ip=None) → _ods_ir¶

class mlir.dialects.gpu.AllocOp(memref, asyncToken, asyncDependencies, dynamicSizes, symbolOperands, *, hostShared=None, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.alloc operation allocates a region of memory on the GPU. It is similar to the memref.alloc op, but supports asynchronous GPU execution.

The op does not execute before all async dependencies have finished executing.

If the async keyword is present, the op is executed asynchronously (i.e. it does not block until the execution has finished on the device). In that case, it also returns a !gpu.async.token.

If the host_shared keyword is present, the memory will be allocated in a memory accessible both on host and on device.

Example:

%memref, %token = gpu.alloc async [%dep] host_shared (%width) : memref<64x?xf32, 1>

OPERATION_NAME = 'gpu.alloc'¶

_ODS_OPERAND_SEGMENTS¶

_ODS_REGIONS = (0, True)¶

asyncDependencies() → _ods_ir¶

dynamicSizes() → _ods_ir¶

symbolOperands() → _ods_ir¶

hostShared() → bool¶

memref() → _ods_ir¶

asyncToken() → _ods_ir | None¶

mlir.dialects.gpu.alloc(memref, async_token, async_dependencies, dynamic_sizes, symbol_operands, *, host_shared=None, loc=None, ip=None) → _ods_ir | _ods_ir | AllocOp¶

class mlir.dialects.gpu.BarrierOp(*, loc=None, ip=None)¶

Bases: _ods_ir

The barrier op synchronizes all work items of a workgroup. It is used to coordinate communication between the work items of the workgroup.

gpu.barrier

waits until all work items in the workgroup have reached this point and all memory accesses made by these work items prior to the op are visible to all work items in the workgroup. Data hazards between work items accessing the same memory can be avoided by synchronizing work items in-between these accesses.

Either none or all work items of a workgroup need to execute this op in convergence.

OPERATION_NAME = 'gpu.barrier'¶

_ODS_REGIONS = (0, True)¶

mlir.dialects.gpu.barrier(*, loc=None, ip=None) → BarrierOp¶

class mlir.dialects.gpu.BinaryOp(sym_name, objects, *, offloadingHandler=None, loc=None, ip=None)¶

Bases: _ods_ir

GPU binaries provide a semantic mechanism for storing GPU objects, e.g. the result of compiling a GPU module to an object file.

This operation has 3 arguments:

The name of the binary.
An optional attribute implementing the offloading LLVM translation interface.
An array of GPU object attributes.

During translation, the offloading attribute will be called for translating GPU binary and launch_func operations. The default offloading handler is: #gpu.select_object, this handler selects the first object from the array and embeds it as a string.

Examples:

// Selects the first object.
gpu.binary @myobject [#gpu.object<...>, #gpu.object<...>]
// Uses the `#foo.my_handler` for handling the binary during translation.
gpu.binary @myobject <#foo.my_handler> [#gpu.object<...>, #gpu.object<...>]
// Selects the object with the `#rocdl.target` target attribute.
gpu.binary @myobject <#gpu.select_object<#rocdl.target>> [#gpu.object<...>, #gpu.object<#rocdl.target, ...>]

OPERATION_NAME = 'gpu.binary'¶

_ODS_REGIONS = (0, True)¶

sym_name() → _ods_ir¶

offloadingHandler() → _ods_ir | None¶

objects() → _ods_ir¶

mlir.dialects.gpu.binary(sym_name, objects, *, offloading_handler=None, loc=None, ip=None) → BinaryOp¶

class mlir.dialects.gpu.BlockDimOp(dimension, *, upper_bound=None, results=None, loc=None, ip=None)¶

Bases: _ods_ir

Returns the number of threads in the thread block (aka the block size) along the x, y, or z dimension.

Example:

%bDimX = gpu.block_dim x

If known_block_size is set on an this operation’s enclosing gpu.func, or gpu.known_block_size is set on an enclosing FunctionOpInterface implementor, or if the enclosing gpu.launch specifies a constant size for dimension’s blocks, these contextual facts may be used to infer that this operation has a constant value, though such a transformation will not be performed by canonicalization or the default constant folder. Executions which cause that constant-value assumption to be false incur undefined behavior.

If upper_bound is set, executions where the bblock size along dimension exceeds upper_bound cause undefined behavior.

There is an implicit upper bound of kMaxDim (currently uint32_t::max).

OPERATION_NAME = 'gpu.block_dim'¶

_ODS_REGIONS = (0, True)¶

dimension() → _ods_ir¶

upper_bound() → _ods_ir | None¶

mlir.dialects.gpu.block_dim(dimension, *, upper_bound=None, results=None, loc=None, ip=None) → _ods_ir¶

class mlir.dialects.gpu.BlockIdOp(dimension, *, upper_bound=None, results=None, loc=None, ip=None)¶

Bases: _ods_ir

Returns the block id, i.e. the index of the current block within the grid along the x, y, or z dimension.

Example:

%bIdY = gpu.block_id y

If upper_bound is set, or if one can be inferred from known_grid_size-type annotations in context, executions where the block index in dimension would be greater than or equal to that bound cause undefined behavior. upper_bound takes priority over bounds inferrable from context.

There is an implicit upper bound of kMaxDim (currently uint32_t::max).

OPERATION_NAME = 'gpu.block_id'¶

_ODS_REGIONS = (0, True)¶

dimension() → _ods_ir¶

upper_bound() → _ods_ir | None¶

mlir.dialects.gpu.block_id(dimension, *, upper_bound=None, results=None, loc=None, ip=None) → _ods_ir¶

class mlir.dialects.gpu.ClusterBlockIdOp(dimension, *, upper_bound=None, results=None, loc=None, ip=None)¶

Bases: _ods_ir

Returns the block id within the cluster along the x, y, or z dimension.

Example:

%cBlockIdY = gpu.cluster_block_id y

If upper_bound is set, then executing (a lowering of) this operation in an environment where the number of thread blocks per cluster along dimension is greater than upper_bound causes undefined behavior.

There is an implicit upper bound of kMaxClusterDim (currently 8).

OPERATION_NAME = 'gpu.cluster_block_id'¶

_ODS_REGIONS = (0, True)¶

dimension() → _ods_ir¶

upper_bound() → _ods_ir | None¶

mlir.dialects.gpu.cluster_block_id(dimension, *, upper_bound=None, results=None, loc=None, ip=None) → _ods_ir¶

class mlir.dialects.gpu.ClusterDimBlocksOp(dimension, *, upper_bound=None, results=None, loc=None, ip=None)¶

Bases: _ods_ir

Returns the number of thread blocks in the cluster along the x, y, or z dimension.

Example:

%cDimBlocksX = gpu.cluster_dim_blocks x

If upper_bound is set, then executing (a lowering of) this operation in an environment where the thread blocks per cluster is greater than upper_bound causes undefined behavior.

There is an implicit upper bound of kMaxClusterDim (currently 8).

OPERATION_NAME = 'gpu.cluster_dim_blocks'¶

_ODS_REGIONS = (0, True)¶

dimension() → _ods_ir¶

upper_bound() → _ods_ir | None¶

mlir.dialects.gpu.cluster_dim_blocks(dimension, *, upper_bound=None, results=None, loc=None, ip=None) → _ods_ir¶

class mlir.dialects.gpu.ClusterDimOp(dimension, *, upper_bound=None, results=None, loc=None, ip=None)¶

Bases: _ods_ir

Returns the number of cluster identifiers per grid along the x, y, or z dimension.

Example:

%cDimX = gpu.cluster_dim x

If upper_bound is set, then executing (a lowering of) this operation in an environment where the clusters per grid is greater than upper_bound causes undefined behavior.

There is an implicit upper bound of kMaxDim (currently uint32_t::max).

OPERATION_NAME = 'gpu.cluster_dim'¶

_ODS_REGIONS = (0, True)¶

dimension() → _ods_ir¶

upper_bound() → _ods_ir | None¶

mlir.dialects.gpu.cluster_dim(dimension, *, upper_bound=None, results=None, loc=None, ip=None) → _ods_ir¶

class mlir.dialects.gpu.ClusterIdOp(dimension, *, upper_bound=None, results=None, loc=None, ip=None)¶

Bases: _ods_ir

Returns the cluster id, i.e. the index of the current cluster within the grid along the x, y, or z dimension.

Example:

%cIdY = gpu.cluster_id y

If upper_bound is set, then executing (a lowering of) this operation in an environment where the number of clusters in the grid along dimension is greater than upper_bound causes undefined behavior.

There is an implicit upper bound of kMaxDim (currently uint32_t::max).

OPERATION_NAME = 'gpu.cluster_id'¶

_ODS_REGIONS = (0, True)¶

dimension() → _ods_ir¶

upper_bound() → _ods_ir | None¶

mlir.dialects.gpu.cluster_id(dimension, *, upper_bound=None, results=None, loc=None, ip=None) → _ods_ir¶

class mlir.dialects.gpu.Create2To4SpMatOp(spMat, asyncToken, asyncDependencies, rows, cols, memref, *, pruneFlag=None, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.create_2to4_spmat operation initializes a sparse matrix in dense format with 2:4 sparsity. The buffers must already be copied from the host to the device prior to using this operation. The operation returns a handle to the sparse matrix descriptor.

If the async keyword is present, the op is executed asynchronously (i.e. it does not block until the execution has finished on the device). In that case, it returns a !gpu.async.token in addition to the environment.

Example:

%spmat, %token = gpu.create_2to4_spmat async [%dep] {PRUNE_AND_CHECK} %rows, %cols, %mem: memref<?xf64>

OPERATION_NAME = 'gpu.create_2to4_spmat'¶

_ODS_REGIONS = (0, True)¶

asyncDependencies() → _ods_ir¶

rows() → _ods_ir¶

cols() → _ods_ir¶

memref() → _ods_ir¶

pruneFlag() → _ods_ir¶

spMat() → _ods_ir¶

asyncToken() → _ods_ir | None¶

mlir.dialects.gpu.create_2to4_spmat(sp_mat, async_token, async_dependencies, rows, cols, memref, *, prune_flag=None, loc=None, ip=None) → _ods_ir | _ods_ir | Create2To4SpMatOp¶

class mlir.dialects.gpu.CreateBsrOp(spmat, asyncToken, asyncDependencies, brows, bcols, bnnz, rBlockSize, cBlockSize, bRowPos, bColIdxs, values, *, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.create_bsr operation initializes a sparse matrix in BSR format with the given sizes for the matrix and blocks from the given position, index, and values buffers. The buffers must already be copied from the host to the device prior to using this operation. The operation returns a handle to the sparse matrix descriptor.

The BSR format is similar to CSR, where the column indices represent two-dimensional blocks instead of a single matrix entry. Note that this operation (currently) only supports storage with square blocks, i.e., rBlockSize == cBlockSize.

Example:

%spmat, %token = gpu.create_bsr async [%dep]
   %brows, %bcols, %bnnz, %rBlockSize, %cBlockSize,
   %bRowPos, %bColIdxs, %values : memref<?xindex>, memref<?xindex>, memref<?xf64>

OPERATION_NAME = 'gpu.create_bsr'¶

_ODS_REGIONS = (0, True)¶

asyncDependencies() → _ods_ir¶

brows() → _ods_ir¶

bcols() → _ods_ir¶

bnnz() → _ods_ir¶

rBlockSize() → _ods_ir¶

cBlockSize() → _ods_ir¶

bRowPos() → _ods_ir¶

bColIdxs() → _ods_ir¶

values() → _ods_ir¶

spmat() → _ods_ir¶

asyncToken() → _ods_ir | None¶

mlir.dialects.gpu.create_bsr(spmat, async_token, async_dependencies, brows, bcols, bnnz, r_block_size, c_block_size, b_row_pos, b_col_idxs, values, *, loc=None, ip=None) → _ods_ir | _ods_ir | CreateBsrOp¶

class mlir.dialects.gpu.CreateCooAoSOp(spmat, asyncToken, asyncDependencies, rows, cols, nnz, idxs, values, *, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.create_coo_aos operation initializes a sparse matrix in COO format with the given sizes from the given index and values buffers. The buffers must already be copied from the host to the device prior to using this operation. The operation returns a handle to the sparse matrix descriptor. Unlike the default gpu.create_coo operation, this operation builds the COO format from a single index buffer in AoS format (note that this feature has been deprecated in cuSparse 11.2).

Example:

%spmat, %token = gpu.create_coo_aos async [%dep] %rows, %cols, %nnz, %idxs,
    %values : memref<?xindex>, memref<?xf64>

OPERATION_NAME = 'gpu.create_coo_aos'¶

_ODS_REGIONS = (0, True)¶

asyncDependencies() → _ods_ir¶

rows() → _ods_ir¶

cols() → _ods_ir¶

nnz() → _ods_ir¶

idxs() → _ods_ir¶

values() → _ods_ir¶

spmat() → _ods_ir¶

asyncToken() → _ods_ir | None¶

mlir.dialects.gpu.create_coo_aos(spmat, async_token, async_dependencies, rows, cols, nnz, idxs, values, *, loc=None, ip=None) → _ods_ir | _ods_ir | CreateCooAoSOp¶

class mlir.dialects.gpu.CreateCooOp(spmat, asyncToken, asyncDependencies, rows, cols, nnz, rowIdxs, colIdxs, values, *, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.create_coo operation initializes a sparse matrix in COO format with the given sizes from the given index and values buffers. The buffers must already be copied from the host to the device prior to using this operation. The operation returns a handle to the sparse matrix descriptor. Note that this operation builds the COO in SoA format.

