Skip to content

GitLab

Commit 4769425e authored by Chao Liu's avatar Chao Liu
Browse files

Merge remote-tracking branch 'origin/develop' into gelu

parents b548c0be ba58a93f
Hide whitespace changes
Inline Side-by-side
Showing with 475 additions and 718 deletions
library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp
View file @ 4769425e
......@@ -10,8 +10,9 @@ namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using F16 = ck::half_t;
using F32 = float;
using DPtrsGlobal = ck::Tuple<F32*, F32*>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
......@@ -21,41 +22,52 @@ using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::reduce::Add<F32>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using ReduceOps = ck::Tuple<ReduceSum, ReduceSum>;
using Identity = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Identity, Identity>;
using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
ck::InMemoryDataOperationEnum::AtomicAdd>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// c[g, m, n] = a[g, m, k] * b[g, n, k]
// d0[g, m] = reduce0(c[g, m, n])
// d1[g, m] = reduce1(c[g, m, n])
using device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instances =
std::tuple<
// clang-format off
//##################################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| D0| D1| D1EleOp| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//##################################| | | | Type| Type| Type| DataType| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//##################################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 4, 4, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 2, 2, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 2, 2, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 2, 2, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>
//##################################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//##################################| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData|Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//##################################| | | | | | | | | | | Operation| Operation| Operation| Operation| | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 4, 4, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 2, 2, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 2, 2, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 2, 2, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>
// clang-format on
>;
void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instances(
std::vector<DeviceGemmReducePtr<PassThrough, PassThrough, PassThrough, Square>>& instances)
std::vector<DeviceGemmReducePtr<DPtrsGlobal,
PassThrough,
PassThrough,
PassThrough,
DInElementOps,
DOutElementOps>>& instances)
{
add_device_operation_instances(
instances,
......
library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp
View file @ 4769425e
......@@ -10,8 +10,9 @@ namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using F16 = ck::half_t;
using F32 = float;
using DPtrsGlobal = ck::Tuple<F32*, F32*>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
......@@ -21,41 +22,52 @@ using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::reduce::Add<F32>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using ReduceOps = ck::Tuple<ReduceSum, ReduceSum>;
using Identity = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Identity, Identity>;
using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
ck::InMemoryDataOperationEnum::AtomicAdd>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// c[g, m, n] = a[g, m, k] * b[g, n, k]
// d0[g, m] = reduce0(c[g, m, n])
// d1[g, m] = reduce1(c[g, m, n])
using device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instances =
std::tuple<
// clang-format off
//##################################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| D0| D1| D1EleOp| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//##################################| | | | Type| Type| Type| DataType| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//##################################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 2, 8, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 2, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 2, 8, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 2, 8, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>
//##################################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//##################################| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData|Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//##################################| | | | | | | | | | | Operation| Operation| Operation| Operation| | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 2, 8, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 2, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 2, 8, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 2, 8, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>
// clang-format on
>;
void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instances(
std::vector<DeviceGemmReducePtr<PassThrough, PassThrough, PassThrough, Square>>& instances)
std::vector<DeviceGemmReducePtr<DPtrsGlobal,
PassThrough,
PassThrough,
PassThrough,
DInElementOps,
DOutElementOps>>& instances)
{
add_device_operation_instances(
instances,
......
library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp
View file @ 4769425e
......@@ -10,8 +10,9 @@ namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using F16 = ck::half_t;
using F32 = float;
using DPtrsGlobal = ck::Tuple<F32*, F32*>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
......@@ -21,41 +22,52 @@ using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::reduce::Add<F32>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using ReduceOps = ck::Tuple<ReduceSum, ReduceSum>;
using Identity = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Identity, Identity>;
using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
ck::InMemoryDataOperationEnum::AtomicAdd>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// c[g, m, n] = a[g, m, k] * b[g, n, k]
// d0[g, m] = reduce0(c[g, m, n])
// d1[g, m] = reduce1(c[g, m, n])
using device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances =
std::tuple<
// clang-format off
//##################################| ALayout| BLayout| CLayout| AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| D0| D1| D1EleOp| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//##################################| | | | Type| Type| Type| DataType| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//##################################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 8, 2, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 8, 2, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 8, 2, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>
//##################################| ALayout| BLayout| CLayout| AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//##################################| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData|Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//##################################| | | | | | | | | | | Operation| Operation| Operation| Operation| | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 8, 2, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 8, 2, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 8, 2, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>
// clang-format on
>;
void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances(
std::vector<DeviceGemmReducePtr<PassThrough, PassThrough, PassThrough, Square>>& instances)
std::vector<DeviceGemmReducePtr<DPtrsGlobal,
PassThrough,
PassThrough,
PassThrough,
DInElementOps,
DOutElementOps>>& instances)
{
add_device_operation_instances(
instances,
......
