Skip to content

GitLab

Commit 32203c74 authored by danyao12's avatar danyao12
Browse files

bwd hd256

parent 339b86e9
Expand all Show whitespace changes
Inline Side-by-side
Showing with 4316 additions and 1 deletion
example/32_batched_gemm_scale_softmax_gemm/batched_multihead_attention_backward_v2.cpp
View file @ 32203c74
...@@ -25,7 +25,7 @@ Kernel outputs: ...@@ -25,7 +25,7 @@ Kernel outputs:
#define PRINT_HOST 0 #define PRINT_HOST 0
#define USING_MASK 0 #define USING_MASK 0
#define DIM 128 // DIM should be a multiple of 8. #define DIM 256 // DIM should be a multiple of 8.
#include <iostream> #include <iostream>
#include <numeric> #include <numeric>
...@@ -38,6 +38,7 @@ Kernel outputs: ...@@ -38,6 +38,7 @@ Kernel outputs:
#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp" #include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_batched_mha_bwd_xdl_cshuffle_qloop_v1.hpp" #include "ck/tensor_operation/gpu/device/impl/device_batched_mha_bwd_xdl_cshuffle_qloop_v1.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_batched_mha_bwd_xdl_cshuffle_qloop_v2.hpp" #include "ck/tensor_operation/gpu/device/impl/device_batched_mha_bwd_xdl_cshuffle_qloop_v2.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_batched_mha_bwd_xdl_cshuffle_qloop_v3.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
...@@ -150,6 +151,19 @@ using DeviceGemmInstance = ...@@ -150,6 +151,19 @@ using DeviceGemmInstance =
// ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, InputDataType, OutputDataType, GemmDataType, ZDataType, LSEDataType, Acc0BiasDataType, Acc1BiasDataType, AccDataType, ShuffleDataType, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 32, 128, 128, 128, 32, 64, 8, 8, 2, 32, 32, 1, 1, 4, 1, S<8, 32, 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, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, Deterministic>; // ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, InputDataType, OutputDataType, GemmDataType, ZDataType, LSEDataType, Acc0BiasDataType, Acc1BiasDataType, AccDataType, ShuffleDataType, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 32, 128, 128, 128, 32, 64, 8, 8, 2, 32, 32, 1, 1, 4, 1, S<8, 32, 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, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, Deterministic>;
// ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, InputDataType, OutputDataType, GemmDataType, ZDataType, LSEDataType, Acc0BiasDataType, Acc1BiasDataType, AccDataType, ShuffleDataType, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 32, 128, 64, 128, 32, 64, 8, 8, 2, 32, 32, 1, 1, 4, 1, S<8, 32, 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, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, Deterministic>; // ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, InputDataType, OutputDataType, GemmDataType, ZDataType, LSEDataType, Acc0BiasDataType, Acc1BiasDataType, AccDataType, ShuffleDataType, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 32, 128, 64, 128, 32, 64, 8, 8, 2, 32, 32, 1, 1, 4, 1, S<8, 32, 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, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, Deterministic>;
// clang-format on // clang-format on
#elif(DIM <= 256)
// clang-format off
using DeviceGemmInstance =
// ########################################################################################| NumDimG| NumDimM| NumDimN| NumDimK| NumDimO| InputDataType| OutputDataType| GemmDataType| ZDataType| LSEDataType| Acc0BiasDataType| Acc1BiasDataType| GemmAcc| CShuffle| A| B| Acc| B1| C| GEMM| ATensorSpec| B0TensorSpec| B1TensorSpec| CTensorSpec| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| Gemm2| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| Gemm2| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| D0BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CShuffleBlockTransferScalarPerVector_NPerBlock| MaskingSpec| Deterministic|
// ########################################################################################| | | | | | | | | | | | | DataType| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| | | | | Prefetch| Size| MPer| NPer| KPer| NPer| KPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcScalar| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| | | |
// ########################################################################################| | | | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | | | | | Stage| | Block| 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| | PerVector| Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| | | |
// ########################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | | | | |
// ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, InputDataType, OutputDataType, GemmDataType, ZDataType, LSEDataType, Acc0BiasDataType, Acc1BiasDataType, AccDataType, ShuffleDataType, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 32, 128, 128, 128, 32, 64, 8, 8, 2, 32, 32, 1, 1, 4, 1, S<8, 32, 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, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, Deterministic>;
// ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, InputDataType, OutputDataType, GemmDataType, ZDataType, LSEDataType, Acc0BiasDataType, Acc1BiasDataType, AccDataType, ShuffleDataType, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 32, 128, 64, 128, 32, 64, 8, 8, 2, 32, 32, 1, 1, 4, 1, S<8, 32, 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, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, Deterministic>;
// ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V3< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, InputDataType, OutputDataType, GemmDataType, ZDataType, LSEDataType, Acc0BiasDataType, Acc1BiasDataType, AccDataType, ShuffleDataType, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 32, 128, 32, 256, 32, 64, 8, 8, 2, 32, 32, 1, 1, 8, 2, S<8, 32, 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, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, Deterministic>;
ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V3< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, InputDataType, OutputDataType, GemmDataType, ZDataType, LSEDataType, Acc0BiasDataType, Acc1BiasDataType, AccDataType, ShuffleDataType, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 32, 128, 32, 256, 32, 64, 8, 8, 2, 32, 32, 1, 1, 8, 2, S<8, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 4, S<1, 32, 1, 8>, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, Deterministic>;
// clang-format on
#endif #endif
// Ref Gemm0: S = alpha * Q * K^T // Ref Gemm0: S = alpha * Q * K^T
......
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