Skip to content

GitLab

Commit 72498367 authored by guangzlu's avatar guangzlu
Browse files

fixed bugs

parent d1280ed5
Hide whitespace changes
Inline Side-by-side
Showing with 13 additions and 13 deletions
library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_mha_bwd_qloop_bf16_bf16_gmk_gnk_gno_gmo_instance.cpp
View file @ 72498367
......@@ -32,10 +32,10 @@ using YElementOp = PassThrough;
using Acc0BiasDataType = ck::Tuple<>;
using Acc1BiasDataType = ck::Tuple<>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
//static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
static constexpr auto TensorDefault = ck::tensor_operation::device::TensorSpecialization::Default;
//static constexpr auto TensorDefault = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecQ = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecK = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecV = ck::tensor_operation::device::TensorSpecialization::Default;
......@@ -57,10 +57,10 @@ using device_batched_mha_bwd_qloop_bf16_bf16_gmk_gnk_gno_gmo_instances =
// ########################################################################################| | | | | | | | | | | | | DataType| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| | | | | Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| | | |
// ########################################################################################| | | | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | | | | | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| | | |
// ########################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | |
ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, unsigned short, F32, Acc0BiasDataType, Acc1BiasDataType, F32, F32, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 64, 128, 64, 64, 32, 8, 8, 2, 32, 32, 2, 1, 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, 2, S<1, 32, 1, 8>, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, Deterministic>,
ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, unsigned short, F32, Acc0BiasDataType, Acc1BiasDataType, F32, F32, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 128, 128, 64, 64, 32, 8, 8, 2, 32, 32, 4, 1, 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, 2, S<1, 32, 1, 8>, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, Deterministic>,
ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, unsigned short, F32, Acc0BiasDataType, Acc1BiasDataType, F32, F32, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 128, 128, 32, 64, 32, 8, 8, 2, 32, 32, 4, 1, 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, 2, S<1, 32, 1, 8>, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, Deterministic>,
ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, unsigned short, F32, Acc0BiasDataType, Acc1BiasDataType, F32, F32, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 128, 128, 64, 64, 32, 8, 8, 2, 32, 32, 4, 1, 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, 2, S<1, 32, 1, 8>, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, Deterministic>
ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, unsigned short, F32, Acc0BiasDataType, Acc1BiasDataType, F32, F32, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 64, 128, 64, 64, 32, 8, 8, 2, 32, 32, 2, 1, 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, 2, S<1, 32, 1, 8>, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, Deterministic>
// ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, unsigned short, F32, Acc0BiasDataType, Acc1BiasDataType, F32, F32, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 128, 128, 64, 64, 32, 8, 8, 2, 32, 32, 4, 1, 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, 2, S<1, 32, 1, 8>, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, Deterministic>,
// ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, unsigned short, F32, Acc0BiasDataType, Acc1BiasDataType, F32, F32, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 128, 128, 32, 64, 32, 8, 8, 2, 32, 32, 4, 1, 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, 2, S<1, 32, 1, 8>, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, Deterministic>,
// ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, BF16, BF16, BF16, unsigned short, F32, Acc0BiasDataType, Acc1BiasDataType, F32, F32, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 128, 128, 64, 64, 32, 8, 8, 2, 32, 32, 4, 1, 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, 2, S<1, 32, 1, 8>, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, Deterministic>
// clang-format on
>;
......
library/src/tensor_operation_instance/gpu/batched_gemm_softmax_gemm_permute/device_batched_mha_bwd_qloop_f16_f16_gmk_gnk_gno_gmo_instance.cpp
View file @ 72498367
......@@ -32,10 +32,10 @@ using YElementOp = PassThrough;
using Acc0BiasDataType = ck::Tuple<>;
using Acc1BiasDataType = ck::Tuple<>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
//static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
static constexpr auto TensorDefault = ck::tensor_operation::device::TensorSpecialization::Default;
//static constexpr auto TensorDefault = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecQ = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecK = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecV = ck::tensor_operation::device::TensorSpecialization::Default;
......@@ -57,16 +57,16 @@ using device_batched_mha_bwd_qloop_f16_f16_gmk_gnk_gno_gmo_instances =
// ########################################################################################| | | | | | | | | | | | | DataType| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| | | | | Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| | | |
// ########################################################################################| | | | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | | | | | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| | | |
// ########################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | |
ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, unsigned short, F32, Acc0BiasDataType, Acc1BiasDataType, F32, F32, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 64, 128, 64, 64, 32, 8, 8, 2, 32, 32, 2, 1, 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, 2, S<1, 32, 1, 8>, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, Deterministic>,
ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, unsigned short, F32, Acc0BiasDataType, Acc1BiasDataType, F32, F32, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 128, 128, 64, 64, 32, 8, 8, 2, 32, 32, 4, 1, 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, 2, S<1, 32, 1, 8>, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, Deterministic>,
ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, unsigned short, F32, Acc0BiasDataType, Acc1BiasDataType, F32, F32, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 128, 128, 32, 64, 32, 8, 8, 2, 32, 32, 4, 1, 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, 2, S<1, 32, 1, 8>, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, Deterministic>,
ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, unsigned short, F32, Acc0BiasDataType, Acc1BiasDataType, F32, F32, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 128, 128, 64, 64, 32, 8, 8, 2, 32, 32, 4, 1, 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, 2, S<1, 32, 1, 8>, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, Deterministic>
ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, unsigned short, F32, Acc0BiasDataType, Acc1BiasDataType, F32, F32, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 64, 128, 64, 64, 32, 8, 8, 2, 32, 32, 2, 1, 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, 2, S<1, 32, 1, 8>, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, Deterministic>
// ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, unsigned short, F32, Acc0BiasDataType, Acc1BiasDataType, F32, F32, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 128, 128, 64, 64, 32, 8, 8, 2, 32, 32, 4, 1, 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, 2, S<1, 32, 1, 8>, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, Deterministic>,
// ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, unsigned short, F32, Acc0BiasDataType, Acc1BiasDataType, F32, F32, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 128, 128, 32, 64, 32, 8, 8, 2, 32, 32, 4, 1, 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, 2, S<1, 32, 1, 8>, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, Deterministic>,
// ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, F16, F16, F16, unsigned short, F32, Acc0BiasDataType, Acc1BiasDataType, F32, F32, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 128, 128, 64, 64, 32, 8, 8, 2, 32, 32, 4, 1, 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, 2, S<1, 32, 1, 8>, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, Deterministic>
// clang-format on
>;
void add_device_batched_mha_bwd_qloop_casual_f16_f16_gmk_gnk_gno_gmo_instances(
std::vector<std::unique_ptr<
DeviceBatchedGemmSoftmaxGemmPermute<2,
DeviceBatchedMultiheadAttentionBackward<2,
1,
1,
1,
......
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