library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp
View file @ 4769425e
......@@ -10,8 +10,9 @@ namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using F16 = ck::half_t;
using F32 = float;
using DPtrsGlobal = ck::Tuple<F32*, F32*>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
......@@ -21,38 +22,49 @@ using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::reduce::Add<F32>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using ReduceOps = ck::Tuple<ReduceSum, ReduceSum>;
using Identity = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Identity, Identity>;
using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
ck::InMemoryDataOperationEnum::AtomicAdd>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// c[g, m, n] = a[g, m, k] * b[g, n, k]
// d0[g, m] = reduce0(c[g, m, n])
// d1[g, m] = reduce1(c[g, m, n])
using device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instances =
std::tuple<
// clang-format off
//##################################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| D0| D1| D1EleOp| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//##################################| | | | Type| Type| Type| DataType| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//##################################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<32, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<32, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<32, 2>, 4, 1>
//##################################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//##################################| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData|Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//##################################| | | | | | | | | | | Operation| Operation| Operation| Operation| | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<32, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<32, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<32, 2>, 4, 1>
// clang-format on
>;
void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instances(
std::vector<DeviceGemmReducePtr<PassThrough, PassThrough, PassThrough, Square>>& instances)
std::vector<DeviceGemmReducePtr<DPtrsGlobal,
PassThrough,
PassThrough,
PassThrough,
DInElementOps,
DOutElementOps>>& instances)
{
add_device_operation_instances(
instances,
......
library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instance.cpp
View file @ 4769425e
......@@ -10,8 +10,9 @@ namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using F16 = ck::half_t;
using F32 = float;
using DPtrsGlobal = ck::Tuple<F32*, F32*>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
......@@ -21,40 +22,51 @@ using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::reduce::Add<F32>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using ReduceOps = ck::Tuple<ReduceSum, ReduceSum>;
using Identity = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Identity, Identity>;
using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
ck::InMemoryDataOperationEnum::AtomicAdd>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// c[m, n] = a[k, m] * b[k, n]
// d0[m] = reduce0(c[m, n])
// d1[m] = reduce1(c[m, n])
using device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instances = std::tuple<
// clang-format off
//###########################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| D0| D1| D1EleOp| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//###########################| | | | Type| Type| Type| DataType| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//###########################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//###########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 2, 2, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 2, 2, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 2, 2, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 2, 2, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>
//###########################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//###########################| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData|Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//###########################| | | | | | | | | | | Operation| Operation| Operation| Operation| | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//###########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 2, 2, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 2, 2, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 2, 2, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 2, 2, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>
// clang-format on
>;
void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instances(
std::vector<DeviceGemmReducePtr<PassThrough, PassThrough, PassThrough, Square>>& instances)
std::vector<DeviceGemmReducePtr<DPtrsGlobal,
PassThrough,
PassThrough,
PassThrough,
DInElementOps,
DOutElementOps>>& instances)
{
add_device_operation_instances(
instances, device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_kn_mn_instances{});
......
library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instance.cpp
View file @ 4769425e
......@@ -10,8 +10,9 @@ namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using F16 = ck::half_t;
using F32 = float;
using DPtrsGlobal = ck::Tuple<F32*, F32*>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
......@@ -21,40 +22,51 @@ using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::reduce::Add<F32>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using ReduceOps = ck::Tuple<ReduceSum, ReduceSum>;
using Identity = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Identity, Identity>;
using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
ck::InMemoryDataOperationEnum::AtomicAdd>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// c[m, n] = a[k, m] * b[n, k]
// d0[m] = reduce0(c[m, n])
// d1[m] = reduce1(c[m, n])
using device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instances = std::tuple<
// clang-format off
//###########################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| D0| D1| D1EleOp| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//###########################| | | | Type| Type| Type| DataType| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//###########################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//###########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 2, 8, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 2, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 2, 8, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 2, 8, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>
//###########################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//###########################| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData|Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//###########################| | | | | | | | | | | Operation| Operation| Operation| Operation| | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//###########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 2, 8, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 2, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 2, 8, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 2, 8, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>
// clang-format on
>;
void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instances(
std::vector<DeviceGemmReducePtr<PassThrough, PassThrough, PassThrough, Square>>& instances)
std::vector<DeviceGemmReducePtr<DPtrsGlobal,
PassThrough,
PassThrough,
PassThrough,
DInElementOps,
DOutElementOps>>& instances)
{
add_device_operation_instances(
instances, device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_km_nk_mn_instances{});
......
