# MLIR

Multi-Level IR Compiler Framework

# 'gpu' Dialect

Note: this dialect is more likely to change than others in the near future; use with caution.

This dialect provides middle-level abstractions for launching GPU kernels following a programming model similar to that of CUDA or OpenCL. It provides abstractions for kernel invocations (and may eventually provide those for device management) that are not present at the lower level (e.g., as LLVM IR intrinsics for GPUs). Its goal is to abstract away device- and driver-specific manipulations to launch a GPU kernel and provide a simple path towards GPU execution from MLIR. It may be targeted, for example, by DSLs using MLIR. The dialect uses gpu as its canonical prefix.

Memory buffers are defined at the function level, either in “gpu.launch” or in “gpu.func” ops. This encoding makes it clear where the memory belongs and makes the lifetime of the memory visible. The memory is only accessible while the kernel is launched/the function is currently invoked. The latter is more strict than actual GPU implementations but using static memory at the function level is just for convenience. It is also always possible to pass pointers to the workgroup memory into other functions, provided they expect the correct memory space.

The buffers are considered live throughout the execution of the GPU function body. The absence of memory attribution syntax means that the function does not require special buffers. Rationale: although the underlying models declare memory buffers at the module level, we chose to do it at the function level to provide some structuring for the lifetime of those buffers; this avoids the incentive to use the buffers for communicating between different kernels or launches of the same kernel, which should be done through function arguments instead; we chose not to use alloca-style approach that would require more complex lifetime analysis following the principles of MLIR that promote structure and representing analysis results in the IR.

## Operations ¶

### gpu.all_reduce (::mlir::gpu::AllReduceOp) ¶

Reduce values among workgroup.

Syntax:

( $dynamicSizes ) ( [$symbolOperands^ ])? attr-dict : type($memref)  The gpu.alloc operation allocates a region of memory on the GPU. It is similar to the memref.alloc op, but supports asynchronous GPU execution. The op does not execute before all async dependencies have finished executing. If the async keyword is present, the op is executed asynchronously (i.e. it does not block until the execution has finished on the device). In that case, it also returns a !gpu.async.token. If the host_shared keyword is present, the memory will be allocated in a memory accessible both on host and on device. Example: %memref, %token = gpu.alloc async [%dep] host_shared (%width) : memref<64x?xf32, 1>  Traits: AttrSizedOperandSegments Interfaces: GPU_AsyncOpInterface #### Attributes: ¶ AttributeMLIR TypeDescription hostShared::mlir::UnitAttrunit attribute #### Operands: ¶ OperandDescription asyncDependenciesasync token type dynamicSizesindex symbolOperandsindex #### Results: ¶ ResultDescription memrefmemref of any type values asyncTokenasync token type ### gpu.barrier (::mlir::gpu::BarrierOp) ¶ Synchronizes all work items of a workgroup. Syntax: operation ::= gpu.barrier attr-dict  The “barrier” op synchronizes all work items of a workgroup. It is used to coordinate communication between the work items of the workgroup. gpu.barrier  waits until all work items in the workgroup have reached this point and all memory accesses made by these work items prior to the op are visible to all work items in the workgroup. Data hazards between work items accessing the same memory can be avoided by synchronizing work items in-between these accesses. Either none or all work items of a workgroup need to execute this op in convergence. ### gpu.block_dim (::mlir::gpu::BlockDimOp) ¶ Syntax: operation ::= gpu.block_dim$dimension attr-dict


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


Interfaces: InferIntRangeInterface, InferTypeOpInterface, NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

#### Attributes: ¶

AttributeMLIR TypeDescription
dimension::mlir::gpu::DimensionAttra dimension, either ‘x’, ‘y’, or ‘z’

