![]() |
CUTLASS
CUDA Templates for Linear Algebra Subroutines and Solvers
|
#include <gemm_splitk_parallel.h>
Classes | |
| struct | Arguments |
| Argument structure. More... | |
Public Types | |
| using | ElementA = ElementA_ |
| using | LayoutA = LayoutA_ |
| using | ElementB = ElementB_ |
| using | LayoutB = LayoutB_ |
| using | ElementC = ElementC_ |
| using | LayoutC = LayoutC_ |
| using | ElementAccumulator = ElementAccumulator_ |
| using | OperatorClass = OperatorClass_ |
| using | ArchTag = ArchTag_ |
| using | ThreadblockShape = ThreadblockShape_ |
| using | WarpShape = WarpShape_ |
| using | InstructionShape = InstructionShape_ |
| using | ConvertScaledOp = ConvertScaledOp_ |
| using | EpilogueOutputOp = EpilogueOutputOp_ |
| using | ReductionOp = ReductionOp_ |
| using | ThreadblockSwizzle = ThreadblockSwizzle_ |
| using | Operator = Operator_ |
| using | GemmKernel = typename kernel::DefaultGemmSplitKParallel< ElementA, LayoutA, kAlignmentA, ElementB, LayoutB, kAlignmentB, ElementAccumulator, LayoutC, ElementAccumulator, OperatorClass, ArchTag, ThreadblockShape, WarpShape, InstructionShape, ConvertScaledOp, ThreadblockSwizzle, kStages, Operator >::GemmKernel |
| GEMM kernel. More... | |
| using | ReductionKernel = cutlass::reduction::kernel::ReduceSplitK< cutlass::MatrixShape< 4, 32 *EpilogueOutputOp::kCount >, EpilogueOutputOp, ReductionOp > |
| Reduction kernel. More... | |
Public Member Functions | |
| GemmSplitKParallel () | |
| Constructs the GEMM. More... | |
| Status | initialize (Arguments const &args, void *workspace) |
| Initializes GEMM state from arguments. More... | |
| Status | update (Arguments const &args, void *workspace=nullptr) |
| Lightweight update given a subset of arguments. More... | |
| Status | run (cudaStream_t stream=nullptr) |
| Runs the kernel using initialized state. More... | |
| Status | operator() (cudaStream_t stream=nullptr) |
| Runs the kernel using initialized state. More... | |
| Status | operator() (Arguments const &args, void *workspace=nullptr, cudaStream_t stream=nullptr) |
| Runs the kernel using initialized state. More... | |
Static Public Member Functions | |
| static Status | can_implement (Arguments const &args) |
| Determines whether the GEMM can execute the given problem. More... | |
| static size_t | get_workspace_size (Arguments const &args) |
| Gets the workspace size. More... | |
Static Public Attributes | |
| static int const | kStages = Stages |
Gemm device-level operator performing parallel reduction over the K partition.
| using cutlass::gemm::device::GemmSplitKParallel< ElementA_, LayoutA_, ElementB_, LayoutB_, ElementC_, LayoutC_, ElementAccumulator_, OperatorClass_, ArchTag_, ThreadblockShape_, WarpShape_, InstructionShape_, EpilogueOutputOp_, ConvertScaledOp_, ReductionOp_, ThreadblockSwizzle_, Stages, kAlignmentA, kAlignmentB, Operator_ >::ArchTag = ArchTag_ |
| using cutlass::gemm::device::GemmSplitKParallel< ElementA_, LayoutA_, ElementB_, LayoutB_, ElementC_, LayoutC_, ElementAccumulator_, OperatorClass_, ArchTag_, ThreadblockShape_, WarpShape_, InstructionShape_, EpilogueOutputOp_, ConvertScaledOp_, ReductionOp_, ThreadblockSwizzle_, Stages, kAlignmentA, kAlignmentB, Operator_ >::ConvertScaledOp = ConvertScaledOp_ |