library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instance.cpp
View file @ 4769425e
......@@ -10,8 +10,9 @@ namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using F16 = ck::half_t;
using F32 = float;
using DPtrsGlobal = ck::Tuple<F32*, F32*>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
......@@ -21,40 +22,51 @@ using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::reduce::Add<F32>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using ReduceOps = ck::Tuple<ReduceSum, ReduceSum>;
using Identity = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Identity, Identity>;
using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
ck::InMemoryDataOperationEnum::AtomicAdd>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// c[m, n] = a[m, k] * b[n, k]
// d0[m] = reduce0(c[m, n])
// d1[m] = reduce1(c[m, n])
using device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances = std::tuple<
// clang-format off
//###########################| ALayout| BLayout| CLayout| AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| D0| D1| D1EleOp| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//###########################| | | | Type| Type| Type| DataType| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//###########################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//###########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 8, 2, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 8, 2, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 8, 2, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>
//###########################| ALayout| BLayout| CLayout| AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//###########################| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData|Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//###########################| | | | | | | | | | | Operation| Operation| Operation| Operation| | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//###########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 8, 2, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 8, 2, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 8, 2, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>
// clang-format on
>;
void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances(
std::vector<DeviceGemmReducePtr<PassThrough, PassThrough, PassThrough, Square>>& instances)
std::vector<DeviceGemmReducePtr<DPtrsGlobal,
PassThrough,
PassThrough,
PassThrough,
DInElementOps,
DOutElementOps>>& instances)
{
add_device_operation_instances(
instances, device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances{});
......
library/src/tensor_operation_instance/gpu/gemm_reduce/device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instance.cpp
View file @ 4769425e
......@@ -10,8 +10,9 @@ namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using F16 = ck::half_t;
using F32 = float;
using DPtrsGlobal = ck::Tuple<F32*, F32*>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
......@@ -21,37 +22,48 @@ using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::reduce::Add<F32>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using ReduceOps = ck::Tuple<ReduceSum, ReduceSum>;
using Identity = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Identity, Identity>;
using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
ck::InMemoryDataOperationEnum::AtomicAdd>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// c[m, n] = a[m, k] * b[n, k]
// d0[m] = reduce0(c[m, n])
// d1[m] = reduce1(c[m, n])
using device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instances = std::tuple<
// clang-format off
//###########################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| D0| D1| D1EleOp| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//###########################| | | | Type| Type| Type| DataType| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//###########################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//###########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<32, 2>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<32, 2>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<32, 2>, 4, 1>
//###########################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//###########################| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//###########################| | | | | | | | | | | Operation| Operation| Operation| Operation| | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//###########################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<32, 2>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<32, 2>, 4, 1>,
DeviceGemmReduce_Xdl_CShuffle< Row, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<32, 2>, 4, 1>
// clang-format on
>;
void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instances(
std::vector<DeviceGemmReducePtr<PassThrough, PassThrough, PassThrough, Square>>& instances)
std::vector<DeviceGemmReducePtr<DPtrsGlobal,
PassThrough,
PassThrough,
PassThrough,
DInElementOps,
DOutElementOps>>& instances)
{
add_device_operation_instances(
instances, device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_nk_mn_instances{});
......
profiler/include/profile_batched_gemm_reduce_impl.hpp
View file @ 4769425e
......@@ -17,11 +17,21 @@ namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F32 = float;
using F16 = ck::half_t;
using DPtrsGlobal = ck::Tuple<F32*, F32*>;
using Identity = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Identity, Identity>;
using DeviceGemmReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmReducePtr<
DPtrsGlobal,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::UnarySquare<float, float, false>>;
DInElementOps,
DOutElementOps>;
void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances(
std::vector<DeviceGemmReduceNoOpPtr>&);
......@@ -119,19 +129,25 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
b_g_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
}
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::reduce::Add<float>;
using D1ReduceOp = ck::reduce::Add<float>;
using D1ElementOp = ck::tensor_operation::element_wise::UnarySquare<float, float, false>;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
const auto d0_reduce_op = D0ReduceOp{};
const auto d1_reduce_op = D1ReduceOp{};
const auto d1_element_op = D1ElementOp{};
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::reduce::Add<float>;
using D1ReduceOp = ck::reduce::Add<float>;
using UnaryIdenticElementOp =
ck::tensor_operation::element_wise::UnaryIdentic<float, float, false>;
using UnarySquareElementOp =
ck::tensor_operation::element_wise::UnarySquare<float, float, false>;
using DxsInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOps = ck::Tuple<UnaryIdenticElementOp, UnaryIdenticElementOp>;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
const auto dxs_in_element_op = DxsInElementOps{};
const auto dxs_out_element_op = DxsOutElementOps{};
const auto d0_reduce_op = D0ReduceOp{};
const auto d1_reduce_op = D1ReduceOp{};
if(do_verification)
{
......@@ -163,7 +179,7 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
float d0_val = ck::type_convert<float>(c_g_m_n_host_result(batch, m, n));
float d1_val;
d1_element_op(d1_val, d0_val);
UnarySquareElementOp{}(d1_val, d0_val);
d0_reduce_op(d0_acc, d0_val);
d1_reduce_op(d1_acc, d1_val);
}
......@@ -180,6 +196,9 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
DeviceMem d0_device_buf(sizeof(DDataType) * d0_g_m_device_result.mDesc.GetElementSpace());
DeviceMem d1_device_buf(sizeof(DDataType) * d1_g_m_device_result.mDesc.GetElementSpace());
auto dxs_global = ck::make_tuple(static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()));
a_device_buf.ToDevice(a_g_m_k.mData.data());
b_device_buf.ToDevice(b_g_k_n.mData.data());
......@@ -241,8 +260,7 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
gemm_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()),
dxs_global,
M,
N,
K,
......@@ -252,7 +270,8 @@ bool profile_batched_gemm_reduce_impl(int do_verification,
a_element_op,
b_element_op,
c_element_op,
d1_element_op,
dxs_in_element_op,
dxs_out_element_op,
BatchCount);
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
......