Example:

%spmat, %token = gpu.create_coo async [%dep] %rows, %cols, %nnz, %rowIdx,
    %colIdx, %values : memref<?xindex>, memref<?xindex>, memref<?xf64>

OPERATION_NAME = 'gpu.create_coo'¶

_ODS_REGIONS = (0, True)¶

asyncDependencies() → _ods_ir¶

rows() → _ods_ir¶

cols() → _ods_ir¶

nnz() → _ods_ir¶

rowIdxs() → _ods_ir¶

colIdxs() → _ods_ir¶

values() → _ods_ir¶

spmat() → _ods_ir¶

asyncToken() → _ods_ir | None¶

mlir.dialects.gpu.create_coo(spmat, async_token, async_dependencies, rows, cols, nnz, row_idxs, col_idxs, values, *, loc=None, ip=None) → _ods_ir | _ods_ir | CreateCooOp¶

class mlir.dialects.gpu.CreateCscOp(spmat, asyncToken, asyncDependencies, rows, cols, nnz, colPos, rowIdxs, values, *, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.create_csc operation initializes a sparse matrix in CSC format with the given sizes from the given position, index, and values buffers. The buffers must already be copied from the host to the device prior to using this operation. The operation returns a handle to the sparse matrix descriptor.

The CSC format has exactly the same memory layout as its transpose in CSR format (and vice versa).

Example:

%spmat, %token = gpu.create_csc async [%dep] %rows, %cols, %nnz, %colPos,
    %rowIdx, %values : memref<?xindex>, memref<?xindex>, memref<?xf64>

OPERATION_NAME = 'gpu.create_csc'¶

_ODS_REGIONS = (0, True)¶

asyncDependencies() → _ods_ir¶

rows() → _ods_ir¶

cols() → _ods_ir¶

nnz() → _ods_ir¶

colPos() → _ods_ir¶

rowIdxs() → _ods_ir¶

values() → _ods_ir¶

spmat() → _ods_ir¶

asyncToken() → _ods_ir | None¶

mlir.dialects.gpu.create_csc(spmat, async_token, async_dependencies, rows, cols, nnz, col_pos, row_idxs, values, *, loc=None, ip=None) → _ods_ir | _ods_ir | CreateCscOp¶

class mlir.dialects.gpu.CreateCsrOp(spmat, asyncToken, asyncDependencies, rows, cols, nnz, rowPos, colIdxs, values, *, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.create_csr operation initializes a sparse matrix in CSR format with the given sizes from the given position, index, and values buffers. The buffers must already be copied from the host to the device prior to using this operation. The operation returns a handle to the sparse matrix descriptor.

The CSR format has exactly the same memory layout as its transpose in CSC format (and vice versa).

Example:

%spmat, %token = gpu.create_csr async [%dep] %rows, %cols, %nnz, %rowPos,
    %colIdx, %values : memref<?xindex>, memref<?xindex>, memref<?xf64>

OPERATION_NAME = 'gpu.create_csr'¶

_ODS_REGIONS = (0, True)¶

asyncDependencies() → _ods_ir¶

rows() → _ods_ir¶

cols() → _ods_ir¶

nnz() → _ods_ir¶

rowPos() → _ods_ir¶

colIdxs() → _ods_ir¶

values() → _ods_ir¶

spmat() → _ods_ir¶

asyncToken() → _ods_ir | None¶

mlir.dialects.gpu.create_csr(spmat, async_token, async_dependencies, rows, cols, nnz, row_pos, col_idxs, values, *, loc=None, ip=None) → _ods_ir | _ods_ir | CreateCsrOp¶

class mlir.dialects.gpu.CreateDnTensorOp(dnTensor, asyncToken, asyncDependencies, memref, dims, *, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.create_dn_tensor operation initializes a dense tensor from the given values buffer and sizes. The buffer must already be copied from the host to the device prior to using this operation. The operation returns a handle to the dense tensor descriptor.

Example:

%dmat, %token = gpu.create_dn_tensor async [%dep] %mem, %dims : index, index into memref<?xf64>

OPERATION_NAME = 'gpu.create_dn_tensor'¶

_ODS_OPERAND_SEGMENTS¶

_ODS_REGIONS = (0, True)¶

asyncDependencies() → _ods_ir¶

memref() → _ods_ir¶

dims() → _ods_ir¶

dnTensor() → _ods_ir¶

asyncToken() → _ods_ir | None¶

mlir.dialects.gpu.create_dn_tensor(dn_tensor, async_token, async_dependencies, memref, dims, *, loc=None, ip=None) → _ods_ir | _ods_ir | CreateDnTensorOp¶

class mlir.dialects.gpu.DeallocOp(asyncToken, asyncDependencies, memref, *, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.dealloc operation frees the region of memory referenced by a memref which was originally created by the gpu.alloc operation. It is similar to the memref.dealloc op, but supports asynchronous GPU execution.

The op does not execute before all async dependencies have finished executing.

If the async keyword is present, the op is executed asynchronously (i.e. it does not block until the execution has finished on the device). In that case, it returns a !gpu.async.token.

Example:

%token = gpu.dealloc async [%dep] %memref : memref<8x64xf32, 1>

OPERATION_NAME = 'gpu.dealloc'¶

_ODS_REGIONS = (0, True)¶

asyncDependencies() → _ods_ir¶

memref() → _ods_ir¶

asyncToken() → _ods_ir | None¶

mlir.dialects.gpu.dealloc(async_token, async_dependencies, memref, *, loc=None, ip=None) → _ods_ir | _ods_ir | DeallocOp¶

class mlir.dialects.gpu.DestroyDnTensorOp(asyncToken, asyncDependencies, dnTensor, *, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.destroy_dn_tensor operation releases all resources of a dense tensor represented by a handle that was previously created by a gpu.create_dn_tensor operation.

Example:

%token = gpu.destroy_dn_tensor async [%dep] %dnTensor

OPERATION_NAME = 'gpu.destroy_dn_tensor'¶

_ODS_REGIONS = (0, True)¶

asyncDependencies() → _ods_ir¶

dnTensor() → _ods_ir¶

asyncToken() → _ods_ir | None¶

mlir.dialects.gpu.destroy_dn_tensor(async_token, async_dependencies, dn_tensor, *, loc=None, ip=None) → _ods_ir | _ods_ir | DestroyDnTensorOp¶

class mlir.dialects.gpu.DestroySpMatOp(asyncToken, asyncDependencies, spmat, *, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.destroy_sp_mat operation releases all resources of a sparse matrix represented by a handle that was previously created by a one of the sparse matrix creation operations.

Example:

%token = gpu.destroy_sp_mat async [%dep] %spmat

OPERATION_NAME = 'gpu.destroy_sp_mat'¶

_ODS_REGIONS = (0, True)¶

asyncDependencies() → _ods_ir¶

spmat() → _ods_ir¶

asyncToken() → _ods_ir | None¶

mlir.dialects.gpu.destroy_sp_mat(async_token, async_dependencies, spmat, *, loc=None, ip=None) → _ods_ir | _ods_ir | DestroySpMatOp¶

class mlir.dialects.gpu.DynamicSharedMemoryOp(resultMemref, *, loc=None, ip=None)¶

Bases: _ods_ir

This operation provides a memref pointer to the start of dynamic shared memory, often referred to as workgroup memory. It’s important to note that this dynamic shared memory needs to be allocated at kernel launch. One can conveniently utilize the dynamic_shared_memory_size parameter of gpu.launch for this purpose.

Examples:

%0 = gpu.dynamic.shared.memory : memref<?xi8, #gpu.address_space<workgroup>>
%1 = memref.view %0[%c8192][] : memref<?xi8, #gpu.address_space<workgroup>>
                        to memref<32x64xf32, #gpu.address_space<workgroup>>
%2 = memref.view %0[%c16384][] : memref<?xi8, #gpu.address_space<workgroup>>
                        to memref<32x64xf32, #gpu.address_space<workgroup>>

OPERATION_NAME = 'gpu.dynamic_shared_memory'¶

_ODS_REGIONS = (0, True)¶

resultMemref() → _ods_ir¶

mlir.dialects.gpu.dynamic_shared_memory(result_memref, *, loc=None, ip=None) → _ods_ir¶

class mlir.dialects.gpu.GPUFuncOp(function_type, *, arg_attrs=None, res_attrs=None, workgroup_attrib_attrs=None, private_attrib_attrs=None, known_block_size=None, known_grid_size=None, loc=None, ip=None)¶

Bases: _ods_ir

Defines a function that can be executed on a GPU. This supports memory attribution and its body has a particular execution model.

GPU functions are either kernels (as indicated by the kernel attribute) or regular functions. The former can be launched from the host side, while the latter are device side only.

The memory attribution defines SSA values that correspond to memory buffers allocated in the memory hierarchy of the GPU (see below).

The operation has one attached region that corresponds to the body of the function. The region arguments consist of the function arguments without modification, followed by buffers defined in memory annotations. The body of a GPU function, when launched, is executed by multiple work items. There are no guarantees on the order in which work items execute, or on the connection between them. In particular, work items are not necessarily executed in lock-step. Synchronization ops such as “gpu.barrier” should be used to coordinate work items. Declarations of GPU functions, i.e. not having the body region, are not supported.

A function may optionally be annotated with the block and/or grid sizes that will be used when it is launched using the known_block_size and known_grid_size attributes, respectively. If set, these attributes must be arrays of three 32-bit integers giving the x, y, and z launch dimensions. Launching a kernel that has these annotations, or that calls a function with these annotations, using a block size or grid size other than what is specified is undefined behavior. These attributes may be set on non-gpu.func functions by using gpu.known_block_size or gpu.known_grid_size, but this carries the risk that they will de discarded.

Syntax:

op ::= `gpu.func` symbol-ref-id `(` argument-list `)` (`->`
function-result-list)?
       memory-attribution `kernel`? function-attributes? region

memory-attribution ::= (`workgroup` `(` ssa-id-and-type-list `)`)?
                       (`private` `(` ssa-id-and-type-list `)`)?

Example:

gpu.func @foo(%arg0: index)
    workgroup(%workgroup: memref<32xf32, 3>)
    private(%private: memref<1xf32, 5>)
    kernel
    attributes {qux: "quux"} {
  gpu.return
}

The generic form illustrates the concept

"gpu.func"(%arg: index) {sym_name: "foo", kernel, qux: "quux"} ({
^bb0(%arg0: index, %workgroup: memref<32xf32, 3>,
     %private: memref<1xf32, 5>):
  "gpu.return"() : () -> ()
}) : (index) -> ()

Note the non-default memory spaces used in memref types in memory attribution.

OPERATION_NAME = 'gpu.func'¶

_ODS_REGIONS = (1, True)¶

function_type() → _ods_ir¶

arg_attrs() → _ods_ir | None¶

res_attrs() → _ods_ir | None¶

workgroup_attrib_attrs() → _ods_ir | None¶

private_attrib_attrs() → _ods_ir | None¶

known_block_size() → _ods_ir | None¶

known_grid_size() → _ods_ir | None¶

body() → _ods_ir¶

mlir.dialects.gpu.func(function_type, *, arg_attrs=None, res_attrs=None, workgroup_attrib_attrs=None, private_attrib_attrs=None, known_block_size=None, known_grid_size=None, loc=None, ip=None) → GPUFuncOp¶

class mlir.dialects.gpu.GPUModuleOp(sym_name, *, targets=None, offloadingHandler=None, loc=None, ip=None)¶

Bases: _ods_ir

GPU module contains code that is intended to be run on a GPU. A host device can launch this code through a gpu.launc_func that creates a fully qualified symbol through the gpu.module’s symbol and a gpu.func symbol contained in the gpu.module.

The module’s top-level scope is modeled by a single region with a single block. GPU modules are required to have a name that is used for symbol resolution by the gpu.launch_func operation.

Using an op with a region to define a GPU module enables “embedding” GPU modules with SIMT execution models in other dialects in a clean manner and allows filtering of code regions to execute passes on only code intended to or not intended to be run on the separate device.

Modules can contain zero or more target attributes. These attributes encode how to transform modules into binary strings and are used by the gpu-module-to-binary pass to transform modules into GPU binaries.