#### Results: ¶

ResultDescription
«unnamed»index

### gpu.block_id (::mlir::gpu::BlockIdOp) ¶

Syntax:

operation ::= gpu.block_id $dimension attr-dict  Returns the block id, i.e. the index of the current block within the grid along the x, y, or z dimension. Example: %bIdY = gpu.block_id y  Interfaces: InferIntRangeInterface, InferTypeOpInterface, NoSideEffect (MemoryEffectOpInterface) Effects: MemoryEffects::Effect{} #### Attributes: ¶ AttributeMLIR TypeDescription dimension::mlir::gpu::DimensionAttra dimension, either ‘x’, ‘y’, or ‘z’ #### Results: ¶ ResultDescription «unnamed»index ### gpu.dealloc (::mlir::gpu::DeallocOp) ¶ GPU memory deallocation operation Syntax: operation ::= gpu.dealloc custom<AsyncDependencies>(type($asyncToken), $asyncDependencies)$memref attr-dict : type($memref)  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>  Interfaces: GPU_AsyncOpInterface #### Operands: ¶ OperandDescription asyncDependenciesasync token type memrefmemref of any type values #### Results: ¶ ResultDescription asyncTokenasync token type ### gpu.func (::mlir::gpu::GPUFuncOp) ¶ Function executable on a GPU 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. 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. Traits: AutomaticAllocationScope, HasParent, IsolatedFromAbove Interfaces: FunctionOpInterface, Symbol #### Attributes: ¶ AttributeMLIR TypeDescription function_type::mlir::TypeAttrtype attribute of function type ### gpu.module (::mlir::gpu::GPUModuleOp) ¶ A top level compilation unit containing code to be run on a GPU. 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.  gpu.module @symbol_name { gpu.func {} ... gpu.module_end }  Traits: HasDefaultDLTIDataLayout, IsolatedFromAbove, SingleBlockImplicitTerminator, SymbolTable Interfaces: DataLayoutOpInterface, Symbol ### gpu.global_id (::mlir::gpu::GlobalIdOp) ¶ Syntax: operation ::= gpu.global_id$dimension attr-dict


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


Interfaces: InferIntRangeInterface, InferTypeOpInterface, NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

#### Attributes: ¶

AttributeMLIR TypeDescription
dimension::mlir::gpu::DimensionAttra dimension, either ‘x’, ‘y’, or ‘z’

#### Results: ¶

ResultDescription
«unnamed»index

### gpu.grid_dim (::mlir::gpu::GridDimOp) ¶

Syntax:

operation ::= gpu.grid_dim $dimension attr-dict  Returns the number of thread blocks in the grid along the x, y, or z dimension. Example: %gDimZ = gpu.grid_dim z  Interfaces: InferIntRangeInterface, InferTypeOpInterface, NoSideEffect (MemoryEffectOpInterface) Effects: MemoryEffects::Effect{} #### Attributes: ¶ AttributeMLIR TypeDescription dimension::mlir::gpu::DimensionAttra dimension, either ‘x’, ‘y’, or ‘z’ #### Results: ¶ ResultDescription «unnamed»index ### gpu.host_register (::mlir::gpu::HostRegisterOp) ¶ Registers a memref for access from device. Syntax: operation ::= gpu.host_register$value attr-dict : type($value)  This op maps the provided host buffer into the device address space. This operation may not be supported in every environment, there is not yet a way to check at runtime whether this feature is supported. Writes from the host are guaranteed to be visible to device kernels that are launched afterwards. Writes from the device are guaranteed to be visible on the host after synchronizing with the device kernel completion. #### Operands: ¶ OperandDescription valueunranked.memref of any type values ### gpu.lane_id (::mlir::gpu::LaneIdOp) ¶ Syntax: operation ::= gpu.lane_id attr-dict  Returns the lane id within the subgroup (warp/wave). Example: %laneId = gpu.lane_id  Interfaces: InferIntRangeInterface, InferTypeOpInterface, NoSideEffect (MemoryEffectOpInterface) Effects: MemoryEffects::Effect{} #### Results: ¶ ResultDescription resultindex ### gpu.launch_func (::mlir::gpu::LaunchFuncOp) ¶ Launches a function as a GPU kernel Syntax: operation ::= gpu.launch_func custom<AsyncDependencies>(type($asyncToken), $asyncDependencies)$kernel
blocks in   ($gridSizeX,$gridSizeY, $gridSizeZ) threads in   ($blockSizeX, $blockSizeY,$blockSizeZ)
(dynamic_shared_memory_size $dynamicSharedMemorySize^)? custom<LaunchFuncOperands>($kernelOperands, type($kernelOperands)) attr-dict  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. 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 "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. 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>) }  Traits: AttrSizedOperandSegments Interfaces: GPU_AsyncOpInterface #### Attributes: ¶ AttributeMLIR TypeDescription kernel::mlir::SymbolRefAttrsymbol reference attribute #### Operands: ¶ OperandDescription asyncDependenciesasync token type gridSizeXindex gridSizeYindex gridSizeZindex blockSizeXindex blockSizeYindex blockSizeZindex dynamicSharedMemorySize32-bit signless integer kernelOperandsany type #### Results: ¶ ResultDescription asyncTokenasync token type ### gpu.launch (::mlir::gpu::LaunchOp) ¶ GPU kernel launch operation 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 twelve arguments, grouped as follows: • 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). Syntax: operation ::= gpu.launch (async ([ ssa-id-list ])? )? block ( ssa-id-list ) in ssa-reassignment threads ( ssa-id-list ) in ssa-reassignment (dynamic_shared_memory_size ssa-use)? region attr-dict? ssa-reassignment ::= ( ssa-id = ssa-use (, ssa-id = ssa-use)* )  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 }  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. Traits: AttrSizedOperandSegments, AutomaticAllocationScope Interfaces: GPU_AsyncOpInterface, InferIntRangeInterface #### Operands: ¶ OperandDescription asyncDependenciesasync token type gridSizeXindex gridSizeYindex gridSizeZindex blockSizeXindex blockSizeYindex blockSizeZindex dynamicSharedMemorySize32-bit signless integer #### Results: ¶ ResultDescription asyncTokenasync token type ### gpu.memcpy (::mlir::gpu::MemcpyOp) ¶ GPU memcpy operation Syntax: operation ::= gpu.memcpy custom<AsyncDependencies>(type($asyncToken), $asyncDependencies)$dst, $src : type($dst), type($src) attr-dict  The gpu.memcpy operation copies the content of one memref to another. The op does not execute before all async dependencies have finished executing. If the async keyword is present, the op is executed asynchronously (i.e. it does not block until the execution has finished on the device). In that case, it returns a !gpu.async.token. Example: %token = gpu.memcpy async [%dep] %dst, %src : memref<?xf32, 1>, memref<?xf32>  Interfaces: GPU_AsyncOpInterface #### Operands: ¶ OperandDescription asyncDependenciesasync token type dstmemref of any type values srcmemref of any type values #### Results: ¶ ResultDescription asyncTokenasync token type ### gpu.memset (::mlir::gpu::MemsetOp) ¶ GPU memset operation Syntax: operation ::= gpu.memset custom<AsyncDependencies>(type($asyncToken), $asyncDependencies)$dst, $value : type($dst), type($value) attr-dict  The gpu.memset operation sets the content of memref to a scalar value. The op does not execute before all async dependencies have finished executing. If the async keyword is present, the op is executed asynchronously (i.e. it does not block until the execution has finished on the device). In that case, it returns a !gpu.async.token. Example: %token = gpu.memset async [%dep] %dst, %value : memref<?xf32, 1>, f32  Interfaces: GPU_AsyncOpInterface #### Operands: ¶ OperandDescription asyncDependenciesasync token type dstmemref of any type values valueany type #### Results: ¶ ResultDescription asyncTokenasync token type ### gpu.module_end (::mlir::gpu::ModuleEndOp) ¶ A pseudo op that marks the end of a gpu.module. Syntax: operation ::= gpu.module_end attr-dict  This op terminates the only block inside the only region of a gpu.module. Traits: HasParent, Terminator ### gpu.num_subgroups (::mlir::gpu::NumSubgroupsOp) ¶ Syntax: operation ::= gpu.num_subgroups attr-dict : type($result)


Returns the number of subgroups within a workgroup.

Example:

%numSg = gpu.num_subgroups : index


Interfaces: InferIntRangeInterface, InferTypeOpInterface, NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

#### Results: ¶

ResultDescription
resultindex

### gpu.printf (::mlir::gpu::PrintfOp) ¶

Device-side printf, as in CUDA or OpenCL, for debugging

Syntax:

operation ::= gpu.printf $format attr-dict ($args^ : type($args))?  gpu.printf takes a literal format string format and an arbitrary number of scalar arguments that should be printed. The format string is a C-style printf string, subject to any restrictions imposed by one’s target platform. Interfaces: MemoryEffectOpInterface (MemoryEffectOpInterface) Effects: MemoryEffects::Effect{MemoryEffects::Write on ::mlir::SideEffects::DefaultResource} #### Attributes: ¶ AttributeMLIR TypeDescription format::mlir::StringAttrstring attribute #### Operands: ¶ OperandDescription argsinteger or index or floating-point ### gpu.return (::mlir::gpu::ReturnOp) ¶ Terminator for GPU functions. Syntax: operation ::= gpu.return attr-dict ($operands^ : type($operands))?  A terminator operation for regions that appear in the body of gpu.func functions. The operands to the gpu.return are the result values returned by an invocation of the gpu.func. Traits: HasParent, Terminator Interfaces: NoSideEffect (MemoryEffectOpInterface) Effects: MemoryEffects::Effect{} #### Operands: ¶ OperandDescription operandsany type ### gpu.set_default_device (::mlir::gpu::SetDefaultDeviceOp) ¶ Set default GPU for operations after this by index Syntax: operation ::= gpu.set_default_device attr-dict$devIndex


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.

Interfaces: MemoryEffectOpInterface (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{MemoryEffects::Write on ::mlir::SideEffects::DefaultResource}

#### Operands: ¶

OperandDescription
devIndex32-bit signless integer

### gpu.shuffle (::mlir::gpu::ShuffleOp) ¶

Shuffles values within a subgroup.

Syntax:

operation ::= gpu.shuffle $mode$value , $offset ,$width attr-dict : type($value)  The “shuffle” op moves values to a different invocation within the same subgroup. Example: %1, %2 = gpu.shuffle %0, %offset, %width xor : f32  For lane k returns the value from lane k ^ offset and true if that lane is smaller than %width. Otherwise it returns an unspecified value and false. A lane is the index of an invocation relative to its subgroup. The width specifies the number of invocations that participate in the shuffle. The width needs to be the same for all invocations that participate in the shuffle. Exactly the first width invocations of a subgroup need to execute this op in convergence. Interfaces: InferTypeOpInterface, NoSideEffect (MemoryEffectOpInterface) Effects: MemoryEffects::Effect{} #### Attributes: ¶ AttributeMLIR TypeDescription mode::mlir::gpu::ShuffleModeAttrIndexing modes supported by gpu.shuffle. #### Operands: ¶ OperandDescription valuei32 or f32 offset32-bit signless integer width32-bit signless integer #### Results: ¶ ResultDescription shuffleResulti32 or f32 valid1-bit signless integer ### gpu.subgroup_id (::mlir::gpu::SubgroupIdOp) ¶ Syntax: operation ::= gpu.subgroup_id attr-dict : type($result)


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

Example:

%sgId = gpu.subgroup_id : index


Interfaces: InferIntRangeInterface, InferTypeOpInterface, NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

#### Results: ¶

ResultDescription
resultindex

### gpu.subgroup_mma_compute (::mlir::gpu::SubgroupMmaComputeOp) ¶

GPU warp synchronous matrix multiply accumulate

Syntax:

operation ::= gpu.subgroup_mma_compute $opA,$opB, $opC attr-dict : type($opA), type($opB) -> type($res)


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_matrixs as arguments: these hold A, B and Coperands 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.

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">


Interfaces: InferTypeOpInterface, NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

#### Operands: ¶

OperandDescription
opAgpu.mma_matrix of 16-bit float or 32-bit float values
opBgpu.mma_matrix of 16-bit float or 32-bit float values
opCgpu.mma_matrix of 16-bit float or 32-bit float values

#### Results: ¶

ResultDescription
resMMAMatrix type

### gpu.subgroup_mma_constant_matrix (::mlir::gpu::SubgroupMmaConstantMatrixOp) ¶

GPU warp synchronous constant matrix

Syntax:

operation ::= gpu.subgroup_mma_constant_matrix $value attr-dict : type($res)


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">


Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

#### Operands: ¶

OperandDescription
value16-bit float or 32-bit float

#### Results: ¶

ResultDescription
resMMAMatrix type

### gpu.subgroup_mma_elementwise (::mlir::gpu::SubgroupMmaElementwiseOp) ¶

GPU warp elementwise operation on a matrix

Syntax:

operation ::= gpu.subgroup_mma_elementwise $opType$args attr-dict : functional-type($args,$res)


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">


Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

#### Attributes: ¶

AttributeMLIR TypeDescription
opType::mlir::gpu::MMAElementwiseOpAttrelementwise operation to apply to mma matrix

#### Operands: ¶

OperandDescription
argsMMAMatrix type

#### Results: ¶

ResultDescription
resMMAMatrix type

### gpu.subgroup_mma_load_matrix (::mlir::gpu::SubgroupMmaLoadMatrixOp) ¶

Syntax:

operation ::= gpu.subgroup_mma_load_matrix $srcMemref[$indices] attr-dict : type($srcMemref) -> type($res)


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.

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">


Interfaces: MemoryEffectOpInterface (MemoryEffectOpInterface)

#### Attributes: ¶

AttributeMLIR TypeDescription
leadDimension::mlir::IntegerAttrindex attribute

#### Operands: ¶

OperandDescription
srcMemrefmemref of 16-bit float or 32-bit float values
indicesindex

#### Results: ¶

ResultDescription
resMMAMatrix type

### gpu.subgroup_mma_store_matrix (::mlir::gpu::SubgroupMmaStoreMatrixOp) ¶

GPU warp synchronous matrix store

Syntax:

operation ::= gpu.subgroup_mma_store_matrix $src,$dstMemref[$indices] attr-dict : type($src), type($dstMemref)  The gpu.subgroup_mma_store_matrix operation stores a matrix collectively using all the threads in a subgroup. This operation takes a !gpu.mma_matrix and a memref as operands. !gpu.mma_matrix is the source value containing the data to be stored into the destination memref which can be in global or shared memory. The store address is determined using the indices provided. The leadDimension attribute specifies the leading dimension of the destination matrix. This op is often meant to be used along with gpu.subgroup_mma_load_matrix and gpu.subgroup_mma_compute. Example: gpu.subgroup_mma_store_matrix %D, %sg[%i,%j] : { leadDimension = 32 : i32} : !gpu.mma_matrix<16x16xf16, "COp">, memref<32x32xf16, 3>  Interfaces: MemoryEffectOpInterface (MemoryEffectOpInterface) Effects: MemoryEffects::Effect{MemoryEffects::Write on ::mlir::SideEffects::DefaultResource} #### Attributes: ¶ AttributeMLIR TypeDescription leadDimension::mlir::IntegerAttrindex attribute #### Operands: ¶ OperandDescription srcgpu.mma_matrix of 16-bit float or 32-bit float values dstMemrefmemref of 16-bit float or 32-bit float values indicesindex ### gpu.subgroup_size (::mlir::gpu::SubgroupSizeOp) ¶ Syntax: operation ::= gpu.subgroup_size attr-dict : type($result)


Returns the number of threads within a subgroup.

Example:

%sgSz = gpu.subgroup_size : index


Interfaces: InferIntRangeInterface, InferTypeOpInterface, NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

#### Results: ¶

ResultDescription
resultindex

### gpu.terminator (::mlir::gpu::TerminatorOp) ¶

Terminator for GPU launch regions.

Syntax:

operation ::= gpu.terminator attr-dict


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.

Traits: HasParent, Terminator

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

### gpu.thread_id (::mlir::gpu::ThreadIdOp) ¶

Syntax:

operation ::= gpu.thread_id $dimension attr-dict  Returns the thread id, i.e. the index of the current thread within the block along the x, y, or z dimension. Example: %tIdX = gpu.thread_id x  Interfaces: InferIntRangeInterface, InferTypeOpInterface, NoSideEffect (MemoryEffectOpInterface) Effects: MemoryEffects::Effect{} #### Attributes: ¶ AttributeMLIR TypeDescription dimension::mlir::gpu::DimensionAttra dimension, either ‘x’, ‘y’, or ‘z’ #### Results: ¶ ResultDescription «unnamed»index ### gpu.wait (::mlir::gpu::WaitOp) ¶ Wait for async gpu ops to complete. Syntax: operation ::= gpu.wait custom<AsyncDependencies>(type($asyncToken), \$asyncDependencies) attr-dict


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]


Interfaces: GPU_AsyncOpInterface

#### Operands: ¶

OperandDescription
asyncDependenciesasync token type

#### Results: ¶

ResultDescription
asyncTokenasync token type

### gpu.yield (::mlir::gpu::YieldOp) ¶

GPU yield operation

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


Traits: Terminator

Interfaces: NoSideEffect (MemoryEffectOpInterface)

Effects: MemoryEffects::Effect{}

#### Operands: ¶

OperandDescription
values`any type