profiler/include/profile_conv_bwd_data_impl.hpp deleted 100644 → 0
View file @ b548c0be
#pragma once
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "device_conv_bwd_data.hpp"
#include "element_wise_operation.hpp"
#include "reference_conv_bwd_data.hpp"
using F16 = ck::half_t;
using F32 = float;
using BF16 = ck::bhalf_t;
using INT8 = int8_t;
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_conv2d_bwd_data_instance {
using DeviceConvBwdDataNoOpPtr =
DeviceConvBwdDataPtr<ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough>;
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
void add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(
std::vector<DeviceConvBwdDataNoOpPtr>&);
} // namespace device_conv2d_bwd_data_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace ck {
namespace profiler {
template <int NDimSpatial,
typename InDataType,
typename WeiDataType,
typename OutDataType,
typename AccDataType,
typename InLayout,
typename WeiLayout,
typename OutLayout>
void profile_conv_bwd_data_impl(int do_verification,
int init_method,
bool do_log,
bool time_kernel,
ck::index_t N,
ck::index_t K,
ck::index_t C,
std::vector<ck::index_t> input_spatial_lengths,
std::vector<ck::index_t> filter_spatial_lengths,
std::vector<ck::index_t> output_spatial_lengths,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads)
{
const ck::index_t Y = filter_spatial_lengths[0];
const ck::index_t X = filter_spatial_lengths[1];
const ck::index_t Hi = input_spatial_lengths[0];
const ck::index_t Wi = input_spatial_lengths[1];
const ck::index_t Ho = output_spatial_lengths[0];
const ck::index_t Wo = output_spatial_lengths[1];
auto f_host_tensor_descriptor =
[](std::size_t N_, std::size_t C_, std::size_t H, std::size_t W, auto layout) {
if constexpr(is_same<decltype(layout), ck::tensor_layout::convolution::NCHW>::value ||
is_same<decltype(layout), ck::tensor_layout::convolution::KCYX>::value ||
is_same<decltype(layout), ck::tensor_layout::convolution::NKHW>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
std::vector<std::size_t>({C_ * H * W, H * W, W, 1}));
}
else if constexpr(is_same<decltype(layout), tensor_layout::convolution::NHWC>::value ||
is_same<decltype(layout), tensor_layout::convolution::KYXC>::value ||
is_same<decltype(layout), tensor_layout::convolution::NHWK>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
std::vector<std::size_t>({C_ * H * W, 1, W * C_, C_}));
}
};
Tensor<InDataType> in_n_c_hi_wi_host_result(f_host_tensor_descriptor(N, C, Hi, Wi, InLayout{}));
Tensor<InDataType> in_n_c_hi_wi_device_result(
f_host_tensor_descriptor(N, C, Hi, Wi, InLayout{}));
Tensor<WeiDataType> wei_k_c_y_x(f_host_tensor_descriptor(K, C, Y, X, WeiLayout{}));
Tensor<OutDataType> out_n_k_ho_wo(f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi_host_result.mDesc << std::endl;
std::cout << "wei_k_c_y_x: " << wei_k_c_y_x.mDesc << std::endl;
std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_2<WeiDataType>{-5, 5});
break;
default:
out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_3<InDataType>{0.0, 1.0});
wei_k_c_y_x.GenerateTensorValue(GeneratorTensor_3<WeiDataType>{-0.5, 0.5});
}
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
const auto in_element_op = InElementOp{};
const auto wei_element_op = WeiElementOp{};
const auto out_element_op = OutElementOp{};
if(do_verification)
{
using ReferenceConvBwdDataInstance =
ck::tensor_operation::host::ReferenceConvBwdData<InDataType,
WeiDataType,
OutDataType,
AccDataType,
InElementOp,
WeiElementOp,
OutElementOp>;
auto ref_conv = ReferenceConvBwdDataInstance{};
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi_host_result,
wei_k_c_y_x,
out_n_k_ho_wo,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op);
ref_invoker.Run(ref_argument);
}
DeviceMem in_device_buf(sizeof(InDataType) *
in_n_c_hi_wi_device_result.mDesc.GetElementSpace());