Modules can contain an optional OffloadingTranslationAttr attribute. This attribute will be used during the gpu-module-to-binary pass to specify the OffloadingTranslationAttr used when creating the gpu.binary operation.

gpu.module @symbol_name {
  gpu.func {}
    ...
}
// Module with offloading handler and target attributes.
gpu.module @symbol_name2 <#gpu.select_object<1>> [
    #nvvm.target,
    #rocdl.target<chip = "gfx90a">] {
  gpu.func {}
    ...
}

OPERATION_NAME = 'gpu.module'¶

_ODS_REGIONS = (1, True)¶

sym_name() → _ods_ir¶

targets() → _ods_ir | None¶

offloadingHandler() → _ods_ir | None¶

bodyRegion() → _ods_ir¶

mlir.dialects.gpu.module(sym_name, *, targets=None, offloading_handler=None, loc=None, ip=None) → GPUModuleOp¶

class mlir.dialects.gpu.GlobalIdOp(dimension, *, upper_bound=None, results=None, loc=None, ip=None)¶

Bases: _ods_ir

Returns the unique global workitem/thread id, i.e., the unique index of the current workitem/thread within all workgroups / grid along the x, y, or z dimension.

Example:

%gidX = gpu.global_id x
%gidX = gpu.global_id x upper_bound 65536

The upper_bound attribute defines an upper bound analogously to the ones on thread_id and block_id. If one is not set, the bound may be inferred from a combination of known_block_size and known_grid_size-type annotations.

OPERATION_NAME = 'gpu.global_id'¶

_ODS_REGIONS = (0, True)¶

dimension() → _ods_ir¶

upper_bound() → _ods_ir | None¶

mlir.dialects.gpu.global_id(dimension, *, upper_bound=None, results=None, loc=None, ip=None) → _ods_ir¶

class mlir.dialects.gpu.GridDimOp(dimension, *, upper_bound=None, results=None, loc=None, ip=None)¶

Bases: _ods_ir

Returns the number of thread blocks in the grid along the x, y, or z dimension.

Example:

%gDimZ = gpu.grid_dim z

If known_grid_size is set on an this operation’s enclosing gpu.func, or gpu.known_grid_size is set on an enclosing FunctionOpInterface implementor, or if the enclosing gpu.launch specifies a constant size for dimension’s grid length, these contextual facts may be used to infer that this operation has a constant value, though such a transformation will not be performed by canonicalization or the default constant folder. Executions which cause that constant-value assumption to be false incur undefined behavior.

If upper_bound is set, executions where the grid size in dimension would exceed upper_bound cause undefined behavior.

There is an implicit upper bound of kMaxDim (currently uint32_t::max).

OPERATION_NAME = 'gpu.grid_dim'¶

_ODS_REGIONS = (0, True)¶

dimension() → _ods_ir¶

upper_bound() → _ods_ir | None¶

mlir.dialects.gpu.grid_dim(dimension, *, upper_bound=None, results=None, loc=None, ip=None) → _ods_ir¶

class mlir.dialects.gpu.HostRegisterOp(value, *, loc=None, ip=None)¶

Bases: _ods_ir

This op maps the provided host buffer into the device address space.

This operation may not be supported in every environment, there is not yet a way to check at runtime whether this feature is supported.

Writes from the host are guaranteed to be visible to device kernels that are launched afterwards. Writes from the device are guaranteed to be visible on the host after synchronizing with the device kernel completion.

OPERATION_NAME = 'gpu.host_register'¶

_ODS_REGIONS = (0, True)¶

value() → _ods_ir¶

mlir.dialects.gpu.host_register(value, *, loc=None, ip=None) → HostRegisterOp¶

class mlir.dialects.gpu.HostUnregisterOp(value, *, loc=None, ip=None)¶

Bases: _ods_ir

This op unmaps the provided host buffer from the device address space.

This operation may not be supported in every environment, there is not yet a way to check at runtime whether this feature is supported.

OPERATION_NAME = 'gpu.host_unregister'¶

_ODS_REGIONS = (0, True)¶

value() → _ods_ir¶

mlir.dialects.gpu.host_unregister(value, *, loc=None, ip=None) → HostUnregisterOp¶

class mlir.dialects.gpu.LaneIdOp(*, upper_bound=None, results=None, loc=None, ip=None)¶

Bases: _ods_ir

Returns the lane id within the subgroup (warp/wave).

Example:

%laneId = gpu.lane_id

If upper_bound is set, executions with more than upper_bound lanes per subgroup cause undefined behavior. In the abscence of upper_bound, the lane id is still assumed to be non-negative and less than the target-independent kMaxSubgroupSize (currently 128).

OPERATION_NAME = 'gpu.lane_id'¶

_ODS_REGIONS = (0, True)¶

upper_bound() → _ods_ir | None¶

result() → _ods_ir¶: Shortcut to get an op result if it has only one (throws an error otherwise).

mlir.dialects.gpu.lane_id(*, upper_bound=None, results=None, loc=None, ip=None) → _ods_ir¶

class mlir.dialects.gpu.LaunchFuncOp(asyncToken, asyncDependencies, kernel, gridSizeX, gridSizeY, gridSizeZ, blockSizeX, blockSizeY, blockSizeZ, kernelOperands, *, clusterSizeX=None, clusterSizeY=None, clusterSizeZ=None, dynamicSharedMemorySize=None, asyncObject=None, loc=None, ip=None)¶

Bases: _ods_ir

Launch a kernel function on the specified grid of thread blocks. gpu.launch operations are lowered to gpu.launch_func operations by outlining the kernel body into a function in a dedicated module, which reflects the separate compilation process. The kernel function is required to have the gpu.kernel attribute. The module containing the kernel function is required to be a gpu.module. And finally, the module containing the kernel module (which thus cannot be the top-level module) is required to have the gpu.container_module attribute. The gpu.launch_func operation has a symbol attribute named kernel to identify the fully specified kernel function to launch (both the gpu.module and func).

The gpu.launch_func supports async dependencies: the kernel does not start executing until the ops producing those async dependencies have completed.

By the default, the host implicitly blocks until kernel execution has completed. If the async keyword is present, the host does not block but instead a !gpu.async.token is returned. Other async GPU ops can take this token as dependency.

The operation requires at least the grid and block sizes along the x,y,z dimensions as arguments. When a lower-dimensional kernel is required, unused sizes must be explicitly set to 1.

The remaining operands are optional. The first optional operand corresponds to the amount of dynamic shared memory a kernel’s workgroup should be allocated; when this operand is not present, a zero size is assumed.

The remaining operands if present are passed as arguments to the kernel function.

The gpu.launch_func also supports kernel launching with clusters if supported by the target architecture. The cluster size can be set by clusterSizeX, clusterSizeY, and clusterSizeZ arguments. When these arguments are present, the Op launches a kernel that clusters the given thread blocks. This feature is exclusive to certain architectures.

Example:

module attributes {gpu.container_module} {

  // This module creates a separate compilation unit for the GPU compiler.
  gpu.module @kernels {
    func.func @kernel_1(%arg0 : f32, %arg1 : memref<?xf32, 1>)
        attributes { nvvm.kernel = true } {

      // Operations that produce block/thread IDs and dimensions are
      // injected when outlining the `gpu.launch` body to a function called
      // by `gpu.launch_func`.
      %tIdX = gpu.thread_id x
      %tIdY = gpu.thread_id y
      %tIdZ = gpu.thread_id z

      %bDimX = gpu.block_dim x
      %bDimY = gpu.block_dim y
      %bDimZ = gpu.block_dim z

      %bIdX = gpu.block_id x
      %bIdY = gpu.block_id y
      %bIdZ = gpu.block_id z

      %gDimX = gpu.grid_dim x
      %gDimY = gpu.grid_dim y
      %gDimZ = gpu.grid_dim z

      // (Optional)  Cluster size only for support architectures
      %cIdX = gpu.cluster_id x
      %cIdY = gpu.cluster_id y
      %cIdZ = gpu.cluster_id z

      %cDimX = gpu.cluster_dim x
      %cDimY = gpu.cluster_dim y
      %cDimZ = gpu.cluster_dim z

      "some_op"(%bx, %tx) : (index, index) -> ()
      %42 = load %arg1[%bx] : memref<?xf32, 1>
    }
  }

  %t0 = gpu.wait async
  gpu.launch_func
      async                           // (Optional) Don't block host, return token.
      [%t0]                           // (Optional) Execute only after %t0 has completed.
      @kernels::@kernel_1             // Kernel function.
      clusters in (%cst, %cst, %cst)  // (Optional) Cluster size only for support architectures.
      blocks in (%cst, %cst, %cst)    // Grid size.
      threads in (%cst, %cst, %cst)   // Block size.
      dynamic_shared_memory_size %s   // (Optional) Amount of dynamic shared
                                      // memory to allocate for a workgroup.
      args(%arg0 : f32,               // (Optional) Kernel arguments.
           %arg1 : memref<?xf32, 1>)
}

OPERATION_NAME = 'gpu.launch_func'¶

_ODS_OPERAND_SEGMENTS¶

_ODS_REGIONS = (0, True)¶

asyncDependencies() → _ods_ir¶

gridSizeX() → _ods_ir¶

gridSizeY() → _ods_ir¶

gridSizeZ() → _ods_ir¶

blockSizeX() → _ods_ir¶

blockSizeY() → _ods_ir¶

blockSizeZ() → _ods_ir¶

clusterSizeX() → _ods_ir | None¶

clusterSizeY() → _ods_ir | None¶

clusterSizeZ() → _ods_ir | None¶

dynamicSharedMemorySize() → _ods_ir | None¶

kernelOperands() → _ods_ir¶

asyncObject() → _ods_ir | None¶

kernel() → _ods_ir¶

asyncToken() → _ods_ir | None¶

mlir.dialects.gpu.launch_func(async_token, async_dependencies, kernel, grid_size_x, grid_size_y, grid_size_z, block_size_x, block_size_y, block_size_z, kernel_operands, *, cluster_size_x=None, cluster_size_y=None, cluster_size_z=None, dynamic_shared_memory_size=None, async_object=None, loc=None, ip=None) → _ods_ir | _ods_ir | LaunchFuncOp¶

class mlir.dialects.gpu.LaunchOp(asyncToken, asyncDependencies, gridSizeX, gridSizeY, gridSizeZ, blockSizeX, blockSizeY, blockSizeZ, *, clusterSizeX=None, clusterSizeY=None, clusterSizeZ=None, dynamicSharedMemorySize=None, module=None, function=None, loc=None, ip=None)¶

Bases: _ods_ir

Launch a kernel on the specified grid of thread blocks. The body of the kernel is defined by the single region that this operation contains. The operation takes an optional list of async dependencies followed by six operands and an optional operand.

The async keyword indicates the kernel should be launched asynchronously; the operation returns a new !gpu.async.token when the keyword is specified. The kernel launched does not start executing until the ops producing its async dependencies (optional operands) have completed.

The first three operands (following any async dependencies) are grid sizes along the x,y,z dimensions and the following three are block sizes along the x,y,z dimensions. When a lower-dimensional kernel is required, unused sizes must be explicitly set to 1. The last operand is optional and corresponds to the amount of dynamic shared memory a kernel’s workgroup should be allocated; when this operand is not present, a zero size is assumed.

The body region has at least twelve arguments, or eighteen if cluster dimensions are present, grouped as follows:

three optional arguments that contain cluster identifiers along x,y,z

dimensions; * three arguments that contain block identifiers along x,y,z dimensions; * three arguments that contain thread identifiers along x,y,z dimensions; * operands of the gpu.launch operation as is (i.e. the operands for grid and block sizes). * a variadic number of Workgroup memory attributions. * a variadic number of Private memory attributions.

The function and module attributes are optional and specifies the kernel name and a module in which the kernel should be outlined.

Syntax:

operation ::= `gpu.launch` (`async` (`[` ssa-id-list `]`)? )?
                         ( `clusters` `(` ssa-id-list `)` `in` ssa-reassignment )?
                         `blocks` `(` ssa-id-list `)` `in` ssa-reassignment
                         `threads` `(` ssa-id-list `)` `in` ssa-reassignment
                         (dynamic_shared_memory_size ssa-use)?
                         (`module(` symbol-ref-id `)`)?
                         (`function(` symbol-ref-id `)`)?
                         memory-attribution
                         region attr-dict?
ssa-reassignment ::= `(` ssa-id `=` ssa-use (`,` ssa-id `=` ssa-use)* `)`
memory-attribution ::= (`workgroup` `(` ssa-id-and-type-list `)`)?
                       (`private` `(` ssa-id-and-type-list `)`)?

Example:

gpu.launch blocks(%bx, %by, %bz) in (%sz_bx = %0, %sz_by = %1, %sz_bz = %2)
           threads(%tx, %ty, %tz) in (%sz_tx = %3, %sz_ty = %4, %sz_tz = %5) {
  // Block and thread identifiers, as well as block/grid sizes are
  // immediately usable inside body region.
  "some_op"(%bx, %tx) : (index, index) -> ()
  // Assuming %val1 is defined outside the gpu.launch region.
  %42 = load %val1[%bx] : memref<?xf32, 1>
}

// Generic syntax explains how the pretty syntax maps to the IR structure.
"gpu.launch"(%cst, %cst, %c1,  // Grid sizes.
             %cst, %c1, %c1)   // Block sizes.