| using cutlass::gemm::device::GemmSplitKParallel< ElementA_, LayoutA_, ElementB_, LayoutB_, ElementC_, LayoutC_, ElementAccumulator_, OperatorClass_, ArchTag_, ThreadblockShape_, WarpShape_, InstructionShape_, EpilogueOutputOp_, ConvertScaledOp_, ReductionOp_, ThreadblockSwizzle_, Stages, kAlignmentA, kAlignmentB, Operator_ >::ElementA = ElementA_ |
| using cutlass::gemm::device::GemmSplitKParallel< ElementA_, LayoutA_, ElementB_, LayoutB_, ElementC_, LayoutC_, ElementAccumulator_, OperatorClass_, ArchTag_, ThreadblockShape_, WarpShape_, InstructionShape_, EpilogueOutputOp_, ConvertScaledOp_, ReductionOp_, ThreadblockSwizzle_, Stages, kAlignmentA, kAlignmentB, Operator_ >::ElementAccumulator = ElementAccumulator_ |
| using cutlass::gemm::device::GemmSplitKParallel< ElementA_, LayoutA_, ElementB_, LayoutB_, ElementC_, LayoutC_, ElementAccumulator_, OperatorClass_, ArchTag_, ThreadblockShape_, WarpShape_, InstructionShape_, EpilogueOutputOp_, ConvertScaledOp_, ReductionOp_, ThreadblockSwizzle_, Stages, kAlignmentA, kAlignmentB, Operator_ >::ElementB = ElementB_ |
| using cutlass::gemm::device::GemmSplitKParallel< ElementA_, LayoutA_, ElementB_, LayoutB_, ElementC_, LayoutC_, ElementAccumulator_, OperatorClass_, ArchTag_, ThreadblockShape_, WarpShape_, InstructionShape_, EpilogueOutputOp_, ConvertScaledOp_, ReductionOp_, ThreadblockSwizzle_, Stages, kAlignmentA, kAlignmentB, Operator_ >::ElementC = ElementC_ |
| using cutlass::gemm::device::GemmSplitKParallel< ElementA_, LayoutA_, ElementB_, LayoutB_, ElementC_, LayoutC_, ElementAccumulator_, OperatorClass_, ArchTag_, ThreadblockShape_, WarpShape_, InstructionShape_, EpilogueOutputOp_, ConvertScaledOp_, ReductionOp_, ThreadblockSwizzle_, Stages, kAlignmentA, kAlignmentB, Operator_ >::EpilogueOutputOp = EpilogueOutputOp_ |
| using cutlass::gemm::device::GemmSplitKParallel< ElementA_, LayoutA_, ElementB_, LayoutB_, ElementC_, LayoutC_, ElementAccumulator_, OperatorClass_, ArchTag_, ThreadblockShape_, WarpShape_, InstructionShape_, EpilogueOutputOp_, ConvertScaledOp_, ReductionOp_, ThreadblockSwizzle_, Stages, kAlignmentA, kAlignmentB, Operator_ >::GemmKernel = typename kernel::DefaultGemmSplitKParallel< ElementA, LayoutA, kAlignmentA, ElementB, LayoutB, kAlignmentB, ElementAccumulator, LayoutC, ElementAccumulator, OperatorClass, ArchTag, ThreadblockShape, WarpShape, InstructionShape, ConvertScaledOp, ThreadblockSwizzle, kStages, Operator >::GemmKernel |
| using cutlass::gemm::device::GemmSplitKParallel< ElementA_, LayoutA_, ElementB_, LayoutB_, ElementC_, LayoutC_, ElementAccumulator_, OperatorClass_, ArchTag_, ThreadblockShape_, WarpShape_, InstructionShape_, EpilogueOutputOp_, ConvertScaledOp_, ReductionOp_, ThreadblockSwizzle_, Stages, kAlignmentA, kAlignmentB, Operator_ >::InstructionShape = InstructionShape_ |
| using cutlass::gemm::device::GemmSplitKParallel< ElementA_, LayoutA_, ElementB_, LayoutB_, ElementC_, LayoutC_, ElementAccumulator_, OperatorClass_, ArchTag_, ThreadblockShape_, WarpShape_, InstructionShape_, EpilogueOutputOp_, ConvertScaledOp_, ReductionOp_, ThreadblockSwizzle_, Stages, kAlignmentA, kAlignmentB, Operator_ >::LayoutA = LayoutA_ |
| using cutlass::gemm::device::GemmSplitKParallel< ElementA_, LayoutA_, ElementB_, LayoutB_, ElementC_, LayoutC_, ElementAccumulator_, OperatorClass_, ArchTag_, ThreadblockShape_, WarpShape_, InstructionShape_, EpilogueOutputOp_, ConvertScaledOp_, ReductionOp_, ThreadblockSwizzle_, Stages, kAlignmentA, kAlignmentB, Operator_ >::LayoutB = LayoutB_ |
| using cutlass::gemm::device::GemmSplitKParallel< ElementA_, LayoutA_, ElementB_, LayoutB_, ElementC_, LayoutC_, ElementAccumulator_, OperatorClass_, ArchTag_, ThreadblockShape_, WarpShape_, InstructionShape_, EpilogueOutputOp_, ConvertScaledOp_, ReductionOp_, ThreadblockSwizzle_, Stages, kAlignmentA, kAlignmentB, Operator_ >::LayoutC = LayoutC_ |
| using cutlass::gemm::device::GemmSplitKParallel< ElementA_, LayoutA_, ElementB_, LayoutB_, ElementC_, LayoutC_, ElementAccumulator_, OperatorClass_, ArchTag_, ThreadblockShape_, WarpShape_, InstructionShape_, EpilogueOutputOp_, ConvertScaledOp_, ReductionOp_, ThreadblockSwizzle_, Stages, kAlignmentA, kAlignmentB, Operator_ >::Operator = Operator_ |
| using cutlass::gemm::device::GemmSplitKParallel< ElementA_, LayoutA_, ElementB_, LayoutB_, ElementC_, LayoutC_, ElementAccumulator_, OperatorClass_, ArchTag_, ThreadblockShape_, WarpShape_, InstructionShape_, EpilogueOutputOp_, ConvertScaledOp_, ReductionOp_, ThreadblockSwizzle_, Stages, kAlignmentA, kAlignmentB, Operator_ >::OperatorClass = OperatorClass_ |
| using cutlass::gemm::device::GemmSplitKParallel< ElementA_, LayoutA_, ElementB_, LayoutB_, ElementC_, LayoutC_, ElementAccumulator_, OperatorClass_, ArchTag_, ThreadblockShape_, WarpShape_, InstructionShape_, EpilogueOutputOp_, ConvertScaledOp_, ReductionOp_, ThreadblockSwizzle_, Stages, kAlignmentA, kAlignmentB, Operator_ >::ReductionKernel = cutlass::reduction::kernel::ReduceSplitK< cutlass::MatrixShape<4, 32 * EpilogueOutputOp::kCount>, EpilogueOutputOp, ReductionOp > |
| using cutlass::gemm::device::GemmSplitKParallel< ElementA_, LayoutA_, ElementB_, LayoutB_, ElementC_, LayoutC_, ElementAccumulator_, OperatorClass_, ArchTag_, ThreadblockShape_, WarpShape_, InstructionShape_, EpilogueOutputOp_, ConvertScaledOp_, ReductionOp_, ThreadblockSwizzle_, Stages, kAlignmentA, kAlignmentB, Operator_ >::ReductionOp = ReductionOp_ |
| using cutlass::gemm::device::GemmSplitKParallel< ElementA_, LayoutA_, ElementB_, LayoutB_, ElementC_, LayoutC_, ElementAccumulator_, OperatorClass_, ArchTag_, ThreadblockShape_, WarpShape_, InstructionShape_, EpilogueOutputOp_, ConvertScaledOp_, ReductionOp_, ThreadblockSwizzle_, Stages, kAlignmentA, kAlignmentB, Operator_ >::ThreadblockShape = ThreadblockShape_ |
| using cutlass::gemm::device::GemmSplitKParallel< ElementA_, LayoutA_, ElementB_, LayoutB_, ElementC_, LayoutC_, ElementAccumulator_, OperatorClass_, ArchTag_, ThreadblockShape_, WarpShape_, InstructionShape_, EpilogueOutputOp_, ConvertScaledOp_, ReductionOp_, ThreadblockSwizzle_, Stages, kAlignmentA, kAlignmentB, Operator_ >::ThreadblockSwizzle = ThreadblockSwizzle_ |
| using cutlass::gemm::device::GemmSplitKParallel< ElementA_, LayoutA_, ElementB_, LayoutB_, ElementC_, LayoutC_, ElementAccumulator_, OperatorClass_, ArchTag_, ThreadblockShape_, WarpShape_, InstructionShape_, EpilogueOutputOp_, ConvertScaledOp_, ReductionOp_, ThreadblockSwizzle_, Stages, kAlignmentA, kAlignmentB, Operator_ >::WarpShape = WarpShape_ |
|
inline |
|
inlinestatic |
|
inlinestatic |
|
inline |
|
inline |
|
inline |
|
inline |
|
inline |
|
static |
1.8.11