DeviceMem wei_device_buf(sizeof(WeiDataType) * wei_k_c_y_x.mDesc.GetElementSpace());
DeviceMem out_device_buf(sizeof(OutDataType) * out_n_k_ho_wo.mDesc.GetElementSpace());
out_device_buf.ToDevice(out_n_k_ho_wo.mData.data());
wei_device_buf.ToDevice(wei_k_c_y_x.mData.data());
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DeviceConvBwdDataNoOpPtr =
ck::tensor_operation::device::DeviceConvBwdDataPtr<PassThrough, PassThrough, PassThrough>;
// add device Conv instances
std::vector<DeviceConvBwdDataNoOpPtr> conv_ptrs;
if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, float> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, float> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, float>)
{
ck::tensor_operation::device::device_conv2d_bwd_data_instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f32_instances(conv_ptrs);
}
else if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, ck::half_t> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, ck::half_t> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, ck::half_t>)
{
ck::tensor_operation::device::device_conv2d_bwd_data_instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_f16_instances(conv_ptrs);
}
else if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, ck::bhalf_t> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, ck::bhalf_t> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, ck::bhalf_t>)
{
ck::tensor_operation::device::device_conv2d_bwd_data_instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_bf16_instances(conv_ptrs);
}
else if constexpr(ck::is_same_v<ck::remove_cv_t<InDataType>, int8_t> &&
ck::is_same_v<ck::remove_cv_t<WeiDataType>, int8_t> &&
ck::is_same_v<ck::remove_cv_t<OutDataType>, int8_t>)
{
ck::tensor_operation::device::device_conv2d_bwd_data_instance::
add_device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk_int8_instances(conv_ptrs);
}
if(conv_ptrs.size() <= 0)
{
throw std::runtime_error("wrong! no device Conv instance found");
}
std::string best_conv_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
// profile device Conv instances
for(auto& conv_ptr : conv_ptrs)
{
auto argument_ptr = conv_ptr->MakeArgumentPointer(
static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
N,
K,
C,
input_spatial_lengths,
filter_spatial_lengths,
output_spatial_lengths,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op);
auto invoker_ptr = conv_ptr->MakeInvokerPointer();
if(conv_ptr->IsSupportedArgument(argument_ptr.get()))
{
std::string conv_name = conv_ptr->GetTypeString();
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamControl{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
std::size_t num_btype = sizeof(InDataType) * (N * C * Hi * Wi) +
sizeof(WeiDataType) * (K * C * Y * X) +
sizeof(OutDataType) * (N * K * Ho * Wo);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec
<< " GB/s, " << conv_name << std::endl;
if(tflops > best_tflops)
{
best_conv_name = conv_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
in_device_buf.FromDevice(in_n_c_hi_wi_device_result.mData.data());
ck::utils::check_err(in_n_c_hi_wi_device_result.mData,
in_n_c_hi_wi_host_result.mData);
if(do_log)
{
LogRangeAsType<float>(std::cout << "in : ", out_n_k_ho_wo.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "wei: ", wei_k_c_y_x.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "out_host : ", in_n_c_hi_wi_host_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(
std::cout << "out_device: ", in_n_c_hi_wi_device_result.mData, ",")
<< std::endl;
}
}
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_conv_name << std::endl;
}
} // namespace profiler
} // namespace ck
profiler/include/profile_gemm_reduce_impl.hpp
View file @ 4769425e
......@@ -16,11 +16,21 @@ namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F32 = float;
using F16 = ck::half_t;
using DPtrsGlobal = ck::Tuple<F32*, F32*>;
using Identity = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Identity, Identity>;