    {/*attributes*/}
    // All sizes and identifiers have "index" size.
    : (index, index, index, index, index, index) -> () {
// The operation passes block and thread identifiers, followed by grid and
// block sizes.
^bb0(%bx : index, %by : index, %bz : index,
     %tx : index, %ty : index, %tz : index,
     %num_bx : index, %num_by : index, %num_bz : index,
     %num_tx : index, %num_ty : index, %num_tz : index)
  "some_op"(%bx, %tx) : (index, index) -> ()
  %3 = "memref.load"(%val1, %bx) : (memref<?xf32, 1>, index) -> f32
}

// Launch with memory attributions.
gpu.launch blocks(%bx, %by, %bz) in (%sz_bx = %0, %sz_by = %1, %sz_bz = %2)
           threads(%tx, %ty, %tz) in (%sz_tx = %3, %sz_ty = %4, %sz_tz = %5)
           workgroup(%workgroup: memref<32xf32, 3>)
           private(%private: memref<1xf32, 5>) {
  // Block and thread identifiers, as well as block/grid sizes are
  // immediately usable inside body region.
  "some_op"(%bx, %tx) : (index, index) -> ()
  // Assuming %val1 is defined outside the gpu.launch region.
  %42 = load %workgroup[%bx] : memref<32xf32, 3>
}

// Launch with clusters.
gpu.launch clusters(%cx, %cy, %cz) in (%sz_cx = %0, %sz_cy = %1, %sz_cz = %2)
           blocks(%bx, %by, %bz) in (%sz_bx = %3, %sz_by = %4, %sz_bz = %5)
           threads(%tx, %ty, %tz) in (%sz_tx = %6, %sz_ty = %7, %sz_tz = %8)
{
  // Cluster, block and thread identifiers, as well as cluster/block/grid
  // sizes are immediately usable inside body region.
  "some_op"(%cx, %bx, %tx) : (index, index, index) -> ()
}

// Launch with module and function attributes.
gpu.launch blocks(%bx, %by, %bz) in (%sz_bx = %0, %sz_by = %1, %sz_bz = %2)
           threads(%tx, %ty, %tz) in (%sz_tx = %3, %sz_ty = %4, %sz_tz = %5)
           module(@kernel_module) function(@kernel_func) {
  "some_op"(%bx, %tx) : (index, index) -> ()
  %42 = load %val1[%bx] : memref<?xf32, 1>
}

Rationale: using operation/block arguments gives analyses a clear way of understanding that a value has additional semantics (e.g., we will need to know what value corresponds to threadIdx.x for coalescing). We can recover these properties by analyzing the operations producing values, but it is easier just to have that information by construction.

OPERATION_NAME = 'gpu.launch'¶

_ODS_OPERAND_SEGMENTS¶

_ODS_REGIONS = (1, True)¶

asyncDependencies() → _ods_ir¶

gridSizeX() → _ods_ir¶

gridSizeY() → _ods_ir¶

gridSizeZ() → _ods_ir¶

blockSizeX() → _ods_ir¶

blockSizeY() → _ods_ir¶

blockSizeZ() → _ods_ir¶

clusterSizeX() → _ods_ir | None¶

clusterSizeY() → _ods_ir | None¶

clusterSizeZ() → _ods_ir | None¶

dynamicSharedMemorySize() → _ods_ir | None¶

module() → _ods_ir | None¶

function() → _ods_ir | None¶

asyncToken() → _ods_ir | None¶

body() → _ods_ir¶

mlir.dialects.gpu.launch(async_token, async_dependencies, grid_size_x, grid_size_y, grid_size_z, block_size_x, block_size_y, block_size_z, *, cluster_size_x=None, cluster_size_y=None, cluster_size_z=None, dynamic_shared_memory_size=None, module=None, function=None, loc=None, ip=None) → _ods_ir | _ods_ir | LaunchOp¶

class mlir.dialects.gpu.MemcpyOp(asyncToken, asyncDependencies, dst, src, *, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.memcpy operation copies the content of one memref to another.

The op does not execute before all async dependencies have finished executing.

If the async keyword is present, the op is executed asynchronously (i.e. it does not block until the execution has finished on the device). In that case, it returns a !gpu.async.token.

Example:

%token = gpu.memcpy async [%dep] %dst, %src : memref<?xf32, 1>, memref<?xf32>

OPERATION_NAME = 'gpu.memcpy'¶

_ODS_REGIONS = (0, True)¶

asyncDependencies() → _ods_ir¶

dst() → _ods_ir¶

src() → _ods_ir¶

asyncToken() → _ods_ir | None¶

mlir.dialects.gpu.memcpy(async_token, async_dependencies, dst, src, *, loc=None, ip=None) → _ods_ir | _ods_ir | MemcpyOp¶

class mlir.dialects.gpu.MemsetOp(asyncToken, asyncDependencies, dst, value, *, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.memset operation sets the content of memref to a scalar value.

The op does not execute before all async dependencies have finished executing.

If the async keyword is present, the op is executed asynchronously (i.e. it does not block until the execution has finished on the device). In that case, it returns a !gpu.async.token.

Example:

%token = gpu.memset async [%dep] %dst, %value : memref<?xf32, 1>, f32

OPERATION_NAME = 'gpu.memset'¶

_ODS_REGIONS = (0, True)¶

asyncDependencies() → _ods_ir¶

dst() → _ods_ir¶

value() → _ods_ir¶

asyncToken() → _ods_ir | None¶

mlir.dialects.gpu.memset(async_token, async_dependencies, dst, value, *, loc=None, ip=None) → _ods_ir | _ods_ir | MemsetOp¶

class mlir.dialects.gpu.NumSubgroupsOp(*, upper_bound=None, results=None, loc=None, ip=None)¶

Bases: _ods_ir

Returns the number of subgroups within a workgroup.

Example:

%numSg = gpu.num_subgroups : index

If upper_bound is set, executions with more than upper_bound subgroups per workgroup cause undefined behavior. There is a default upper bound of kMaxDim (currently uint32_t::max).

OPERATION_NAME = 'gpu.num_subgroups'¶

_ODS_REGIONS = (0, True)¶

upper_bound() → _ods_ir | None¶

result() → _ods_ir¶: Shortcut to get an op result if it has only one (throws an error otherwise).

mlir.dialects.gpu.num_subgroups(*, upper_bound=None, results=None, loc=None, ip=None) → _ods_ir¶

class mlir.dialects.gpu.PrintfOp(format, args, *, loc=None, ip=None)¶

Bases: _ods_ir

gpu.printf takes a literal format string format and an arbitrary number of scalar arguments that should be printed.

The format string is a C-style printf string, subject to any restrictions imposed by one’s target platform.

OPERATION_NAME = 'gpu.printf'¶

_ODS_REGIONS = (0, True)¶

args() → _ods_ir¶

format() → _ods_ir¶

mlir.dialects.gpu.printf(format, args, *, loc=None, ip=None) → PrintfOp¶

class mlir.dialects.gpu.ReturnOp(operands_, *, loc=None, ip=None)¶

Bases: _ods_ir

A terminator operation for regions that appear in the body of gpu.func functions. The operands to the gpu.return are the result values returned by an invocation of the gpu.func.

OPERATION_NAME = 'gpu.return'¶

_ODS_REGIONS = (0, True)¶

operands_() → _ods_ir¶

mlir.dialects.gpu.return_(operands_, *, loc=None, ip=None) → ReturnOp¶

class mlir.dialects.gpu.RotateOp(value, offset, width, *, results=None, loc=None, ip=None)¶

Bases: _ods_ir

The “rotate” op moves values across lanes in a subgroup (a.k.a., local invocations) within the same subgroup. The width attribute specifies the number of lanes that participate in the rotation, and must be uniform across all participating lanes. Further, the first width lanes of the subgroup must be active.

width must be a power of two, and offset must be in the range [0, width).

Return the rotateResult of the invocation whose id within the group is calculated as follows:

Invocation ID = ((LaneId + offset) & (width - 1)) + (LaneId & ~(width - 1))

Returns the rotateResult and true if the current lane id is smaller than width, and poison value and false otherwise.

example:

%1, %2 = gpu.rotate %0, 1, 16 : f32

For lane k, returns the value from lane (k + cst1) % width.

OPERATION_NAME = 'gpu.rotate'¶

_ODS_REGIONS = (0, True)¶

value() → _ods_ir¶

offset() → _ods_ir¶

width() → _ods_ir¶

rotateResult() → _ods_ir¶

valid() → _ods_ir¶

mlir.dialects.gpu.rotate(value, offset, width, *, results=None, loc=None, ip=None) → _ods_ir¶

class mlir.dialects.gpu.SDDMMBufferSizeOp(bufferSz, asyncToken, asyncDependencies, dnmatA, dnmatB, spmatC, computeType, *, modeA=None, modeB=None, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.sddmm_buffer_size operation returns the buffer size required to perform the SDDMM operation on the given sparse and dense matrices. The operation expects handles returned by previous sparse operations to construct an environment and the operands for SDDMM.

Example:

%buffersz, %token = gpu.sddmm_buffer_size async [%dep] %dnmatA{TRANSPOSE}, %dnmatB{TRANSPOSE}, %spmatC into f32

The matrix arguments can also be associated with one of the following operators: NON_TRANSPOSE, TRANSPOSE, CONJUGATE_TRANSPOSE. The default value is NON_TRANSPOSE.

OPERATION_NAME = 'gpu.sddmm_buffer_size'¶

_ODS_REGIONS = (0, True)¶

asyncDependencies() → _ods_ir¶

dnmatA() → _ods_ir¶

dnmatB() → _ods_ir¶

spmatC() → _ods_ir¶

modeA() → _ods_ir¶

modeB() → _ods_ir¶

computeType() → _ods_ir¶

bufferSz() → _ods_ir¶

asyncToken() → _ods_ir | None¶

mlir.dialects.gpu.sddmm_buffer_size(buffer_sz, async_token, async_dependencies, dnmat_a, dnmat_b, spmat_c, compute_type, *, mode_a=None, mode_b=None, loc=None, ip=None) → _ods_ir | _ods_ir | SDDMMBufferSizeOp¶

class mlir.dialects.gpu.SDDMMOp(asyncToken, asyncDependencies, dnmatA, dnmatB, spmatC, computeType, buffer, *, modeA=None, modeB=None, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.sddmm operation performs the SDDMM operation on the given sparse and dense matrices, and buffer. The operation expects handles returned by previous sparse operations to construct an environment and the operands for SDDMM. The buffer must have been allocated on the device.

Example:

%token = gpu.sddmm async [%dep] %dnmatA{TRANSPOSE}, %dnmatB{TRANSPOSE}, %spmatC, %buffer into f32

The matrix arguments can also be associated with one of the following operators: NON_TRANSPOSE, TRANSPOSE, CONJUGATE_TRANSPOSE. The default value is NON_TRANSPOSE.

OPERATION_NAME = 'gpu.sddmm'¶

_ODS_REGIONS = (0, True)¶

asyncDependencies() → _ods_ir¶

dnmatA() → _ods_ir¶

dnmatB() → _ods_ir¶

spmatC() → _ods_ir¶

buffer() → _ods_ir¶

modeA() → _ods_ir¶

modeB() → _ods_ir¶

computeType() → _ods_ir¶

asyncToken() → _ods_ir | None¶

mlir.dialects.gpu.sddmm(async_token, async_dependencies, dnmat_a, dnmat_b, spmat_c, compute_type, buffer, *, mode_a=None, mode_b=None, loc=None, ip=None) → _ods_ir | _ods_ir | SDDMMOp¶

class mlir.dialects.gpu.SetCsrPointersOp(asyncToken, asyncDependencies, spmat, positions, coordinates, values, *, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.set_csr_pointers assigns the given positions, coordinates, and values buffer that reside on the device directly to the given sparse matrix descriptor in csr format.

If the async keyword is present, the op is executed asynchronously (i.e. it does not block until the execution has finished on the device). In that case, it returns a !gpu.async.token in addition to the environment.

Example:

%token = gpu.set_csr_pointers async [%dep] %positions, %coordinates, %values
      : memref<?xf32>, memref<?xindex>, memref<?xindex>

OPERATION_NAME = 'gpu.set_csr_pointers'¶

_ODS_REGIONS = (0, True)¶

asyncDependencies() → _ods_ir¶

spmat() → _ods_ir¶

positions() → _ods_ir¶

coordinates() → _ods_ir¶

values() → _ods_ir¶

asyncToken() → _ods_ir | None¶

mlir.dialects.gpu.set_csr_pointers(async_token, async_dependencies, spmat, positions, coordinates, values, *, loc=None, ip=None) → _ods_ir | _ods_ir | SetCsrPointersOp¶

class mlir.dialects.gpu.SetDefaultDeviceOp(devIndex, *, loc=None, ip=None)¶

Bases: _ods_ir

Operation that sets the current default GPU, using a zero-based index into the set of GPUs on the system. The default GPU setting may be thread-local.

OPERATION_NAME = 'gpu.set_default_device'¶

_ODS_REGIONS = (0, True)¶

devIndex() → _ods_ir¶

mlir.dialects.gpu.set_default_device(dev_index, *, loc=None, ip=None) → SetDefaultDeviceOp¶

class mlir.dialects.gpu.ShuffleOp(value, offset, width, mode, *, results=None, loc=None, ip=None)¶

Bases: _ods_ir

The “shuffle” op moves values across lanes in a subgroup (a.k.a., local invocation) within the same subgroup. The width argument specifies the number of lanes that participate in the shuffle, and must be uniform across all lanes. Further, the first width lanes of the subgroup must be active.

The intepretation of the offset arguments depends on the selected mode.

Returns the shuffleResult and true if the current lane id is smaller than width, and an unspecified value and false otherwise.

xor example:

%1, %2 = gpu.shuffle xor %0, %offset, %width : f32

For lane k, returns the value %0 from lane k ^ offset. Every lane trades value with exactly one other lane.

down example:

%cst1 = arith.constant 1 : i32
%3, %4 = gpu.shuffle down %0, %cst1, %width : f32

For lane k, returns the value from lane (k + cst1). If (k + cst1) is bigger than or equal to width, the value is poison and valid is false.

up example:

%cst1 = arith.constant 1 : i32
%5, %6 = gpu.shuffle up %0, %cst1, %width : f32

For lane k, returns the value from lane (k - cst1). If (k - cst1) is smaller than 0, the value is poison and valid is false.

idx example:

%cst0 = arith.constant 0 : i32
%7, %8 = gpu.shuffle idx %0, %cst0, %width : f32

Broadcasts the value from lane 0 to all lanes.

OPERATION_NAME = 'gpu.shuffle'¶

_ODS_REGIONS = (0, True)¶

value() → _ods_ir¶

offset() → _ods_ir¶

width() → _ods_ir¶

mode() → _ods_ir¶

shuffleResult() → _ods_ir¶

valid() → _ods_ir¶

mlir.dialects.gpu.shuffle(value, offset, width, mode, *, results=None, loc=None, ip=None) → _ods_ir¶

class mlir.dialects.gpu.SpGEMMCopyOp(asyncToken, asyncDependencies, desc, spmatA, spmatB, spmatC, computeType, *, modeA=None, modeB=None, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.spgemm_copy operation copies the sparse matrix result of a SpGEMM computation.

Example:

gpu.spgemm_copy %spmatA, %spmatB, %spmatC, %spgemmDesc: f32

The matrix arguments can also be associated with one of the following operators: NON_TRANSPOSE, TRANSPOSE, CONJUGATE_TRANSPOSE. The default value is NON_TRANSPOSE.

OPERATION_NAME = 'gpu.spgemm_copy'¶

_ODS_REGIONS = (0, True)¶

asyncDependencies() → _ods_ir¶

desc() → _ods_ir¶

spmatA() → _ods_ir¶

spmatB() → _ods_ir¶

spmatC() → _ods_ir¶

modeA() → _ods_ir¶

modeB() → _ods_ir¶

computeType() → _ods_ir¶

asyncToken() → _ods_ir | None¶

mlir.dialects.gpu.spgemm_copy(async_token, async_dependencies, desc, spmat_a, spmat_b, spmat_c, compute_type, *, mode_a=None, mode_b=None, loc=None, ip=None) → _ods_ir | _ods_ir | SpGEMMCopyOp¶

class mlir.dialects.gpu.SpGEMMCreateDescrOp(desc, asyncToken, asyncDependencies, *, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.spgemm_create_descr creates a descriptor for the SpGEMM operation. The descriptor describes the SpGEMM operation and stores the internal data throughout the computation. It needs to be passed as an argument to spgemm_* operations.

Example:

%desc, %token = gpu.spgemm_create_descr async [%dep]

OPERATION_NAME = 'gpu.spgemm_create_descr'¶

_ODS_REGIONS = (0, True)¶

asyncDependencies() → _ods_ir¶

desc() → _ods_ir¶

asyncToken() → _ods_ir | None¶

mlir.dialects.gpu.spgemm_create_descr(desc, async_token, async_dependencies, *, loc=None, ip=None) → _ods_ir | _ods_ir | SpGEMMCreateDescrOp¶

class mlir.dialects.gpu.SpGEMMDestroyDescrOp(asyncToken, asyncDependencies, desc, *, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.spgemm_destroy_descr destroys the SpGEMM operation descriptor.

Example:

%token = gpu.spgemm_destroy_descr async [%dep] %desc

OPERATION_NAME = 'gpu.spgemm_destroy_descr'¶

_ODS_REGIONS = (0, True)¶

asyncDependencies() → _ods_ir¶

desc() → _ods_ir¶

asyncToken() → _ods_ir | None¶

mlir.dialects.gpu.spgemm_destroy_descr(async_token, async_dependencies, desc, *, loc=None, ip=None) → _ods_ir | _ods_ir | SpGEMMDestroyDescrOp¶

class mlir.dialects.gpu.SpGEMMWorkEstimationOrComputeOp(bufferSzNew, asyncToken, asyncDependencies, desc, spmatA, spmatB, spmatC, computeType, bufferSz, buffer, kind, *, modeA=None, modeB=None, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.spgemm_work_estimation_or_compute is used to call cusparseSpGEMM_workEstimation or cusparseSpGEMM_compute. Both of them are for both determining the buffer size and performing the actual computation. The operation expects handles returned by previous sparse operations to construct an environment and the operands for SpGEMM. The buffer must have been allocated on the device.

C’ = alpha * op(A) * op(B) + beta * C

Example:

%bufferSz, %token = gpu.spgemm_work_estimation_or_compute async [%dep] {COMPUTE}
                      %desc, %spmatA{NON_TRANSPOSE}, %spmatB{NON_TRANSPOSE},
                      %spmatC, %spgemmDesc, %c0, %alloc: f32 into
                      memref<0xi8>

The matrix arguments can also be associated with one of the following operators: NON_TRANSPOSE, TRANSPOSE, CONJUGATE_TRANSPOSE. The default value is NON_TRANSPOSE.

OPERATION_NAME = 'gpu.spgemm_work_estimation_or_compute'¶

_ODS_REGIONS = (0, True)¶

asyncDependencies() → _ods_ir¶

desc() → _ods_ir¶

spmatA() → _ods_ir¶

spmatB() → _ods_ir¶

spmatC() → _ods_ir¶

bufferSz() → _ods_ir¶

buffer() → _ods_ir¶

modeA() → _ods_ir¶

modeB() → _ods_ir¶

computeType() → _ods_ir¶

kind() → _ods_ir¶

bufferSzNew() → _ods_ir¶

asyncToken() → _ods_ir | None¶

mlir.dialects.gpu.spgemm_work_estimation_or_compute(buffer_sz_new, async_token, async_dependencies, desc, spmat_a, spmat_b, spmat_c, compute_type, buffer_sz, buffer, kind, *, mode_a=None, mode_b=None, loc=None, ip=None) → _ods_ir | _ods_ir | SpGEMMWorkEstimationOrComputeOp¶

class mlir.dialects.gpu.SpMMBufferSizeOp(bufferSzs, asyncToken, asyncDependencies, spmatA, dnmatB, dnmatC, computeType, *, modeA=None, modeB=None, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.spmm_buffer_size operation returns the buffer size required to perform the SpMM operation on the given sparse and dense matrix. The operation expects handles returned by previous sparse operations to construct an environment and the operands for SpMM.

The matrix arguments can also be associated with one of the following operators: NON_TRANSPOSE, TRANSPOSE, CONJUGATE_TRANSPOSE. The default value is NON_TRANSPOSE.

Example:

%bufferszs, %token = gpu.spmm_buffer_size async [%dep] %spmatA{TRANSPOSE}, %dnmatB{TRANSPOSE}, %dnmatC : i64 into f32

OPERATION_NAME = 'gpu.spmm_buffer_size'¶

_ODS_RESULT_SEGMENTS¶

_ODS_REGIONS = (0, True)¶

asyncDependencies() → _ods_ir¶

spmatA() → _ods_ir¶

dnmatB() → _ods_ir¶

dnmatC() → _ods_ir¶

modeA() → _ods_ir¶

modeB() → _ods_ir¶

computeType() → _ods_ir¶

bufferSzs() → _ods_ir¶

asyncToken() → _ods_ir | None¶

mlir.dialects.gpu.spmm_buffer_size(buffer_szs, async_token, async_dependencies, spmat_a, dnmat_b, dnmat_c, compute_type, *, mode_a=None, mode_b=None, loc=None, ip=None) → _ods_ir | _ods_ir | SpMMBufferSizeOp¶

class mlir.dialects.gpu.SpMMOp(asyncToken, asyncDependencies, spmatA, dnmatB, dnmatC, computeType, buffers, *, modeA=None, modeB=None, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.spmm operation performs the SpMM operation on the given sparse and dense matrix, and buffer. The operation expects handles returned by previous sparse operations to construct an environment and the operands for SpMM. The buffer must have been allocated on the device.

The matrix arguments can also be associated with one of the following operators: NON_TRANSPOSE, TRANSPOSE, CONJUGATE_TRANSPOSE. The default value is NON_TRANSPOSE.

Example:

%token = gpu.spmm async [%dep] %spmatA{TRANSPOSE}, %dnmatB{TRANSPOSE}, %dnmatC, %buffers : type($buffers) into f32

OPERATION_NAME = 'gpu.spmm'¶

_ODS_OPERAND_SEGMENTS¶

_ODS_REGIONS = (0, True)¶

asyncDependencies() → _ods_ir¶

spmatA() → _ods_ir¶

dnmatB() → _ods_ir¶

dnmatC() → _ods_ir¶

buffers() → _ods_ir¶

modeA() → _ods_ir¶

modeB() → _ods_ir¶

computeType() → _ods_ir¶

asyncToken() → _ods_ir | None¶

mlir.dialects.gpu.spmm(async_token, async_dependencies, spmat_a, dnmat_b, dnmat_c, compute_type, buffers, *, mode_a=None, mode_b=None, loc=None, ip=None) → _ods_ir | _ods_ir | SpMMOp¶

class mlir.dialects.gpu.SpMVBufferSizeOp(bufferSz, asyncToken, asyncDependencies, spmatA, dnX, dnY, computeType, *, modeA=None, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.spmv_buffer_size operation returns the buffer size required to perform the SpMV operation on the given sparse matrix and dense vectors. The operation expects handles returned by previous sparse operations to construct an environment and the operands for SpMV.

The matrix arguments can also be associated with one of the following operators: NON_TRANSPOSE, TRANSPOSE, CONJUGATE_TRANSPOSE. The default value is NON_TRANSPOSE.

Example:

%buffersz, %token = gpu.spmv_buffer_size async [%dep] %spmatA{TRANSPOSE}, %dnX, %dnY into f32

OPERATION_NAME = 'gpu.spmv_buffer_size'¶

_ODS_REGIONS = (0, True)¶

asyncDependencies() → _ods_ir¶

spmatA() → _ods_ir¶

dnX() → _ods_ir¶

dnY() → _ods_ir¶

modeA() → _ods_ir¶

computeType() → _ods_ir¶

bufferSz() → _ods_ir¶

asyncToken() → _ods_ir | None¶

mlir.dialects.gpu.spmv_buffer_size(buffer_sz, async_token, async_dependencies, spmat_a, dn_x, dn_y, compute_type, *, mode_a=None, loc=None, ip=None) → _ods_ir | _ods_ir | SpMVBufferSizeOp¶

class mlir.dialects.gpu.SpMVOp(asyncToken, asyncDependencies, spmatA, dnX, dnY, computeType, buffer, *, modeA=None, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.spmv operation performs the SpMV operation on the given sparse matrix, dense vectors, and buffer. The operation expects handles returned by previous sparse operations to construct an environment and the operands for SpMV. The buffer must have been allocated on the device.

The matrix arguments can also be associated with one of the following operators: NON_TRANSPOSE, TRANSPOSE, CONJUGATE_TRANSPOSE. The default value is NON_TRANSPOSE.

Example:

%token = gpu.spmv async [%dep] %spmatA{TRANSPOSE}, %dnX, %dnY : memref<?xf64> into bf16

OPERATION_NAME = 'gpu.spmv'¶

_ODS_REGIONS = (0, True)¶

asyncDependencies() → _ods_ir¶

spmatA() → _ods_ir¶

dnX() → _ods_ir¶

dnY() → _ods_ir¶

buffer() → _ods_ir¶

modeA() → _ods_ir¶

computeType() → _ods_ir¶

asyncToken() → _ods_ir | None¶

mlir.dialects.gpu.spmv(async_token, async_dependencies, spmat_a, dn_x, dn_y, compute_type, buffer, *, mode_a=None, loc=None, ip=None) → _ods_ir | _ods_ir | SpMVOp¶

class mlir.dialects.gpu.SpMatGetSizeOp(rows, cols, nnz, asyncToken, asyncDependencies, spmat, *, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.spmat_get_size operation retrieves the number of rows, number of columns, and number of non-zero elements of a sparse matrix.

