Adding ThreadblockSwizzle

Hi,

I have the following code CUTLASS kernel that uses GemmSplitKParallel (code 1) and I want to add ThreadblockSwizzle and Stages to the GemmSplitKParallel. I changed it as code 2 (in the following). But it is not compiled. Please help me how to fix the bug. Thank you.

code 1 ———————————-

using ElementAccumulator = float; // accumulator type

using ElementComputeEpilogue = ElementAccumulator; // epilogue compute type

using ElementInputA = cutlass::half_t; // input A type (FP16)

using ElementInputB = cutlass::half_t; // input B type (FP16)

using ElementOutput = float; // output type (FP32)

using LayoutInputA = cutlass::layout::ColumnMajor;

using LayoutInputB = cutlass::layout::RowMajor;

using LayoutOutput = cutlass::layout::RowMajor;

using MMAOp = cutlass::arch::OpClassTensorOp;

using SmArch = cutlass::arch::Sm80;

using ShapeMMAThreadBlock =

cutlass::gemm::GemmShape<128, 128, 32>; // threadblock tile (M,N,K) //default

using ShapeMMAWarp = cutlass::gemm::GemmShape<64, 64, 32>; // warp tile //default

using ShapeMMAOp = cutlass::gemm::GemmShape<16, 8, 8>; // MMA Op tile (may be tuned)
using EpilogueOp = cutlass::epilogue::thread::LinearCombination<

ElementOutput,                                     // output element type

128 / cutlass::sizeof_bits::value, // elements per vectorized access (128 bits / bits-per-element)

ElementAccumulator,                                // accumulator type

ElementComputeEpilogue>;                           // compute type for alpha/beta

// CUTLASS GemmSplitKParallel instantiation

using Gemm = cutlass::gemm::device::GemmSplitKParallel<

ElementInputA,

LayoutInputA,

ElementInputB,

LayoutInputB,

ElementOutput,

LayoutOutput,

ElementAccumulator,

MMAOp,

SmArch,

ShapeMMAThreadBlock,

ShapeMMAWarp,

ShapeMMAOp,

EpilogueOp>;

code 2————————————————————————-

using ElementAccumulator = float; // accumulator type

using ElementComputeEpilogue = ElementAccumulator; // epilogue compute type

using ElementInputA = cutlass::half_t; // input A type (FP16)

using ElementInputB = cutlass::half_t; // input B type (FP16)

using ElementOutput = float; // output type (FP32)

using LayoutInputA = cutlass::layout::ColumnMajor;

using LayoutInputB = cutlass::layout::RowMajor;

using LayoutOutput = cutlass::layout::RowMajor;

using MMAOp = cutlass::arch::OpClassTensorOp;

using SmArch = cutlass::arch::Sm80;

using ShapeMMAThreadBlock =

cutlass::gemm::GemmShape<128, 128, 32>; // threadblock tile (M,N,K) //default

using ShapeMMAWarp = cutlass::gemm::GemmShape<64, 64, 32>; // warp tile //default

using ShapeMMAOp = cutlass::gemm::GemmShape<16, 8, 8>; // MMA Op tile (may be tuned)
using EpilogueOp = cutlass::epilogue::thread::LinearCombination<

ElementOutput,                                     // output element type

128 / cutlass::sizeof_bits::value, // elements per vectorized access (128 bits / bits-per-element)

ElementAccumulator,                                // accumulator type

ElementComputeEpilogue>;                           // compute type for alpha/beta

using ThreadblockSwizzle = cutlass::gemm::threadblock::GemmIdentityThreadblockSwizzle<>;

constexpr int Stages = 2;   // typical values: 2 or 3 (choose based on latency / occupancy)

// CUTLASS GemmSplitKParallel instantiation

using Gemm = cutlass::gemm::device::GemmSplitKParallel<

ElementInputA,

LayoutInputA,

ElementInputB,

LayoutInputB,

ElementOutput,

LayoutOutput,

ElementAccumulator,

MMAOp,

SmArch,

ShapeMMAThreadBlock,

ShapeMMAWarp,

ShapeMMAOp,

EpilogueOp,

ThreadblockSwizzle,

Stages

>;