using DeviceGemmReduceNoOpPtr = ck::tensor_operation::device::DeviceGemmReducePtr<
DPtrsGlobal,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::UnarySquare<float, float, false>>;
DInElementOps,
DOutElementOps>;
void add_device_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_mk_kn_mn_instances(
std::vector<DeviceGemmReduceNoOpPtr>&);
......@@ -112,19 +122,25 @@ bool profile_gemm_reduce_impl(int do_verification,
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}, num_thread);
}
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::reduce::Add<float>;
using D1ReduceOp = ck::reduce::Add<float>;
using D1ElementOp = ck::tensor_operation::element_wise::UnarySquare<float, float, false>;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
const auto d0_reduce_op = D0ReduceOp{};
const auto d1_reduce_op = D1ReduceOp{};
const auto d1_element_op = D1ElementOp{};
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::reduce::Add<float>;
using D1ReduceOp = ck::reduce::Add<float>;
using UnaryIdenticElementOp =
ck::tensor_operation::element_wise::UnaryIdentic<float, float, false>;
using UnarySquareElementOp =
ck::tensor_operation::element_wise::UnarySquare<float, float, false>;
using DxsInElementOps = ck::Tuple<UnaryIdenticElementOp, UnarySquareElementOp>;
using DxsOutElementOps = ck::Tuple<UnaryIdenticElementOp, UnaryIdenticElementOp>;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto c_element_op = CElementOp{};
const auto dxs_in_element_op = DxsInElementOps{};
const auto dxs_out_element_op = DxsOutElementOps{};
const auto d0_reduce_op = D0ReduceOp{};
const auto d1_reduce_op = D1ReduceOp{};
if(do_verification)
{
......@@ -149,7 +165,7 @@ bool profile_gemm_reduce_impl(int do_verification,
float d0_val = ck::type_convert<float>(c_m_n_host_result(m, n));
float d1_val;
d1_element_op(d1_val, d0_val);
UnarySquareElementOp{}(d1_val, d0_val);
d0_reduce_op(d0_acc, d0_val);
d1_reduce_op(d1_acc, d1_val);
}
......@@ -165,6 +181,9 @@ bool profile_gemm_reduce_impl(int do_verification,
DeviceMem d0_device_buf(sizeof(DDataType) * d0_m_device_result.mDesc.GetElementSpace());
DeviceMem d1_device_buf(sizeof(DDataType) * d1_m_device_result.mDesc.GetElementSpace());
auto dxs_global = ck::make_tuple(static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()));
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
......@@ -226,8 +245,7 @@ bool profile_gemm_reduce_impl(int do_verification,
gemm_ptr->MakeArgumentPointer(static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d0_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d1_device_buf.GetDeviceBuffer()),
dxs_global,
M,
N,
K,
......@@ -237,7 +255,8 @@ bool profile_gemm_reduce_impl(int do_verification,
a_element_op,
b_element_op,
c_element_op,
d1_element_op);
dxs_in_element_op,
dxs_out_element_op);
auto invoker_ptr = gemm_ptr->MakeInvokerPointer();
......
profiler/src/profile_conv_bwd_data.cpp deleted 100644 → 0
View file @ b548c0be
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "profile_conv_bwd_data_impl.hpp"
enum struct ConvDataType
{
F32_F32_F32, // 0
F16_F16_F16, // 1
BF16_BF16_BF16, // 2
INT8_INT8_INT8, // 3
};
enum struct ConvInputLayout
{
NCHW, // 0
NHWC, // 1
};
enum struct ConvWeightLayout
{
KCYX, // 0
KYXC, // 1
};
enum struct ConvOutputLayout
{
NKHW, // 0
NHWK, // 1
};
int profile_conv_bwd_data(int argc, char* argv[])
{
if(argc != 25)
{
printf("arg1: tensor operation (conv_bwd: BackwardConvolution)\n");
printf("arg2: data type (0: fp32; 1: fp16)\n");
printf("arg3: input tensor layout (0: NCHW; 1: NHWC)\n");
printf("arg4: weight tensor layout (0: KCYX; 1: KYXC)\n");
printf("arg5: output tensor layout (0: NKHW; 1: NHWK)\n");
printf("arg6: verification (0: no; 1: yes)\n");
printf("arg7: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg8: print tensor value (0: no; 1: yes)\n");
printf("arg9: time kernel (0=n0, 1=yes)\n");