Example:

%rows, %cols, %nnz, %token = gpu.spmat_get_size async [%dep] %spmatC

OPERATION_NAME = 'gpu.spmat_get_size'¶

_ODS_REGIONS = (0, True)¶

asyncDependencies() → _ods_ir¶

spmat() → _ods_ir¶

rows() → _ods_ir¶

cols() → _ods_ir¶

nnz() → _ods_ir¶

asyncToken() → _ods_ir | None¶

mlir.dialects.gpu.spmat_get_size(rows, cols, nnz, async_token, async_dependencies, spmat, *, loc=None, ip=None) → _ods_ir | _ods_ir | SpMatGetSizeOp¶

class mlir.dialects.gpu.SubgroupBroadcastOp(src, broadcast_type, *, lane=None, results=None, loc=None, ip=None)¶

Bases: _ods_ir

Broadcasts a value from one lane to all active lanes in a subgroup. The result is guaranteed to be uniform across the active lanes in subgroup.

The possible broadcast types are:

first_active_lane - broadcasts the value from the first active lane

in the subgroup. * specific_lane - broadcasts from the specified lane. The lane index must be uniform and within the subgroup size. The result is poison if the lane index is invalid, non subgroup-uniform, or if the source lane is not active.

OPERATION_NAME = 'gpu.subgroup_broadcast'¶

_ODS_REGIONS = (0, True)¶

src() → _ods_ir¶

lane() → _ods_ir | None¶

broadcast_type() → _ods_ir¶

result() → _ods_ir¶: Shortcut to get an op result if it has only one (throws an error otherwise).

mlir.dialects.gpu.subgroup_broadcast(src, broadcast_type, *, lane=None, results=None, loc=None, ip=None) → _ods_ir¶

class mlir.dialects.gpu.SubgroupIdOp(*, upper_bound=None, results=None, loc=None, ip=None)¶

Bases: _ods_ir

Returns the subgroup id, i.e., the index of the current subgroup within the workgroup.

Example:

%sgId = gpu.subgroup_id : index

Executions where there are more than upper_bound subgroups per workgroup cause undefined behavior. There is an implicit upper bound of kMaxDim (currently uint32_t::max).

OPERATION_NAME = 'gpu.subgroup_id'¶

_ODS_REGIONS = (0, True)¶

upper_bound() → _ods_ir | None¶

result() → _ods_ir¶: Shortcut to get an op result if it has only one (throws an error otherwise).

mlir.dialects.gpu.subgroup_id(*, upper_bound=None, results=None, loc=None, ip=None) → _ods_ir¶

class mlir.dialects.gpu.SubgroupMmaComputeOp(opA, opB, opC, *, a_transpose=None, b_transpose=None, results=None, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.subgroup_mma_compute operation performs a matrix-multiply accumulate (mma) operation using all the threads in a subgroup.

This operation takes three !gpu.mma_matrix``s as arguments: these hold ``A, B and C``operands for the mma operation. The operation performed is represented as ``C += A * B. The op returns a !gpu.mma_matrix which contains the result of the operation held by all threads in a subgroup. a_transpose or b_transpose if present, signify that the respective operand was loaded in a transposed manner. The transpose operands are required to map to correct underlying intrisics but they currently do not seem to affect correctness even if they are absent given that the operands were loaded correctly using the transpose attribute in gpu.subgroup_mma_load_matrix op.

For integer types, the A and B matrices carry their signedness with their types. The accumulator type is expected to be signless and imply a signed integer with a greater width than the other two operands.

This op is meant to be used along with gpu.subgroup_mma_store_matrix and gpu.subgroup_mma_load_matrix ops.

Example:

%D = gpu.subgroup_mma_compute_matrix %A, %B, %C :
  !gpu.mma_matrix<16x16xf16, "AOp">, !gpu.mma_matrix<16x16xf16, "BOp">>
  -> !gpu.mma_matrix<16x16xf16, "COp">

OPERATION_NAME = 'gpu.subgroup_mma_compute'¶

_ODS_REGIONS = (0, True)¶

opA() → _ods_ir¶

opB() → _ods_ir¶

opC() → _ods_ir¶

a_transpose() → bool¶

b_transpose() → bool¶

res() → _ods_ir¶

mlir.dialects.gpu.subgroup_mma_compute(op_a, op_b, op_c, *, a_transpose=None, b_transpose=None, results=None, loc=None, ip=None) → _ods_ir¶

class mlir.dialects.gpu.SubgroupMmaConstantMatrixOp(res, value, *, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.subgroup_mma_constant_matrix creates a !gpu.mma_matrix with constant elements.

The operation takes a scalar input and return a !gpu.mma_matrix where each element of is equal to the operand constant. The destination mma_matrix type must have elememt type equal to the constant type. Since the layout of !gpu.mma_matrix is opaque this only support setting all the elements to the same value.

This op is meant to be used along with gpu.subgroup_mma_compute.

Example:

%0 = gpu.subgroup_mma_constant_matrix %a :
  !gpu.mma_matrix<16x16xf16, "AOp">
%1 = gpu.subgroup_mma_constant_matrix %b :
  !gpu.mma_matrix<16x16xf32, "COp">

OPERATION_NAME = 'gpu.subgroup_mma_constant_matrix'¶

_ODS_REGIONS = (0, True)¶

value() → _ods_ir¶

res() → _ods_ir¶

mlir.dialects.gpu.subgroup_mma_constant_matrix(res, value, *, loc=None, ip=None) → _ods_ir¶

class mlir.dialects.gpu.SubgroupMmaElementwiseOp(res, args, opType, *, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.subgroup_mma_elementwise takes !gpu.mma_matrix inputs and compute a new !gpu.mma_matrix by applying an elementwise operation to each element.

Since the operation is elementwise and the matrix type must match, the matrix elements are processed independently of the matrix layout.

This op is meant to be used along with gpu.subgroup_mma_compute.

Example:

%0 =  %A, %B { opType = "ADD" } :
 (!gpu.mma_matrix<16x16xf16, "COp">, !gpu.mma_matrix<16x16xf16, "COp">)
 -> !gpu.mma_matrix<16x16xf16, "COp">

OPERATION_NAME = 'gpu.subgroup_mma_elementwise'¶

_ODS_REGIONS = (0, True)¶

args() → _ods_ir¶

opType() → _ods_ir¶

res() → _ods_ir¶

mlir.dialects.gpu.subgroup_mma_elementwise(res, args, op_type, *, loc=None, ip=None) → _ods_ir¶

class mlir.dialects.gpu.SubgroupMmaExtractThreadLocalOp(matrix, indices, *, results=None, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.subgroup_mma_extract_thread_local operation extracts a value from !gpu.mma_matrix that is stored at subgroup level.

This operation takes !gpu.mma_matrix as its first operand. It is the source matrix across a subgroup. The op returns a scalar value stored in the invocation in the subgroup.

Since matrix is packed into the the threads within a subgroup, indices are the indices into the values stored by each thread. That is, an index of 0 (or [0, 0]) does not necessarily refer to the first element of the matrix, but the first element that a particular thread holds.

The mapping of matrix elements to threads is not defined by this operation and may not be defined by some lowerings (such as the lowering to SPIR-V). However, if the size of the subgroup is S, then subgroup_mma_extract_thread_local at each index in [0, (M * N) / S) will have the entire matrix extracted across the subgroup.

Example:

%c0 = arith.constant 0 : index
%val = gpu.subgroup_mma_extract_thread_local %m[%c0] : !gpu.mma_matrix<16x16xf32, "AOp"> -> f32

OPERATION_NAME = 'gpu.subgroup_mma_extract_thread_local'¶

_ODS_REGIONS = (0, True)¶

matrix() → _ods_ir¶

indices() → _ods_ir¶

res() → _ods_ir¶

mlir.dialects.gpu.subgroup_mma_extract_thread_local(matrix, indices, *, results=None, loc=None, ip=None) → _ods_ir¶

class mlir.dialects.gpu.SubgroupMmaInsertThreadLocalOp(res, value, matrix, indices, *, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.subgroup_mma_insert_thread_local operation inserts a value to !gpu.mma_matrix that is stored at subgroup level.

This operation takes scalar value as its first operand and !gpu.mma_matrix as its second operand. The op inserts the scalar value to the matrix.

The mapping of matrix elements to threads is not defined by this operation and may not be defined by some lowerings (such as the lowering to SPIR-V). However, if the size of the subgroup is S, then subgroup_mma_insert_thread_local at each index in [0, (M * N) / S) will have the entire matrix inserted across the subgroup.

The op returns !gpu.mma_matrix with the updated value.

Example:

%c0 = arith.constant 0 : index
%s0 = gpu.subgroup_mma_insert_thread_local %val, %m[%c0] : f16, !gpu.mma_matrix<16x16xf16, "COp">
        -> !gpu.mma_matrix<16x16xf16, "COp">

OPERATION_NAME = 'gpu.subgroup_mma_insert_thread_local'¶

_ODS_REGIONS = (0, True)¶

value() → _ods_ir¶

matrix() → _ods_ir¶

indices() → _ods_ir¶

res() → _ods_ir¶

mlir.dialects.gpu.subgroup_mma_insert_thread_local(res, value, matrix, indices, *, loc=None, ip=None) → _ods_ir¶

class mlir.dialects.gpu.SubgroupMmaLoadMatrixOp(res, srcMemref, indices, leadDimension, *, transpose=None, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.subgroup_mma_load_matrix operation loads a matrix collectively using all the threads in a subgroup.

This operation takes a memref as its first operand: it is the source matrix from which data is to be loaded. The op returns a !gpu.mma_matrix. The source memref can be in global memory or shared memory. The load address is determined using indices. The matrix being loaded into is the result. The leadDimension attribute specifies the leading dimension size of the source matrix which eventually allows the lowering to determine the size of each row. If the transpose attribute is present then the op does a transposed load.

For integer types, the resulting !gpu.mma_matrix type needs to specify the signedness of the data if the matrix type is an A or B operand for gpu.subgroup_mma_compute.

This op is often meant to be used along with gpu.subgroup_mma_store_matrix and gpu.subgroup_mma_compute.

Example:

%0 = gpu.subgroup_mma_load_matrix src[%i,%j] : {leadDimension = 32 : i32}
     : memref<32x32xf16, 3>, !gpu.mma_matrix<16x16xf16, "AOp">

OPERATION_NAME = 'gpu.subgroup_mma_load_matrix'¶

_ODS_REGIONS = (0, True)¶

srcMemref() → _ods_ir¶

indices() → _ods_ir¶

leadDimension() → _ods_ir¶

transpose() → bool¶

res() → _ods_ir¶

mlir.dialects.gpu.subgroup_mma_load_matrix(res, src_memref, indices, lead_dimension, *, transpose=None, loc=None, ip=None) → _ods_ir¶

class mlir.dialects.gpu.SubgroupMmaStoreMatrixOp(src, dstMemref, indices, leadDimension, *, transpose=None, loc=None, ip=None)¶

Bases: _ods_ir

The gpu.subgroup_mma_store_matrix operation stores a matrix collectively using all the threads in a subgroup.

This operation takes a !gpu.mma_matrix and a memref as operands. !gpu.mma_matrix is the source value containing the data to be stored into the destination memref which can be in global or shared memory. The store address is determined using the indices provided. The leadDimension attribute specifies the leading dimension of the destination matrix. If the transpose attribute is present then the op does a transposed store.

This op is often meant to be used along with gpu.subgroup_mma_load_matrix and gpu.subgroup_mma_compute.

Example:

gpu.subgroup_mma_store_matrix %D, %sg[%i,%j] : { leadDimension = 32 : i32}
                : !gpu.mma_matrix<16x16xf16, "COp">, memref<32x32xf16, 3>

OPERATION_NAME = 'gpu.subgroup_mma_store_matrix'¶

_ODS_REGIONS = (0, True)¶

src() → _ods_ir¶

dstMemref() → _ods_ir¶

indices() → _ods_ir¶

leadDimension() → _ods_ir¶

transpose() → bool¶

mlir.dialects.gpu.subgroup_mma_store_matrix(src, dst_memref, indices, lead_dimension, *, transpose=None, loc=None, ip=None) → SubgroupMmaStoreMatrixOp¶

class mlir.dialects.gpu.SubgroupReduceOp(value, op, *, uniform=None, cluster_size=None, cluster_stride=None, results=None, loc=None, ip=None)¶

Bases: _ods_ir

The subgroup_reduce op reduces the values of lanes (work items) across a subgroup.

The subgroup is divided into clusters starting at lane index 0. Within each cluster, there are size lanes, and the lane index advances by stride. A reduction is done for each cluster in parallel: every lane in the cluster is reduced, and the result is equal for all lanes in the cluster. If size is omitted, there is a single cluster covering the entire subgroup. If stride is omitted, the stride is 1 (the cluster’s lanes are contiguous).

When the reduced value is of a vector type, each vector element is reduced independently. Only 1-d vector types are allowed.

Example:

%1 = gpu.subgroup_reduce add %a : (f32) -> f32
%2 = gpu.subgroup_reduce add %b : (vector<4xf16>) -> vector<4xf16>
%3 = gpu.subgroup_reduce add %c cluster(size = 4) : (f32) -> f32
%3 = gpu.subgroup_reduce add %c cluster(size = 4, stride = 2) : (f32) -> f32

If uniform flag is set either none or all lanes of a subgroup need to execute this op in convergence.

The reduction operation must be one of:

Integer types: add, mul, minui, minsi, maxui, maxsi, and,

or, xor * Floating point types: add, mul, minnumf, maxnumf, minimumf, maximumf

OPERATION_NAME = 'gpu.subgroup_reduce'¶

_ODS_REGIONS = (0, True)¶

value() → _ods_ir¶

op() → _ods_ir¶

uniform() → bool¶

cluster_size() → _ods_ir | None¶

cluster_stride() → _ods_ir¶

result() → _ods_ir¶: Shortcut to get an op result if it has only one (throws an error otherwise).

mlir.dialects.gpu.subgroup_reduce(value, op, *, uniform=None, cluster_size=None, cluster_stride=None, results=None, loc=None, ip=None) → _ods_ir¶

class mlir.dialects.gpu.SubgroupSizeOp(*, upper_bound=None, results=None, loc=None, ip=None)¶

Bases: _ods_ir

Returns the number of threads within a subgroup.

Example:

%sgSz = gpu.subgroup_size : index

Executions where the number of threads per subgroup exceed upper_bound cause undefined behavior. When no upper_bound is specified, range analyses and similar machinery assume the default bound of kMaxSubgroupSize, currently 128.

OPERATION_NAME = 'gpu.subgroup_size'¶

_ODS_REGIONS = (0, True)¶

upper_bound() → _ods_ir | None¶

result() → _ods_ir¶: Shortcut to get an op result if it has only one (throws an error otherwise).

mlir.dialects.gpu.subgroup_size(*, upper_bound=None, results=None, loc=None, ip=None) → _ods_ir¶

class mlir.dialects.gpu.TerminatorOp(*, loc=None, ip=None)¶

Bases: _ods_ir

A terminator operation for regions that appear in the body of gpu.launch operation. These regions are not expected to return any value so the terminator takes no operands.

OPERATION_NAME = 'gpu.terminator'¶

_ODS_REGIONS = (0, True)¶

mlir.dialects.gpu.terminator(*, loc=None, ip=None) → TerminatorOp¶

class mlir.dialects.gpu.ThreadIdOp(dimension, *, upper_bound=None, results=None, loc=None, ip=None)¶

Bases: _ods_ir

Returns the thread id, i.e. the index of the current thread within the block along the x, y, or z dimension.

Example:

%tIdX = gpu.thread_id x

If upper_bound is set, or if one can be inferred from known_block_size-type annotations in context, executions where the thread index would be greater than or equal to that bound cause undefined behavior.

There is an implicit upper bound of kMaxDim (currently uint32_t::max).

OPERATION_NAME = 'gpu.thread_id'¶

_ODS_REGIONS = (0, True)¶

dimension() → _ods_ir¶

upper_bound() → _ods_ir | None¶

mlir.dialects.gpu.thread_id(dimension, *, upper_bound=None, results=None, loc=None, ip=None) → _ods_ir¶

class mlir.dialects.gpu.WaitOp(asyncToken, asyncDependencies, *, loc=None, ip=None)¶

Bases: _ods_ir

This op synchronizes the host or the device with a list of dependent ops.

If the op contains the async keyword, it returns a new async token which is synchronized with the op arguments. This new token is merely a shortcut to the argument list, and one could replace the uses of the result with the arguments for the same effect. The async version of this op is primarily used to make each async token have a single use during lowering and thereby make forks in async execution explicit. Example usage:

%t0 = gpu.foo async : !gpu.async.token
%t1 = gpu.bar async : !gpu.async.token
%t2 = gpu.wait async [%t0, %t1]
// gpu.baz doesn't run until gpu.foo and gpu.bar have both completed, just
// as if the async dependencies were [%t0, %t1].
%t3 = gpu.baz async [%t2]

If the op does not contain the async keyword, it does not return a new async token but blocks until all ops producing the async dependency tokens finished execution. All dependent memory operations are visible to the host once this op completes. Example usage:

%t0 = gpu.foo async : !gpu.async.token
%t1 = gpu.bar async : !gpu.async.token
// The gpu.wait op blocks until gpu.foo and gpu.bar have completed.
gpu.wait [%t0, %t1]

OPERATION_NAME = 'gpu.wait'¶

_ODS_REGIONS = (0, True)¶

asyncDependencies() → _ods_ir¶

asyncToken() → _ods_ir | None¶

mlir.dialects.gpu.wait(async_token, async_dependencies, *, loc=None, ip=None) → _ods_ir | _ods_ir | WaitOp¶

class mlir.dialects.gpu.WarpExecuteOnLane0Op(results_, laneid, warp_size, args, *, loc=None, ip=None)¶

Bases: _ods_ir

warp_execute_on_lane_0 is an operation used to bridge the gap between vector programming and SPMD programming model like GPU SIMT. It allows to trivially convert a region of vector code meant to run on a multiple threads into a valid SPMD region and then allows incremental transformation to distribute vector operations on the threads.

Any code present in the region would only be executed on first thread/lane based on the laneid operand. The laneid operand is an integer ID between [0, warp_size). The warp_size attribute indicates the number of lanes in a warp.

Operands are vector values distributed on all lanes that may be used by the single lane execution. The matching region argument is a vector of all the values of those lanes available to the single active lane. The distributed dimension is implicit based on the shape of the operand and argument. the properties of the distribution may be described by extra attributes (e.g. affine map).

Return values are distributed on all lanes using laneId as index. The vector is distributed based on the shape ratio between the vector type of the yield and the result type. If the shapes are the same this means the value is broadcasted to all lanes. In the future the distribution can be made more explicit using affine_maps and will support having multiple Ids.

Therefore the warp_execute_on_lane_0 operations allow to implicitly copy between lane0 and the lanes of the warp. When distributing a vector from lane0 to all the lanes, the data are distributed in a block cyclic way. For example vector<64xf32> gets distributed on 32 threads and map to vector<2xf32> where thread 0 contains vector[0] and vector[1].

During lowering values passed as operands and return value need to be visible to different lanes within the warp. This would usually be done by going through memory.

The region is not isolated from above. For values coming from the parent region not going through operands only the lane 0 value will be accesible so it generally only make sense for uniform values.

Example:

// Execute in parallel on all threads/lanes.
gpu.warp_execute_on_lane_0 (%laneid)[32] {
  // Serial code running only on thread/lane 0.
  ...
}
// Execute in parallel on all threads/lanes.

This may be lowered to an scf.if region as below:

// Execute in parallel on all threads/lanes.
%cnd = arith.cmpi eq, %laneid, %c0 : index
scf.if %cnd {
  // Serial code running only on thread/lane 0.
  ...
}
// Execute in parallel on all threads/lanes.

When the region has operands and/or return values:

// Execute in parallel on all threads/lanes.
%0 = gpu.warp_execute_on_lane_0(%laneid)[32]
args(%v0 : vector<4xi32>) -> (vector<1xf32>) {
^bb0(%arg0 : vector<128xi32>) :
  // Serial code running only on thread/lane 0.
  ...
  gpu.yield %1 : vector<32xf32>
}
// Execute in parallel on all threads/lanes.

values at the region boundary would go through memory:

// Execute in parallel on all threads/lanes.
...
// Store the data from each thread into memory and Synchronization.
%tmp0 = memreg.alloc() : memref<128xf32>
%tmp1 = memreg.alloc() : memref<32xf32>
%cnd = arith.cmpi eq, %laneid, %c0 : index
vector.store %v0, %tmp0[%laneid] : memref<128xf32>, vector<4xf32>
some_synchronization_primitive
scf.if %cnd {
  // Serialized code running only on thread 0.
  // Load the data from all the threads into a register from thread 0. This
  // allow threads 0 to access data from all the threads.
  %arg0 = vector.load %tmp0[%c0] : memref<128xf32>, vector<128xf32>
  ...
  // Store the data from thread 0 into memory.
  vector.store %1, %tmp1[%c0] : memref<32xf32>, vector<32xf32>
}
// Synchronization and load the data in a block cyclic way so that the
// vector is distributed on all threads.
some_synchronization_primitive
%0 = vector.load %tmp1[%laneid] : memref<32xf32>, vector<32xf32>
// Execute in parallel on all threads/lanes.

OPERATION_NAME = 'gpu.warp_execute_on_lane_0'¶

_ODS_REGIONS = (1, True)¶

laneid() → _ods_ir¶

args() → _ods_ir¶

warp_size() → _ods_ir¶

results_() → _ods_ir¶

warpRegion() → _ods_ir¶

mlir.dialects.gpu.warp_execute_on_lane_0(results_, laneid, warp_size, args, *, loc=None, ip=None) → _ods_ir | _ods_ir | WarpExecuteOnLane0Op¶

class mlir.dialects.gpu.YieldOp(values, *, loc=None, ip=None)¶

Bases: _ods_ir

gpu.yield is a special terminator operation for blocks inside regions in gpu ops. It returns values to the immediately enclosing gpu op.

Example:

gpu.yield %f0, %f1 : f32, f32

OPERATION_NAME = 'gpu.yield'¶

_ODS_REGIONS = (0, True)¶

values() → _ods_ir¶

mlir.dialects.gpu.yield_(values, *, loc=None, ip=None) → YieldOp¶

class mlir.dialects.gpu._Dialect(descriptor: object)¶

Bases: _ods_ir

DIALECT_NAMESPACE = 'gpu'¶

mlir.dialects.gpu.register_attribute_builder(kind, replace=False)¶

mlir.dialects.gpu._ods_ir¶

class mlir.dialects.gpu.AddressSpace¶

Bases: enum.IntEnum

GPU address space

Global = 1¶

Workgroup = 2¶

Private = 3¶

__str__()¶: Return str(self).

mlir.dialects.gpu._gpu_addressspaceenum(x, context)¶

class mlir.dialects.gpu.AllReduceOperation¶

Bases: enum.IntEnum

built-in reduction operations supported by gpu.allreduce.

ADD = 0¶

MUL = 1¶

MINUI = 2¶

MINSI = 3¶

MINNUMF = 4¶

MAXUI = 5¶

MAXSI = 6¶

MAXNUMF = 7¶

AND = 8¶

OR = 9¶

XOR = 10¶

MINIMUMF = 11¶

MAXIMUMF = 12¶

__str__()¶: Return str(self).

mlir.dialects.gpu._gpu_allreduceoperation(x, context)¶

class mlir.dialects.gpu.BroadcastType¶

Bases: enum.IntEnum

a lane to broadcast from

first_active_lane = 0¶

specific_lane = 1¶

__str__()¶: Return str(self).

mlir.dialects.gpu._gpu_broadcasttype(x, context)¶

class mlir.dialects.gpu.CompilationTarget¶

Bases: enum.IntEnum

GPU compilation format

Offload = 1¶

Assembly = 2¶

Binary = 3¶

Fatbin = 4¶

__str__()¶: Return str(self).

mlir.dialects.gpu._gpu_compilationtargetenum(x, context)¶

class mlir.dialects.gpu.Dimension¶

Bases: enum.IntEnum

a dimension, either ‘x’, ‘y’, or ‘z’

x = 0¶

y = 1¶

z = 2¶

__str__()¶: Return str(self).

mlir.dialects.gpu._gpu_dimension(x, context)¶

class mlir.dialects.gpu.Prune2To4SpMatFlag¶

Bases: enum.IntEnum

pruning strategy for 2:4 sparse matrix

NONE = 0¶

PRUNE_ONLY = 1¶

PRUNE_AND_CHECK = 2¶

__str__()¶: Return str(self).

mlir.dialects.gpu._gpu_prune2to4spmatflag(x, context)¶

class mlir.dialects.gpu.ShuffleMode¶

Bases: enum.IntEnum

Indexing modes supported by gpu.shuffle.

XOR = 0¶

UP = 2¶

DOWN = 1¶

IDX = 3¶

__str__()¶: Return str(self).

mlir.dialects.gpu._gpu_shufflemode(x, context)¶

class mlir.dialects.gpu.SpGEMMWorkEstimationOrComputeKind¶

Bases: enum.IntEnum

choose whether spgemm_work_estimation_or_compute does work estimation or compute

WORK_ESTIMATION = 0¶

COMPUTE = 1¶

__str__()¶: Return str(self).

mlir.dialects.gpu._gpu_spgemmworkestimationorcomputekind(x, context)¶

class mlir.dialects.gpu.TransposeMode¶

Bases: enum.IntEnum

transpose mode of sparse matrix supported by sparse tensor ops

NON_TRANSPOSE = 0¶

TRANSPOSE = 1¶

CONJUGATE_TRANSPOSE = 2¶

__str__()¶: Return str(self).

mlir.dialects.gpu._gpu_transposemode(x, context)¶

class mlir.dialects.gpu.MMAElementwiseOp¶

Bases: enum.IntEnum

elementwise operation to apply to mma matrix

ADDF = 0¶

MULF = 1¶

SUBF = 2¶

MAXF = 3¶

MINF = 4¶

DIVF = 5¶

ADDI = 6¶

MULI = 7¶

SUBI = 8¶

DIVS = 9¶

DIVU = 10¶

NEGATEF = 11¶

NEGATES = 12¶

EXTF = 13¶

__str__()¶: Return str(self).

mlir.dialects.gpu._mmaelementwise(x, context)¶

class mlir.dialects.gpu.MappingId¶

Bases: enum.IntEnum

Mapping ids for loop mapping

DimX = 0¶

DimY = 1¶

DimZ = 2¶

LinearDim0 = 3¶

LinearDim1 = 4¶

LinearDim2 = 5¶

LinearDim3 = 6¶

LinearDim4 = 7¶

LinearDim5 = 8¶

LinearDim6 = 9¶

LinearDim7 = 10¶

LinearDim8 = 11¶

LinearDim9 = 12¶

__str__()¶: Return str(self).

mlir.dialects.gpu._mappingidenum(x, context)¶

class mlir.dialects.gpu.Processor¶

Bases: enum.IntEnum

processor for loop mapping

BlockX = 0¶

BlockY = 1¶

BlockZ = 2¶

ThreadX = 3¶

ThreadY = 4¶

ThreadZ = 5¶

Sequential = 6¶

__str__()¶: Return str(self).

mlir.dialects.gpu._processorenum(x, context)¶

mlir.dialects.gpu._gpu_addressspaceattr(x, context)¶

mlir.dialects.gpu._gpu_allreduceoperationattr(x, context)¶

mlir.dialects.gpu._gpu_broadcasttypeattr(x, context)¶

mlir.dialects.gpu._gpu_dimensionattr(x, context)¶

mlir.dialects.gpu._gpu_prune2to4spmatflagattr(x, context)¶

mlir.dialects.gpu._gpu_shufflemodeattr(x, context)¶

mlir.dialects.gpu._gpu_spgemmworkestimationorcomputekindattr(x, context)¶

mlir.dialects.gpu._gpu_transposemodeattr(x, context)¶

mlir.dialects.gpu._mmaelementwiseattr(x, context)¶

mlir.dialects.gpu.gpu_async_token()¶

class mlir.dialects.gpu.GPUFuncOp(function_type: mlir.ir.FunctionType | mlir.ir.TypeAttr, sym_name: str | mlir.ir.StringAttr | None = None, kernel: bool | None = None, workgroup_attrib_attrs: Sequence[dict] | None = None, private_attrib_attrs: Sequence[dict] | None = None, known_block_size: Sequence[int] | mlir.ir.DenseI32ArrayAttr | None = None, known_grid_size: Sequence[int] | mlir.ir.DenseI32ArrayAttr | None = None, loc=None, ip=None, body_builder: Callable[[GPUFuncOp], None] | None = None)¶

Bases: GPUFuncOp

Defines a function that can be executed on a GPU. This supports memory attribution and its body has a particular execution model.

GPU functions are either kernels (as indicated by the kernel attribute) or regular functions. The former can be launched from the host side, while the latter are device side only.

The memory attribution defines SSA values that correspond to memory buffers allocated in the memory hierarchy of the GPU (see below).

Syntax:

op ::= `gpu.func` symbol-ref-id `(` argument-list `)` (`->`
function-result-list)?
       memory-attribution `kernel`? function-attributes? region

memory-attribution ::= (`workgroup` `(` ssa-id-and-type-list `)`)?
                       (`private` `(` ssa-id-and-type-list `)`)?

Example:

gpu.func @foo(%arg0: index)
    workgroup(%workgroup: memref<32xf32, 3>)
    private(%private: memref<1xf32, 5>)
    kernel
    attributes {qux: "quux"} {
  gpu.return
}

The generic form illustrates the concept

"gpu.func"(%arg: index) {sym_name: "foo", kernel, qux: "quux"} ({
^bb0(%arg0: index, %workgroup: memref<32xf32, 3>,
     %private: memref<1xf32, 5>):
  "gpu.return"() : () -> ()
}) : (index) -> ()

Note the non-default memory spaces used in memref types in memory attribution.

__doc__¶

KERNEL_ATTR_NAME = 'gpu.kernel'¶

KNOWN_BLOCK_SIZE_ATTR_NAME = 'known_block_size'¶

KNOWN_GRID_SIZE_ATTR_NAME = 'known_grid_size'¶

FUNCTION_TYPE_ATTR_NAME = 'function_type'¶

SYM_NAME_ATTR_NAME = 'sym_name'¶

ARGUMENT_ATTR_NAME = 'arg_attrs'¶

RESULT_ATTR_NAME = 'res_attrs'¶

property name: mlir.ir.StringAttr¶: Returns the fully qualified name of the operation.

property is_kernel: bool¶

add_entry_block() → mlir.ir.Block¶

property entry_block: mlir.ir.Block¶

property arguments: Sequence[mlir.ir.Type]¶

mlir.dialects.gpu._convert_literal_to_constant(value: int | mlir.dialects.arith.ConstantOp | mlir.ir.Value) → mlir.ir.Value¶

class mlir.dialects.gpu.LaunchFuncOp(kernel: List[str], grid_size: Tuple[Any, Any, Any], block_size: Tuple[Any, Any, Any], kernel_operands: List[mlir.ir.Value] | None = None, async_dependencies: List[mlir.ir.Value] | None = None, dynamic_shared_memory_size: mlir.ir.Value | None = None, async_object=None, *, loc=None, ip=None)¶

Bases: LaunchFuncOp

The gpu.launch_func supports async dependencies: the kernel does not start executing until the ops producing those async dependencies have completed.

The operation requires at least the grid and block sizes along the x,y,z dimensions as arguments. When a lower-dimensional kernel is required, unused sizes must be explicitly set to 1.

The remaining operands if present are passed as arguments to the kernel function.

Example:

module attributes {gpu.container_module} {

  // This module creates a separate compilation unit for the GPU compiler.
  gpu.module @kernels {
    func.func @kernel_1(%arg0 : f32, %arg1 : memref<?xf32, 1>)
        attributes { nvvm.kernel = true } {

      // Operations that produce block/thread IDs and dimensions are
      // injected when outlining the `gpu.launch` body to a function called
      // by `gpu.launch_func`.
      %tIdX = gpu.thread_id x
      %tIdY = gpu.thread_id y
      %tIdZ = gpu.thread_id z

      %bDimX = gpu.block_dim x
      %bDimY = gpu.block_dim y
      %bDimZ = gpu.block_dim z

      %bIdX = gpu.block_id x
      %bIdY = gpu.block_id y
      %bIdZ = gpu.block_id z

      %gDimX = gpu.grid_dim x
      %gDimY = gpu.grid_dim y
      %gDimZ = gpu.grid_dim z

      // (Optional)  Cluster size only for support architectures
      %cIdX = gpu.cluster_id x
      %cIdY = gpu.cluster_id y
      %cIdZ = gpu.cluster_id z

      %cDimX = gpu.cluster_dim x
      %cDimY = gpu.cluster_dim y
      %cDimZ = gpu.cluster_dim z

      "some_op"(%bx, %tx) : (index, index) -> ()
      %42 = load %arg1[%bx] : memref<?xf32, 1>
    }
  }

  %t0 = gpu.wait async
  gpu.launch_func
      async                           // (Optional) Don't block host, return token.
      [%t0]                           // (Optional) Execute only after %t0 has completed.
      @kernels::@kernel_1             // Kernel function.
      clusters in (%cst, %cst, %cst)  // (Optional) Cluster size only for support architectures.
      blocks in (%cst, %cst, %cst)    // Grid size.
      threads in (%cst, %cst, %cst)   // Block size.
      dynamic_shared_memory_size %s   // (Optional) Amount of dynamic shared
                                      // memory to allocate for a workgroup.
      args(%arg0 : f32,               // (Optional) Kernel arguments.
           %arg1 : memref<?xf32, 1>)
}

__doc__¶

mlir.dialects.gpu.launch_func(kernel: List[str], grid_size: Tuple[Any, Any, Any], block_size: Tuple[Any, Any, Any], kernel_operands: List[mlir.ir.Value] | None = None, async_dependencies: List[mlir.ir.Value] | None = None, dynamic_shared_memory_size: mlir.ir.Value | None = None, async_object=None, *, loc=None, ip=None) → mlir.ir.Value | List[mlir.ir.Value] | LaunchFuncOp¶

mlir.dialects.gpu.wait(async_dependencies: List[mlir.ir.Value] | None = None, *, loc=None, ip=None) → mlir.ir.Value | List[mlir.ir.Value] | mlir.dialects._gpu_ops_gen.WaitOp¶

class mlir.dialects.gpu.LaunchOp(grid_size: Tuple[Any, Any, Any], block_size: Tuple[Any, Any, Any], async_dependencies=None, dynamic_shared_memory_size: mlir.ir.Value | None = None, *, loc=None, ip=None)¶

Bases: LaunchOp

The body region has at least twelve arguments, or eighteen if cluster dimensions are present, grouped as follows:

three optional arguments that contain cluster identifiers along x,y,z

The function and module attributes are optional and specifies the kernel name and a module in which the kernel should be outlined.

Syntax:

operation ::= `gpu.launch` (`async` (`[` ssa-id-list `]`)? )?
                         ( `clusters` `(` ssa-id-list `)` `in` ssa-reassignment )?
                         `blocks` `(` ssa-id-list `)` `in` ssa-reassignment
                         `threads` `(` ssa-id-list `)` `in` ssa-reassignment
                         (dynamic_shared_memory_size ssa-use)?
                         (`module(` symbol-ref-id `)`)?
                         (`function(` symbol-ref-id `)`)?
                         memory-attribution
                         region attr-dict?
ssa-reassignment ::= `(` ssa-id `=` ssa-use (`,` ssa-id `=` ssa-use)* `)`
memory-attribution ::= (`workgroup` `(` ssa-id-and-type-list `)`)?
                       (`private` `(` ssa-id-and-type-list `)`)?

Example:

gpu.launch blocks(%bx, %by, %bz) in (%sz_bx = %0, %sz_by = %1, %sz_bz = %2)
           threads(%tx, %ty, %tz) in (%sz_tx = %3, %sz_ty = %4, %sz_tz = %5) {
  // Block and thread identifiers, as well as block/grid sizes are
  // immediately usable inside body region.
  "some_op"(%bx, %tx) : (index, index) -> ()
  // Assuming %val1 is defined outside the gpu.launch region.
  %42 = load %val1[%bx] : memref<?xf32, 1>
}

// Generic syntax explains how the pretty syntax maps to the IR structure.
"gpu.launch"(%cst, %cst, %c1,  // Grid sizes.
             %cst, %c1, %c1)   // Block sizes.

    {/*attributes*/}
    // All sizes and identifiers have "index" size.
    : (index, index, index, index, index, index) -> () {
// The operation passes block and thread identifiers, followed by grid and
// block sizes.
^bb0(%bx : index, %by : index, %bz : index,
     %tx : index, %ty : index, %tz : index,
     %num_bx : index, %num_by : index, %num_bz : index,
     %num_tx : index, %num_ty : index, %num_tz : index)
  "some_op"(%bx, %tx) : (index, index) -> ()
  %3 = "memref.load"(%val1, %bx) : (memref<?xf32, 1>, index) -> f32
}

// Launch with memory attributions.
gpu.launch blocks(%bx, %by, %bz) in (%sz_bx = %0, %sz_by = %1, %sz_bz = %2)
           threads(%tx, %ty, %tz) in (%sz_tx = %3, %sz_ty = %4, %sz_tz = %5)
           workgroup(%workgroup: memref<32xf32, 3>)
           private(%private: memref<1xf32, 5>) {
  // Block and thread identifiers, as well as block/grid sizes are
  // immediately usable inside body region.
  "some_op"(%bx, %tx) : (index, index) -> ()
  // Assuming %val1 is defined outside the gpu.launch region.
  %42 = load %workgroup[%bx] : memref<32xf32, 3>
}

// Launch with clusters.
gpu.launch clusters(%cx, %cy, %cz) in (%sz_cx = %0, %sz_cy = %1, %sz_cz = %2)
           blocks(%bx, %by, %bz) in (%sz_bx = %3, %sz_by = %4, %sz_bz = %5)
           threads(%tx, %ty, %tz) in (%sz_tx = %6, %sz_ty = %7, %sz_tz = %8)
{
  // Cluster, block and thread identifiers, as well as cluster/block/grid
  // sizes are immediately usable inside body region.
  "some_op"(%cx, %bx, %tx) : (index, index, index) -> ()
}

// Launch with module and function attributes.
gpu.launch blocks(%bx, %by, %bz) in (%sz_bx = %0, %sz_by = %1, %sz_bz = %2)
           threads(%tx, %ty, %tz) in (%sz_tx = %3, %sz_ty = %4, %sz_tz = %5)
           module(@kernel_module) function(@kernel_func) {
  "some_op"(%bx, %tx) : (index, index) -> ()
  %42 = load %val1[%bx] : memref<?xf32, 1>
}

__doc__¶

mlir.dialects.gpu.launch_(grid_size: Tuple[Any, Any, Any], block_size: Tuple[Any, Any, Any], async_dependencies=None, dynamic_shared_memory_size: mlir.ir.Value | None = None, *, loc=None, ip=None)¶

mlir.dialects.gpu.launch¶

mlir.dialects.gpu._printf¶

mlir.dialects.gpu.printf(format, *args, loc=None, ip=None)¶