printf("arg10 to 24: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
"RightPx\n");
exit(1);
}
const auto data_type = static_cast<ConvDataType>(std::stoi(argv[2]));
const auto in_layout = static_cast<ConvInputLayout>(std::stoi(argv[3]));
const auto wei_layout = static_cast<ConvWeightLayout>(std::stoi(argv[4]));
const auto out_layout = static_cast<ConvOutputLayout>(std::stoi(argv[5]));
const bool do_verification = std::stoi(argv[6]);
const int init_method = std::stoi(argv[7]);
const bool do_log = std::stoi(argv[8]);
const bool time_kernel = std::stoi(argv[9]);
const ck::index_t N = std::stoi(argv[10]);
const ck::index_t K = std::stoi(argv[11]);
const ck::index_t C = std::stoi(argv[12]);
const ck::index_t Y = std::stoi(argv[13]);
const ck::index_t X = std::stoi(argv[14]);
const ck::index_t Hi = std::stoi(argv[15]);
const ck::index_t Wi = std::stoi(argv[16]);
const ck::index_t conv_stride_h = std::stoi(argv[17]);
const ck::index_t conv_stride_w = std::stoi(argv[18]);
const ck::index_t conv_dilation_h = std::stoi(argv[19]);
const ck::index_t conv_dilation_w = std::stoi(argv[20]);
const ck::index_t in_left_pad_h = std::stoi(argv[21]);
const ck::index_t in_left_pad_w = std::stoi(argv[22]);
const ck::index_t in_right_pad_h = std::stoi(argv[23]);
const ck::index_t in_right_pad_w = std::stoi(argv[24]);
const ck::index_t YEff = (Y - 1) * conv_dilation_h + 1;
const ck::index_t XEff = (X - 1) * conv_dilation_w + 1;
const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - YEff) / conv_stride_h + 1;
const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
if(data_type == ConvDataType::F32_F32_F32 && in_layout == ConvInputLayout::NHWC &&
wei_layout == ConvWeightLayout::KYXC && out_layout == ConvOutputLayout::NHWK)
{
ck::profiler::profile_conv_bwd_data_impl<2,
float,
float,
float,
float,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
do_verification,
init_method,
do_log,
StreamControl{nullptr, time_kernel},
N,
K,
C,
std::vector<ck::index_t>{Hi, Wi},
std::vector<ck::index_t>{Y, X},
std::vector<ck::index_t>{Ho, Wo},
std::vector<ck::index_t>{conv_stride_h, conv_stride_w},
std::vector<ck::index_t>{conv_dilation_h, conv_dilation_w},
std::vector<ck::index_t>{in_left_pad_h, in_left_pad_w},
std::vector<ck::index_t>{in_right_pad_h, in_right_pad_w});
}
else if(data_type == ConvDataType::F16_F16_F16 && in_layout == ConvInputLayout::NHWC &&
wei_layout == ConvWeightLayout::KYXC && out_layout == ConvOutputLayout::NHWK)
{
ck::profiler::profile_conv_bwd_data_impl<2,
ck::half_t,
ck::half_t,
ck::half_t,
float,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
do_verification,
init_method,
do_log,
StreamControl{nullptr, time_kernel},
N,
K,
C,
std::vector<ck::index_t>{Hi, Wi},
std::vector<ck::index_t>{Y, X},
std::vector<ck::index_t>{Ho, Wo},
std::vector<ck::index_t>{conv_stride_h, conv_stride_w},
std::vector<ck::index_t>{conv_dilation_h, conv_dilation_w},
std::vector<ck::index_t>{in_left_pad_h, in_left_pad_w},
std::vector<ck::index_t>{in_right_pad_h, in_right_pad_w});
}
else if(data_type == ConvDataType::BF16_BF16_BF16 && in_layout == ConvInputLayout::NHWC &&
wei_layout == ConvWeightLayout::KYXC && out_layout == ConvOutputLayout::NHWK)
{
ck::profiler::profile_conv_bwd_data_impl<2,
uint16_t,
uint16_t,
uint16_t,
float,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
do_verification,
init_method,
do_log,
StreamControl{nullptr, time_kernel},
N,
K,
C,
std::vector<ck::index_t>{Hi, Wi},
std::vector<ck::index_t>{Y, X},
std::vector<ck::index_t>{Ho, Wo},
std::vector<ck::index_t>{conv_stride_h, conv_stride_w},
std::vector<ck::index_t>{conv_dilation_h, conv_dilation_w},
std::vector<ck::index_t>{in_left_pad_h, in_left_pad_w},
std::vector<ck::index_t>{in_right_pad_h, in_right_pad_w});
}
else if(data_type == ConvDataType::INT8_INT8_INT8 && in_layout == ConvInputLayout::NHWC &&
wei_layout == ConvWeightLayout::KYXC && out_layout == ConvOutputLayout::NHWK)
{
ck::profiler::profile_conv_bwd_data_impl<2,
int8_t,
int8_t,
int8_t,
int32_t,
ck::tensor_layout::convolution::NHWC,
ck::tensor_layout::convolution::KYXC,
ck::tensor_layout::convolution::NHWK>(
do_verification,
init_method,
do_log,
StreamControl{nullptr, time_kernel},
N,
K,
C,
std::vector<ck::index_t>{Hi, Wi},
std::vector<ck::index_t>{Y, X},
std::vector<ck::index_t>{Ho, Wo},
std::vector<ck::index_t>{conv_stride_h, conv_stride_w},
std::vector<ck::index_t>{conv_dilation_h, conv_dilation_w},
std::vector<ck::index_t>{in_left_pad_h, in_left_pad_w},
std::vector<ck::index_t>{in_right_pad_h, in_right_pad_w});
}
else
{
throw std::runtime_error("wrong! this Conv data_type & layout is not implemented");
}
return 1;
}
test/CMakeLists.txt
View file @ 4769425e
......@@ -63,4 +63,5 @@ add_subdirectory(convnd_fwd)
add_subdirectory(reduce)
add_subdirectory(conv2d_bwd_weight)
add_subdirectory(convnd_bwd_data)
# DONOT add client_app, that is tested via CI independently
\ No newline at end of file
add_subdirectory(block_to_ctile_map)
# DONOT add client_app, that is tested via CI independently
test/block_to_ctile_map/CMakeLists.txt 0 → 100644
View file @ 4769425e
add_gtest_executable(test_block_to_ctile_map test_block_to_ctile_map.cpp)
\ No newline at end of file
test/block_to_ctile_map/test_block_to_ctile_map.cpp 0 → 100644
View file @ 4769425e
#include <ck/config.hpp>
#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
#include "gtest/gtest.h"
#include <iostream>
#include <vector>
using namespace ck;
static auto I0 = Number<0>{};
static auto I1 = Number<1>{};
TEST(BlockToCTileMap, TestBlockToCTileMap_M00_N00_M01_N01_DeviceCTileIndexCheck1)
{
const index_t M = 384;
const index_t N = 384;
const index_t MPerBlock = 128;
const index_t NPerBlock = 128;
const index_t MBlock = M / MPerBlock;
const index_t NBlock = N / NPerBlock;
const index_t M01 = 4;
const index_t N01 = 4;
auto c_grid_desc_m_n = make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, I1));
printf("(M, N, MPerBlock, NPerBlock, M01, N01) = (%d, %d, %d, %d, %d, %d)\n",
M,
N,
MPerBlock,
NPerBlock,
M01,
N01);
BlockToCTileMap_M00_N00_M01_N01<MPerBlock, NPerBlock, decltype(c_grid_desc_m_n), true> tile_map(
c_grid_desc_m_n, M01, N01);
EXPECT_TRUE(tile_map.CheckValidity(c_grid_desc_m_n) == true);
EXPECT_TRUE(tile_map.CalculateGridSize(c_grid_desc_m_n) == 16);
// clang-format off
std::vector<std::vector<int>> expected = {
{0, 0, 1},
{0, 1, 1},
{0, 2, 1},
{0, 3, 0},
{1, 0, 1},
{1, 1, 1},
{1, 2, 1},
{1, 3, 0},
{2, 0, 1},
{2, 1, 1},
{2, 2, 1},
{2, 3, 0},
{3, 0, 0},
{3, 1, 0},
{3, 2, 0},
{3, 3, 0}
};
// clang-format on
for(index_t i = 0; i < tile_map.CalculateGridSize(c_grid_desc_m_n); i++)
{
auto m0n0_idx = tile_map.CalculateBottomIndex(make_multi_index(i));
std::cout << "block_1d_id = " << i << ", m0, n0 = " << m0n0_idx[I0] << ", " << m0n0_idx[I1];
std::cout << ", valid = " << tile_map.ValidCTileIndex(m0n0_idx, make_tuple(MBlock, NBlock))
<< std::endl;
bool equal =
expected[i] ==
std::vector<int>{m0n0_idx[I0],
m0n0_idx[I1],
tile_map.ValidCTileIndex(m0n0_idx, make_tuple(MBlock, NBlock))};
EXPECT_TRUE(equal);
}
}
TEST(BlockToCTileMap, TestBlockToCTileMap_M00_N00_M01_N01_DeviceCTileIndexCheck0)
{
const index_t M = 384;
const index_t N = 384;
const index_t MPerBlock = 128;
const index_t NPerBlock = 128;
// const index_t MBlock = M / MPerBlock;
// const index_t NBlock = N / NPerBlock;
const index_t M01 = 4;
const index_t N01 = 4;
auto c_grid_desc_m_n = make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, I1));
printf("(M, N, MPerBlock, NPerBlock, M01, N01) = (%d, %d, %d, %d, %d, %d)\n",
M,
N,
MPerBlock,
NPerBlock,
M01,
N01);
BlockToCTileMap_M00_N00_M01_N01<MPerBlock, NPerBlock, decltype(c_grid_desc_m_n), false>
tile_map(c_grid_desc_m_n, M01, N01);
EXPECT_TRUE(tile_map.CheckValidity(c_grid_desc_m_n) == false);
}
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment