Skip to content

GitLab

Commit f752739c authored by danyao12's avatar danyao12
Browse files

Merge branch 'mha-train-develop' into mha-train-ldsbypass

parents b3a96764 26fa4782
Show whitespace changes
Inline Side-by-side
Showing with 1864 additions and 216 deletions
example/32_batched_gemm_scale_softmax_gemm/batched_multihead_attention_forward_v2.cpp
View file @ f752739c
...@@ -52,8 +52,8 @@ using CShuffleDataType = F32; ...@@ -52,8 +52,8 @@ using CShuffleDataType = F32;
using CDataType = DataType; using CDataType = DataType;
using ZDataType = U16; // INT32 using ZDataType = U16; // INT32
using LSEDataType = F32; using LSEDataType = F32;
using Acc0BiasDataType = ck::Tuple<>; using Acc0BiasDataType = void;
using Acc1BiasDataType = ck::Tuple<>; using Acc1BiasDataType = void;
static constexpr ck::index_t NumDimG = 2; static constexpr ck::index_t NumDimG = 2;
static constexpr ck::index_t NumDimM = 1; static constexpr ck::index_t NumDimM = 1;
...@@ -121,6 +121,7 @@ using DeviceGemmInstance = ...@@ -121,6 +121,7 @@ using DeviceGemmInstance =
1, // MXdlPerWave 1, // MXdlPerWave
4, // NXdlPerWave 4, // NXdlPerWave
1, // Gemm1NXdlPerWave 1, // Gemm1NXdlPerWave
1, // DropoutStep
S<4, 64, 1>, // ABlockTransfer S<4, 64, 1>, // ABlockTransfer
S<1, 0, 2>, S<1, 0, 2>,
S<1, 0, 2>, S<1, 0, 2>,
...@@ -135,6 +136,7 @@ using DeviceGemmInstance = ...@@ -135,6 +136,7 @@ using DeviceGemmInstance =
8, 8,
8, 8,
true, true,
4,
S<16, 16, 1>, // B1BlockTransfer S<16, 16, 1>, // B1BlockTransfer
S<0, 2, 1>, S<0, 2, 1>,
S<0, 2, 1>, S<0, 2, 1>,
...@@ -146,6 +148,7 @@ using DeviceGemmInstance = ...@@ -146,6 +148,7 @@ using DeviceGemmInstance =
1, // CShuffleNXdlPerWavePerShuffle 1, // CShuffleNXdlPerWavePerShuffle
S<1, 64, 1, 4>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock S<1, 64, 1, 4>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock 8, // CShuffleBlockTransferScalarPerVector_NPerBlock
4,
MaskingSpec, // MaskingSpecialization MaskingSpec, // MaskingSpecialization
Deterministic>; Deterministic>;
#elif(DIM <= 64) #elif(DIM <= 64)
...@@ -192,6 +195,7 @@ using DeviceGemmInstance = ...@@ -192,6 +195,7 @@ using DeviceGemmInstance =
1, // MXdlPerWave 1, // MXdlPerWave
4, // NXdlPerWave 4, // NXdlPerWave
2, // Gemm1NXdlPerWave 2, // Gemm1NXdlPerWave
1, // DropoutStep
S<4, 64, 1>, // ABlockTransfer S<4, 64, 1>, // ABlockTransfer
S<1, 0, 2>, S<1, 0, 2>,
S<1, 0, 2>, S<1, 0, 2>,
...@@ -206,6 +210,7 @@ using DeviceGemmInstance = ...@@ -206,6 +210,7 @@ using DeviceGemmInstance =
8, 8,
8, 8,
true, true,
4,
S<16, 16, 1>, // B1BlockTransfer S<16, 16, 1>, // B1BlockTransfer
S<0, 2, 1>, S<0, 2, 1>,
S<0, 2, 1>, S<0, 2, 1>,
...@@ -217,6 +222,7 @@ using DeviceGemmInstance = ...@@ -217,6 +222,7 @@ using DeviceGemmInstance =
2, // CShuffleNXdlPerWavePerShuffle 2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock 8, // CShuffleBlockTransferScalarPerVector_NPerBlock
4,
MaskingSpec, // MaskingSpecialization MaskingSpec, // MaskingSpecialization
Deterministic>; Deterministic>;
#elif(DIM <= 128) #elif(DIM <= 128)
...@@ -253,7 +259,7 @@ using DeviceGemmInstance = ...@@ -253,7 +259,7 @@ using DeviceGemmInstance =
128, // MPerBlock 128, // MPerBlock
128, // NPerBlock 128, // NPerBlock
32, // KPerBlock 32, // KPerBlock
128, // Gemm1NPerBlock 64, // Gemm1NPerBlock
32, // Gemm1KPerBlock 32, // Gemm1KPerBlock
8, // AK1 8, // AK1
8, // BK1 8, // BK1
...@@ -262,7 +268,8 @@ using DeviceGemmInstance = ...@@ -262,7 +268,8 @@ using DeviceGemmInstance =
32, // NPerXDL 32, // NPerXDL
1, // MXdlPerWave 1, // MXdlPerWave
4, // NXdlPerWave 4, // NXdlPerWave
4, // Gemm1NXdlPerWave 2, // Gemm1NXdlPerWave
1, // DropoutStep
S<4, 64, 1>, // ABlockTransfer S<4, 64, 1>, // ABlockTransfer
S<1, 0, 2>, S<1, 0, 2>,
S<1, 0, 2>, S<1, 0, 2>,
...@@ -277,7 +284,8 @@ using DeviceGemmInstance = ...@@ -277,7 +284,8 @@ using DeviceGemmInstance =
8, 8,
8, 8,
true, true,
S<8, 32, 1>, // B1BlockTransfer 4,
S<16, 16, 1>, // B1BlockTransfer
S<0, 2, 1>, S<0, 2, 1>,
S<0, 2, 1>, S<0, 2, 1>,
1, 1,
...@@ -288,6 +296,7 @@ using DeviceGemmInstance = ...@@ -288,6 +296,7 @@ using DeviceGemmInstance =
2, // CShuffleNXdlPerWavePerShuffle 2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock 8, // CShuffleBlockTransferScalarPerVector_NPerBlock
4,
MaskingSpec, // MaskingSpecialization MaskingSpec, // MaskingSpecialization
Deterministic>; Deterministic>;
#endif #endif
......
example/32_batched_gemm_scale_softmax_gemm/batched_multihead_attention_train_v2.cpp
View file @ f752739c
...@@ -79,8 +79,8 @@ using AccDataType = F32; ...@@ -79,8 +79,8 @@ using AccDataType = F32;
using ShuffleDataType = F32; using ShuffleDataType = F32;
using LSEDataType = F32; using LSEDataType = F32;
using ZDataType = U16; // INT32 using ZDataType = U16; // INT32
using Acc0BiasDataType = ck::Tuple<>; using Acc0BiasDataType = void;
using Acc1BiasDataType = ck::Tuple<>; using Acc1BiasDataType = void;
static constexpr ck::index_t NumDimG = 2; static constexpr ck::index_t NumDimG = 2;
static constexpr ck::index_t NumDimM = 1; static constexpr ck::index_t NumDimM = 1;
...@@ -113,11 +113,11 @@ static constexpr bool Deterministic = false; ...@@ -113,11 +113,11 @@ static constexpr bool Deterministic = false;
#if(DIM <= 32) #if(DIM <= 32)
// clang-format off // clang-format off
using DeviceGemmInstanceFWD = using DeviceGemmInstanceFWD =
// #################################################################################| NumDimG| NumDimM| NumDimN| NumDimK| NumDimO| ADataType| BDataType| B1DataType| CDataType| GemmDataType| ZDataType| LSEDataType| Acc0BiasDataType| Acc1BiasDataType| GemmAcc| CShuffle| A| B| Acc| B1| C| GEMM| ATensorSpec| B0TensorSpec| B1TensorSpec| CTensorSpec| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CShuffleBlockTransferScalarPerVector| MaskingSpec| Deterministic| // #################################################################################| NumDimG| NumDimM| NumDimN| NumDimK| NumDimO| ADataType| BDataType| B1DataType| CDataType| GemmDataType| ZDataType| LSEDataType| Acc0BiasDataType| Acc1BiasDataType| GemmAcc| CShuffle| A| B| Acc| B1| C| GEMM| ATensorSpec| B0TensorSpec| B1TensorSpec| CTensorSpec| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1|Dropout| 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| D1BlockTransfer| MaskingSpec| Deterministic|
// #################################################################################| | | | | | | | | | | | | | | DataType| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| | | | | Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| _NPerBlock| | | // #################################################################################| | | | | | | | | | | | | | | DataType| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| | | | | Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| Step| 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| _NPerBlock| SrcScalar| | |
// #################################################################################| | | | | | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | | | | | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| | | | // #################################################################################| | | | | | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | | | | | Stage| | Block| Block| Block| Block| Block| | | | | | 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| | PerVector| | |
// #################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | | | | // #################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
ck::tensor_operation::device::DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, InputDataType, InputDataType, InputDataType, InputDataType, GemmDataType, ZDataType, LSEDataType, Acc0BiasDataType, Acc1BiasDataType, AccDataType, ShuffleDataType, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 128, 128, 32, 32, 32, 8, 8, 2, 32, 32, 1, 4, 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, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, false, 1, 1, S<1, 64, 1, 4>, 8, MaskingSpec, Deterministic>; ck::tensor_operation::device::DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, InputDataType, InputDataType, InputDataType, InputDataType, GemmDataType, ZDataType, LSEDataType, Acc0BiasDataType, Acc1BiasDataType, AccDataType, ShuffleDataType, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 128, 128, 32, 32, 32, 8, 8, 2, 32, 32, 1, 4, 1, 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, 4, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, false, 1, 1, S<1, 64, 1, 4>, 8, 4, MaskingSpec, Deterministic>;
using DeviceGemmInstanceBWD = using DeviceGemmInstanceBWD =
// ########################################################################################| 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| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| Gemm2| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CShuffleBlockTransferScalarPerVector_NPerBlock| MaskingSpec| Deterministic| // ########################################################################################| 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| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| Gemm2| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CShuffleBlockTransferScalarPerVector_NPerBlock| MaskingSpec| Deterministic|
...@@ -129,11 +129,11 @@ using DeviceGemmInstanceBWD = ...@@ -129,11 +129,11 @@ using DeviceGemmInstanceBWD =
#elif(DIM <= 64) #elif(DIM <= 64)
// clang-format off // clang-format off
using DeviceGemmInstanceFWD = using DeviceGemmInstanceFWD =
// #################################################################################| NumDimG| NumDimM| NumDimN| NumDimK| NumDimO| ADataType| BDataType| B1DataType| CDataType| GemmDataType| ZDataType| LSEDataType| Acc0BiasDataType| Acc1BiasDataType| GemmAcc| CShuffle| A| B| Acc| B1| C| GEMM| ATensorSpec| B0TensorSpec| B1TensorSpec| CTensorSpec| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CShuffleBlockTransferScalarPerVector| MaskingSpec| Deterministic| // #################################################################################| NumDimG| NumDimM| NumDimN| NumDimK| NumDimO| ADataType| BDataType| B1DataType| CDataType| GemmDataType| ZDataType| LSEDataType| Acc0BiasDataType| Acc1BiasDataType| GemmAcc| CShuffle| A| B| Acc| B1| C| GEMM| ATensorSpec| B0TensorSpec| B1TensorSpec| CTensorSpec| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1|Dropout| 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| D1BlockTransfer| MaskingSpec| Deterministic|
// #################################################################################| | | | | | | | | | | | | | | DataType| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| | | | | Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| _NPerBlock| | | // #################################################################################| | | | | | | | | | | | | | | DataType| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| | | | | Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| Step| 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| _NPerBlock| SrcScalar| | |
// #################################################################################| | | | | | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | | | | | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| | | | // #################################################################################| | | | | | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | | | | | Stage| | Block| Block| Block| Block| Block| | | | | | 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| | PerVector| | |
// #################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | | | | // #################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
ck::tensor_operation::device::DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, InputDataType, InputDataType, InputDataType, InputDataType, GemmDataType, ZDataType, LSEDataType, Acc0BiasDataType, Acc1BiasDataType, AccDataType, ShuffleDataType, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 128, 128, 32, 64, 32, 8, 8, 2, 32, 32, 1, 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, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec, Deterministic>; ck::tensor_operation::device::DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, InputDataType, InputDataType, InputDataType, InputDataType, GemmDataType, ZDataType, LSEDataType, Acc0BiasDataType, Acc1BiasDataType, AccDataType, ShuffleDataType, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 128, 128, 32, 64, 32, 8, 8, 2, 32, 32, 1, 4, 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, 4, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, 4, MaskingSpec, Deterministic>;
using DeviceGemmInstanceBWD = using DeviceGemmInstanceBWD =
// ########################################################################################| 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| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| Gemm2| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CShuffleBlockTransferScalarPerVector_NPerBlock| MaskingSpec| Deterministic| // ########################################################################################| 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| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| Gemm2| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CShuffleBlockTransferScalarPerVector_NPerBlock| MaskingSpec| Deterministic|
...@@ -152,11 +152,11 @@ using DeviceGemmInstanceBWD = ...@@ -152,11 +152,11 @@ using DeviceGemmInstanceBWD =
#elif(DIM <= 128) #elif(DIM <= 128)
// clang-format off // clang-format off
using DeviceGemmInstanceFWD = using DeviceGemmInstanceFWD =
// #################################################################################| NumDimG| NumDimM| NumDimN| NumDimK| NumDimO| ADataType| BDataType| B1DataType| CDataType| GemmDataType| ZDataType| LSEDataType| Acc0BiasDataType| Acc1BiasDataType| GemmAcc| CShuffle| A| B| Acc| B1| C| GEMM| ATensorSpec| B0TensorSpec| B1TensorSpec| CTensorSpec| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CShuffleBlockTransferScalarPerVector| MaskingSpec| Deterministic| // #################################################################################| NumDimG| NumDimM| NumDimN| NumDimK| NumDimO| ADataType| BDataType| B1DataType| CDataType| GemmDataType| ZDataType| LSEDataType| Acc0BiasDataType| Acc1BiasDataType| GemmAcc| CShuffle| A| B| Acc| B1| C| GEMM| ATensorSpec| B0TensorSpec| B1TensorSpec| CTensorSpec| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1|Dropout| 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| D1BlockTransfer| MaskingSpec| Deterministic|
// #################################################################################| | | | | | | | | | | | | | | DataType| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| | | | | Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| _NPerBlock| | | // #################################################################################| | | | | | | | | | | | | | | DataType| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| | | | | Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| Step| 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| _NPerBlock| SrcScalar| | |
// #################################################################################| | | | | | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | | | | | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| | | | // #################################################################################| | | | | | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | | | | | Stage| | Block| Block| Block| Block| Block| | | | | | 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| | PerVector| | |
// #################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | | | | // #################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
ck::tensor_operation::device::DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, InputDataType, InputDataType, InputDataType, InputDataType, GemmDataType, ZDataType, LSEDataType, Acc0BiasDataType, Acc1BiasDataType, AccDataType, ShuffleDataType, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 128, 128, 32, 128, 32, 8, 8, 2, 32, 32, 1, 4, 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, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec, Deterministic>; ck::tensor_operation::device::DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, InputDataType, InputDataType, InputDataType, InputDataType, GemmDataType, ZDataType, LSEDataType, Acc0BiasDataType, Acc1BiasDataType, AccDataType, ShuffleDataType, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 128, 128, 32, 64, 32, 8, 8, 2, 32, 32, 1, 4, 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, 4, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, 4, MaskingSpec, Deterministic>;
using DeviceGemmInstanceBWD = using DeviceGemmInstanceBWD =
// ########################################################################################| 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| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| Gemm2| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CShuffleBlockTransferScalarPerVector_NPerBlock| MaskingSpec| Deterministic| // ########################################################################################| 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| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| Gemm2| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CShuffleBlockTransferScalarPerVector_NPerBlock| MaskingSpec| Deterministic|
...@@ -534,8 +534,8 @@ int run(int argc, char* argv[]) ...@@ -534,8 +534,8 @@ int run(int argc, char* argv[])
static_cast<InputDataType*>(y_device_buf.GetDeviceBuffer()), static_cast<InputDataType*>(y_device_buf.GetDeviceBuffer()),
static_cast<ZDataType*>(nullptr), static_cast<ZDataType*>(nullptr),
static_cast<LSEDataType*>(lse_device_buf.GetDeviceBuffer()), static_cast<LSEDataType*>(lse_device_buf.GetDeviceBuffer()),
{}, // std::array<void*, 1> p_acc0_biases; nullptr, // p_acc0_biases;
{}, // std::array<void*, 1> p_acc1_biases; nullptr, // p_acc1_biases;
q_gs_ms_ks_lengths, q_gs_ms_ks_lengths,
q_gs_ms_ks_strides, q_gs_ms_ks_strides,
k_gs_ns_ks_lengths, k_gs_ns_ks_lengths,
...@@ -594,8 +594,8 @@ int run(int argc, char* argv[]) ...@@ -594,8 +594,8 @@ int run(int argc, char* argv[])
static_cast<OutputDataType*>(qgrad_device_buf.GetDeviceBuffer()), static_cast<OutputDataType*>(qgrad_device_buf.GetDeviceBuffer()),
static_cast<OutputDataType*>(kgrad_device_buf.GetDeviceBuffer()), static_cast<OutputDataType*>(kgrad_device_buf.GetDeviceBuffer()),
static_cast<OutputDataType*>(vgrad_device_buf.GetDeviceBuffer()), static_cast<OutputDataType*>(vgrad_device_buf.GetDeviceBuffer()),
{}, // std::array<void*, 1> p_acc0_biases; nullptr, // p_acc0_biases;
{}, // std::array<void*, 1> p_acc1_biases; nullptr, // p_acc1_biases;
q_gs_ms_ks_lengths, q_gs_ms_ks_lengths,
q_gs_ms_ks_strides, q_gs_ms_ks_strides,
k_gs_ns_ks_lengths, k_gs_ns_ks_lengths,
......
example/32_batched_gemm_scale_softmax_gemm/grouped_multihead_attention_forward_v2.cpp
View file @ f752739c
...@@ -52,8 +52,8 @@ using CShuffleDataType = F32; ...@@ -52,8 +52,8 @@ using CShuffleDataType = F32;
using CDataType = DataType; using CDataType = DataType;
using ZDataType = U16; // INT32 using ZDataType = U16; // INT32
using LSEDataType = F32; using LSEDataType = F32;
using Acc0BiasDataType = ck::Tuple<>; using Acc0BiasDataType = void;
using Acc1BiasDataType = ck::Tuple<>; using Acc1BiasDataType = void;
static constexpr ck::index_t NumDimG = 2; static constexpr ck::index_t NumDimG = 2;
static constexpr ck::index_t NumDimM = 1; static constexpr ck::index_t NumDimM = 1;
...@@ -121,6 +121,7 @@ using DeviceGemmInstance = ...@@ -121,6 +121,7 @@ using DeviceGemmInstance =
1, // MXdlPerWave 1, // MXdlPerWave
4, // NXdlPerWave 4, // NXdlPerWave
1, // Gemm1NXdlPerWave 1, // Gemm1NXdlPerWave
1, // DropoutStep
S<4, 64, 1>, // ABlockTransfer S<4, 64, 1>, // ABlockTransfer
S<1, 0, 2>, S<1, 0, 2>,
S<1, 0, 2>, S<1, 0, 2>,
...@@ -135,6 +136,7 @@ using DeviceGemmInstance = ...@@ -135,6 +136,7 @@ using DeviceGemmInstance =
8, 8,
8, 8,
true, true,
1,
S<16, 16, 1>, // B1BlockTransfer S<16, 16, 1>, // B1BlockTransfer
S<0, 2, 1>, S<0, 2, 1>,
S<0, 2, 1>, S<0, 2, 1>,
...@@ -146,6 +148,7 @@ using DeviceGemmInstance = ...@@ -146,6 +148,7 @@ using DeviceGemmInstance =
1, // CShuffleNXdlPerWavePerShuffle 1, // CShuffleNXdlPerWavePerShuffle
S<1, 64, 1, 4>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock S<1, 64, 1, 4>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock 8, // CShuffleBlockTransferScalarPerVector_NPerBlock
1,
MaskingSpec, // MaskingSpecialization MaskingSpec, // MaskingSpecialization
Deterministic>; Deterministic>;
#elif(DIM <= 64) #elif(DIM <= 64)
...@@ -192,6 +195,7 @@ using DeviceGemmInstance = ...@@ -192,6 +195,7 @@ using DeviceGemmInstance =
1, // MXdlPerWave 1, // MXdlPerWave
4, // NXdlPerWave 4, // NXdlPerWave
2, // Gemm1NXdlPerWave 2, // Gemm1NXdlPerWave
1, // DropoutStep
S<4, 64, 1>, // ABlockTransfer S<4, 64, 1>, // ABlockTransfer
S<1, 0, 2>, S<1, 0, 2>,
S<1, 0, 2>, S<1, 0, 2>,
...@@ -206,6 +210,7 @@ using DeviceGemmInstance = ...@@ -206,6 +210,7 @@ using DeviceGemmInstance =
8, 8,
8, 8,
true, true,
1,
S<16, 16, 1>, // B1BlockTransfer S<16, 16, 1>, // B1BlockTransfer
S<0, 2, 1>, S<0, 2, 1>,
S<0, 2, 1>, S<0, 2, 1>,
...@@ -217,6 +222,7 @@ using DeviceGemmInstance = ...@@ -217,6 +222,7 @@ using DeviceGemmInstance =
2, // CShuffleNXdlPerWavePerShuffle 2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock 8, // CShuffleBlockTransferScalarPerVector_NPerBlock
1,
MaskingSpec, // MaskingSpecialization MaskingSpec, // MaskingSpecialization
Deterministic>; Deterministic>;
#elif(DIM <= 128) #elif(DIM <= 128)
...@@ -253,7 +259,7 @@ using DeviceGemmInstance = ...@@ -253,7 +259,7 @@ using DeviceGemmInstance =
128, // MPerBlock 128, // MPerBlock
128, // NPerBlock 128, // NPerBlock
32, // KPerBlock 32, // KPerBlock
128, // Gemm1NPerBlock 64, // Gemm1NPerBlock
32, // Gemm1KPerBlock 32, // Gemm1KPerBlock
8, // AK1 8, // AK1
8, // BK1 8, // BK1
...@@ -262,7 +268,8 @@ using DeviceGemmInstance = ...@@ -262,7 +268,8 @@ using DeviceGemmInstance =
32, // NPerXDL 32, // NPerXDL
1, // MXdlPerWave 1, // MXdlPerWave
4, // NXdlPerWave 4, // NXdlPerWave
4, // Gemm1NXdlPerWave 2, // Gemm1NXdlPerWave
1, // DropoutStep
S<4, 64, 1>, // ABlockTransfer S<4, 64, 1>, // ABlockTransfer
S<1, 0, 2>, S<1, 0, 2>,
S<1, 0, 2>, S<1, 0, 2>,
...@@ -277,7 +284,8 @@ using DeviceGemmInstance = ...@@ -277,7 +284,8 @@ using DeviceGemmInstance =
8, 8,
8, 8,
true, true,
S<8, 32, 1>, // B1BlockTransfer 1,
S<16, 16, 1>, // B1BlockTransfer
S<0, 2, 1>, S<0, 2, 1>,
S<0, 2, 1>, S<0, 2, 1>,
1, 1,
...@@ -288,6 +296,7 @@ using DeviceGemmInstance = ...@@ -288,6 +296,7 @@ using DeviceGemmInstance =
2, // CShuffleNXdlPerWavePerShuffle 2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock 8, // CShuffleBlockTransferScalarPerVector_NPerBlock
1,
MaskingSpec, // MaskingSpecialization MaskingSpec, // MaskingSpecialization
Deterministic>; Deterministic>;
#endif #endif
......
example/32_batched_gemm_scale_softmax_gemm/grouped_multihead_attention_train_v2.cpp
View file @ f752739c
...@@ -78,8 +78,8 @@ using AccDataType = F32; ...@@ -78,8 +78,8 @@ using AccDataType = F32;
using ShuffleDataType = F32; using ShuffleDataType = F32;
using LSEDataType = F32; using LSEDataType = F32;
using ZDataType = U16; // INT32 using ZDataType = U16; // INT32
using Acc0BiasDataType = ck::Tuple<>; using Acc0BiasDataType = void;
using Acc1BiasDataType = ck::Tuple<>; using Acc1BiasDataType = void;
static constexpr ck::index_t NumDimG = 2; static constexpr ck::index_t NumDimG = 2;
static constexpr ck::index_t NumDimM = 1; static constexpr ck::index_t NumDimM = 1;
...@@ -112,11 +112,11 @@ static constexpr bool Deterministic = false; ...@@ -112,11 +112,11 @@ static constexpr bool Deterministic = false;
#if(DIM <= 32) #if(DIM <= 32)
// clang-format off // clang-format off
using DeviceGemmInstanceFWD = using DeviceGemmInstanceFWD =
// #################################################################################| NumDimG| NumDimM| NumDimN| NumDimK| NumDimO| ADataType| BDataType| B1DataType| CDataType| GemmDataType| ZDataType| LSEDataType| Acc0BiasDataType| Acc1BiasDataType| GemmAcc| CShuffle| A| B| Acc| B1| C| GEMM| ATensorSpec| B0TensorSpec| B1TensorSpec| CTensorSpec| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CShuffleBlockTransferScalarPerVector| MaskingSpec| Deterministic| // #################################################################################| NumDimG| NumDimM| NumDimN| NumDimK| NumDimO| ADataType| BDataType| B1DataType| CDataType| GemmDataType| ZDataType| LSEDataType| Acc0BiasDataType| Acc1BiasDataType| GemmAcc| CShuffle| A| B| Acc| B1| C| GEMM| ATensorSpec| B0TensorSpec| B1TensorSpec| CTensorSpec| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1|Dropout| 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| D1BlockTransfer| MaskingSpec| Deterministic|
// #################################################################################| | | | | | | | | | | | | | | DataType| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| | | | | Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| _NPerBlock| | | // #################################################################################| | | | | | | | | | | | | | | DataType| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| | | | | Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| Step| 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| _NPerBlock| SrcScalar| | |
// #################################################################################| | | | | | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | | | | | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| | | | // #################################################################################| | | | | | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | | | | | Stage| | Block| Block| Block| Block| Block| | | | | | 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| | PerVector| | |
// #################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | | | | // #################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
ck::tensor_operation::device::DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, InputDataType, InputDataType, InputDataType, InputDataType, GemmDataType, ZDataType, LSEDataType, Acc0BiasDataType, Acc1BiasDataType, AccDataType, ShuffleDataType, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 128, 128, 32, 32, 32, 8, 8, 2, 32, 32, 1, 4, 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, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, false, 1, 1, S<1, 64, 1, 4>, 8, MaskingSpec, Deterministic>; ck::tensor_operation::device::DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, InputDataType, InputDataType, InputDataType, InputDataType, GemmDataType, ZDataType, LSEDataType, Acc0BiasDataType, Acc1BiasDataType, AccDataType, ShuffleDataType, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 128, 128, 32, 32, 32, 8, 8, 2, 32, 32, 1, 4, 1, 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, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, false, 1, 1, S<1, 64, 1, 4>, 8, 1, MaskingSpec, Deterministic>;
using DeviceGemmInstanceBWD = using DeviceGemmInstanceBWD =
// ########################################################################################| 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| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| Gemm2| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CShuffleBlockTransferScalarPerVector_NPerBlock| MaskingSpec| Deterministic| // ########################################################################################| 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| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| Gemm2| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CShuffleBlockTransferScalarPerVector_NPerBlock| MaskingSpec| Deterministic|
...@@ -128,11 +128,11 @@ using DeviceGemmInstanceBWD = ...@@ -128,11 +128,11 @@ using DeviceGemmInstanceBWD =
#elif(DIM <= 64) #elif(DIM <= 64)
// clang-format off // clang-format off
using DeviceGemmInstanceFWD = using DeviceGemmInstanceFWD =
// #################################################################################| NumDimG| NumDimM| NumDimN| NumDimK| NumDimO| ADataType| BDataType| B1DataType| CDataType| GemmDataType| ZDataType| LSEDataType| Acc0BiasDataType| Acc1BiasDataType| GemmAcc| CShuffle| A| B| Acc| B1| C| GEMM| ATensorSpec| B0TensorSpec| B1TensorSpec| CTensorSpec| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CShuffleBlockTransferScalarPerVector| MaskingSpec| Deterministic| // #################################################################################| NumDimG| NumDimM| NumDimN| NumDimK| NumDimO| ADataType| BDataType| B1DataType| CDataType| GemmDataType| ZDataType| LSEDataType| Acc0BiasDataType| Acc1BiasDataType| GemmAcc| CShuffle| A| B| Acc| B1| C| GEMM| ATensorSpec| B0TensorSpec| B1TensorSpec| CTensorSpec| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1|Dropout| 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| D1BlockTransfer| MaskingSpec| Deterministic|
// #################################################################################| | | | | | | | | | | | | | | DataType| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| | | | | Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| _NPerBlock| | | // #################################################################################| | | | | | | | | | | | | | | DataType| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| | | | | Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| Step| 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| _NPerBlock| SrcScalar| | |
// #################################################################################| | | | | | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | | | | | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| | | | // #################################################################################| | | | | | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | | | | | Stage| | Block| Block| Block| Block| Block| | | | | | 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| | PerVector| | |
// #################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | | | | // #################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
ck::tensor_operation::device::DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, InputDataType, InputDataType, InputDataType, InputDataType, GemmDataType, ZDataType, LSEDataType, Acc0BiasDataType, Acc1BiasDataType, AccDataType, ShuffleDataType, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 128, 128, 32, 64, 32, 8, 8, 2, 32, 32, 1, 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, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec, Deterministic>; ck::tensor_operation::device::DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, InputDataType, InputDataType, InputDataType, InputDataType, GemmDataType, ZDataType, LSEDataType, Acc0BiasDataType, Acc1BiasDataType, AccDataType, ShuffleDataType, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 128, 128, 32, 64, 32, 8, 8, 2, 32, 32, 1, 4, 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, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, 1, MaskingSpec, Deterministic>;
using DeviceGemmInstanceBWD = using DeviceGemmInstanceBWD =
// ########################################################################################| 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| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| Gemm2| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CShuffleBlockTransferScalarPerVector_NPerBlock| MaskingSpec| Deterministic| // ########################################################################################| 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| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| Gemm2| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CShuffleBlockTransferScalarPerVector_NPerBlock| MaskingSpec| Deterministic|
...@@ -151,11 +151,11 @@ using DeviceGemmInstanceBWD = ...@@ -151,11 +151,11 @@ using DeviceGemmInstanceBWD =
#elif(DIM <= 128) #elif(DIM <= 128)
// clang-format off // clang-format off
using DeviceGemmInstanceFWD = using DeviceGemmInstanceFWD =
// #################################################################################| NumDimG| NumDimM| NumDimN| NumDimK| NumDimO| ADataType| BDataType| B1DataType| CDataType| GemmDataType| ZDataType| LSEDataType| Acc0BiasDataType| Acc1BiasDataType| GemmAcc| CShuffle| A| B| Acc| B1| C| GEMM| ATensorSpec| B0TensorSpec| B1TensorSpec| CTensorSpec| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CShuffleBlockTransferScalarPerVector| MaskingSpec| Deterministic| // #################################################################################| NumDimG| NumDimM| NumDimN| NumDimK| NumDimO| ADataType| BDataType| B1DataType| CDataType| GemmDataType| ZDataType| LSEDataType| Acc0BiasDataType| Acc1BiasDataType| GemmAcc| CShuffle| A| B| Acc| B1| C| GEMM| ATensorSpec| B0TensorSpec| B1TensorSpec| CTensorSpec| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1|Dropout| 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| D1BlockTransfer| MaskingSpec| Deterministic|
// #################################################################################| | | | | | | | | | | | | | | DataType| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| | | | | Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| _NPerBlock| | | // #################################################################################| | | | | | | | | | | | | | | DataType| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| | | | | Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| Step| 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| _NPerBlock| SrcScalar| | |
// #################################################################################| | | | | | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | | | | | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| | | | // #################################################################################| | | | | | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | | | | | Stage| | Block| Block| Block| Block| Block| | | | | | 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| | PerVector| | |
// #################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | | | | // #################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
ck::tensor_operation::device::DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, InputDataType, InputDataType, InputDataType, InputDataType, GemmDataType, ZDataType, LSEDataType, Acc0BiasDataType, Acc1BiasDataType, AccDataType, ShuffleDataType, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 128, 128, 32, 128, 32, 8, 8, 2, 32, 32, 1, 4, 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, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, MaskingSpec, Deterministic>; ck::tensor_operation::device::DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, InputDataType, InputDataType, InputDataType, InputDataType, GemmDataType, ZDataType, LSEDataType, Acc0BiasDataType, Acc1BiasDataType, AccDataType, ShuffleDataType, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 128, 128, 32, 64, 32, 8, 8, 2, 32, 32, 1, 4, 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, S<16, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 2, S<1, 32, 1, 8>, 8, 1, MaskingSpec, Deterministic>;
using DeviceGemmInstanceBWD = using DeviceGemmInstanceBWD =
// ########################################################################################| 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| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| Gemm2| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CShuffleBlockTransferScalarPerVector_NPerBlock| MaskingSpec| Deterministic| // ########################################################################################| 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| AK1| BK1| B1K1| MPer| NPer| Gemm0| Gemm0| Gemm1| Gemm2| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockTransfer| B0BlockLds| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockTransfer| B1BlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CShuffleBlockTransferScalarPerVector_NPerBlock| MaskingSpec| Deterministic|
......
example/32_batched_gemm_scale_softmax_gemm/run_batched_multihead_attention_forward.inc
View file @ f752739c
...@@ -177,8 +177,8 @@ int run(int argc, char* argv[]) ...@@ -177,8 +177,8 @@ int run(int argc, char* argv[])
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()), static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
static_cast<ZDataType*>(nullptr), static_cast<ZDataType*>(nullptr),
static_cast<LSEDataType*>(lse_device_buf.GetDeviceBuffer()), static_cast<LSEDataType*>(lse_device_buf.GetDeviceBuffer()),
{}, // std::array<void*, 1> p_acc0_biases; nullptr, // std::array<void*, 1> p_acc0_biases;
{}, // std::array<void*, 1> p_acc1_biases; nullptr, // std::array<void*, 1> p_acc1_biases;
a_gs_ms_ks_lengths, a_gs_ms_ks_lengths,
a_gs_ms_ks_strides, a_gs_ms_ks_strides,
b0_gs_ns_ks_lengths, b0_gs_ns_ks_lengths,
......
example/52_flash_atten_bias/CMakeLists.txt 0 → 100644
View file @ f752739c
add_example_executable(example_batched_multihead_attention_bias_forward_v2 batched_multihead_attention_bias_forward_v2.cpp)
add_example_executable(example_grouped_multihead_attention_bias_forward_v2 grouped_multihead_attention_bias_forward_v2.cpp)
\ No newline at end of file
example/52_flash_atten_bias/batched_multihead_attention_bias_forward_v2.cpp 0 → 100644
View file @ f752739c
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
/*
Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g_k_n) * B1_g_n_o
|-----------------|
Gemm0
|-------------------------------------|
Gemm1
*/
#define DIM 128 // DIM should be a multiple of 8.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_batched_mha_fwd_xdl_cshuffle_v2.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_dropout.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using F32 = float;
using U16 = unsigned short;
using INT32 = int32_t;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DataType = F16;
using GemmDataType = F16;
using ADataType = DataType;
using B0DataType = DataType;
using B1DataType = DataType;
using AccDataType = F32;
using CShuffleDataType = F32;
using CDataType = DataType;
using DDataType = F16;
using ZDataType = U16; // INT32
using LSEDataType = F32;
using Acc0BiasDataType = DDataType;
using Acc1BiasDataType = void;
static constexpr ck::index_t NumDimG = 2;
static constexpr ck::index_t NumDimM = 1;
static constexpr ck::index_t NumDimN = 1;
static constexpr ck::index_t NumDimK = 1;
static constexpr ck::index_t NumDimO = 1;
using AElementOp = PassThrough;
using B0ElementOp = PassThrough;
using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
using B1ElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
static constexpr auto MaskingSpec =
ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
static constexpr auto TensorSpecA = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecC = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr bool Deterministic = false;
#if(DIM <= 32)
using DeviceGemmInstance =
ck::tensor_operation::device::DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2<
NumDimG,
NumDimM,
NumDimN,
NumDimK,
NumDimO,
ADataType,
B0DataType,
B1DataType,
CDataType,
GemmDataType,
ZDataType,
LSEDataType,
Acc0BiasDataType,
Acc1BiasDataType,
AccDataType,
CShuffleDataType,
AElementOp,
B0ElementOp,
Acc0ElementOp,
B1ElementOp,
CElementOp,
GemmSpec,
TensorSpecA,
TensorSpecB0,
TensorSpecB1,
TensorSpecC,
1,
256,
128, // MPerBlock
128, // NPerBlock
32, // KPerBlock
32, // Gemm1NPerBlock
32, // Gemm1KPerBlock
8, // AK1
8, // BK1
2, // B1K1
32, // MPerXDL
32, // NPerXDL
1, // MXdlPerWave
4, // NXdlPerWave
1, // Gemm1NXdlPerWave
1, // DropoutStep
S<4, 64, 1>, // ABlockTransfer
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<4, 64, 1>, // BBlockTransfer
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
4,
S<16, 16, 1>, // B1BlockTransfer
S<0, 2, 1>,
S<0, 2, 1>,
1,
2,
2,
false,
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 64, 1, 4>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock
4,
MaskingSpec, // MaskingSpecialization
Deterministic>;
#elif(DIM <= 64)
using DeviceGemmInstance =
ck::tensor_operation::device::DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2<
NumDimG,
NumDimM,
NumDimN,
NumDimK,
NumDimO,
ADataType,
B0DataType,
B1DataType,
CDataType,
GemmDataType,
ZDataType,
LSEDataType,
Acc0BiasDataType,
Acc1BiasDataType,
AccDataType,
CShuffleDataType,
AElementOp,
B0ElementOp,
Acc0ElementOp,
B1ElementOp,
CElementOp,
GemmSpec,
TensorSpecA,
TensorSpecB0,
TensorSpecB1,
TensorSpecC,
1,
256,
128, // MPerBlock
128, // NPerBlock
32, // KPerBlock
64, // Gemm1NPerBlock
32, // Gemm1KPerBlock
8, // AK1
8, // BK1
2, // B1K1
32, // MPerXDL
32, // NPerXDL
1, // MXdlPerWave
4, // NXdlPerWave
2, // Gemm1NXdlPerWave
1, // DropoutStep
S<4, 64, 1>, // ABlockTransfer
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<4, 64, 1>, // BBlockTransfer
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
4,
S<16, 16, 1>, // B1BlockTransfer
S<0, 2, 1>,
S<0, 2, 1>,
1,
4,
2,
false,
1, // CShuffleMXdlPerWavePerShuffle
2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock
4,
MaskingSpec, // MaskingSpecialization
Deterministic>;
#elif(DIM <= 128)
using DeviceGemmInstance =
ck::tensor_operation::device::DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2<
NumDimG,
NumDimM,
NumDimN,
NumDimK,
NumDimO,
ADataType,
B0DataType,
B1DataType,
CDataType,
GemmDataType,
ZDataType,
LSEDataType,
Acc0BiasDataType,
Acc1BiasDataType,
AccDataType,
CShuffleDataType,
AElementOp,
B0ElementOp,
Acc0ElementOp,
B1ElementOp,
CElementOp,
GemmSpec,
TensorSpecA,
TensorSpecB0,
TensorSpecB1,
TensorSpecC,
1,
256,
128, // MPerBlock
128, // NPerBlock
32, // KPerBlock
128, // Gemm1NPerBlock
32, // Gemm1KPerBlock
8, // AK1
8, // BK1
2, // B1K1
32, // MPerXDL
32, // NPerXDL
1, // MXdlPerWave
4, // NXdlPerWave
4, // Gemm1NXdlPerWave
1, // DropoutStep
S<4, 64, 1>, // ABlockTransfer
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<4, 64, 1>, // BBlockTransfer
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
4,
S<8, 32, 1>, // B1BlockTransfer
S<0, 2, 1>,
S<0, 2, 1>,
1,
4,
2,
false,
1, // CShuffleMXdlPerWavePerShuffle
2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock
4,
MaskingSpec, // MaskingSpecialization
Deterministic>;
#endif
// Ref Gemm0: DataType in, AccDataType out
using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
B0DataType,
AccDataType,
AccDataType,
AElementOp,
B0ElementOp,
Acc0ElementOp>;
// Ref Softmax: AccDataType in, DataType out
using ReferenceSoftmaxInstance =
ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
// Ref Gemm1: DataType in, DataType out
using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
B1DataType,
CDataType,
AccDataType,
AElementOp,
B1ElementOp,
CElementOp>;
// Ref dropout
using ReferenceDropoutInstance =
ck::tensor_operation::host::ReferenceDropout<ZDataType, ADataType, ADataType>;
#include "run_batched_multihead_attention_bias_forward.inc"
int main(int argc, char* argv[]) { return run(argc, argv); }
example/52_flash_atten_bias/grouped_multihead_attention_bias_forward_v2.cpp 0 → 100644
View file @ f752739c
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
/*
Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g_k_n) * B1_g_n_o
|-----------------|
Gemm0
|-------------------------------------|
Gemm1
*/
#define DIM 64 // DIM should be a multiple of 8.
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_mha_fwd_xdl_cshuffle_v2.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_batched_gemm.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_softmax.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_dropout.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using F32 = float;
using U16 = unsigned short;
using INT32 = int32_t;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using DataType = F16;
using GemmDataType = F16;
using ADataType = DataType;
using B0DataType = DataType;
using B1DataType = DataType;
using AccDataType = F32;
using DDataType = F16;
using CShuffleDataType = F32;
using CDataType = DataType;
using ZDataType = U16; // INT32
using LSEDataType = F32;
using Acc0BiasDataType = DDataType;
using Acc1BiasDataType = void;
static constexpr ck::index_t NumDimG = 2;
static constexpr ck::index_t NumDimM = 1;
static constexpr ck::index_t NumDimN = 1;
static constexpr ck::index_t NumDimK = 1;
static constexpr ck::index_t NumDimO = 1;
using AElementOp = PassThrough;
using B0ElementOp = PassThrough;
using Acc0ElementOp = ck::tensor_operation::element_wise::Scale;
using B1ElementOp = PassThrough;
using CElementOp = PassThrough;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKOPadding;
static constexpr auto MaskingSpec =
ck::tensor_operation::device::MaskingSpecialization::MaskDisabled;
static constexpr auto TensorSpecA = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecB0 = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecB1 = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr auto TensorSpecC = ck::tensor_operation::device::TensorSpecialization::Default;
static constexpr bool Deterministic = false;
#if(DIM <= 32)
using DeviceGemmInstance =
ck::tensor_operation::device::DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2<
NumDimG,
NumDimM,
NumDimN,
NumDimK,
NumDimO,
ADataType,
B0DataType,
B1DataType,
CDataType,
GemmDataType,
ZDataType,
LSEDataType,
Acc0BiasDataType,
Acc1BiasDataType,
AccDataType,
CShuffleDataType,
AElementOp,
B0ElementOp,
Acc0ElementOp,
B1ElementOp,
CElementOp,
GemmSpec,
TensorSpecA,
TensorSpecB0,
TensorSpecB1,
TensorSpecC,
1,
256,
128, // MPerBlock
128, // NPerBlock
32, // KPerBlock
32, // Gemm1NPerBlock
32, // Gemm1KPerBlock
8, // AK1
8, // BK1
2, // B1K1
32, // MPerXDL
32, // NPerXDL
1, // MXdlPerWave
4, // NXdlPerWave
1, // Gemm1NXdlPerWave
1, // DropoutStep
S<4, 64, 1>, // ABlockTransfer
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<4, 64, 1>, // BBlockTransfer
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
1,
S<16, 16, 1>, // B1BlockTransfer
S<0, 2, 1>,
S<0, 2, 1>,
1,
2,
2,
false,
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 64, 1, 4>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock
1,
MaskingSpec, // MaskingSpecialization
Deterministic>;
#elif(DIM <= 64)
using DeviceGemmInstance =
ck::tensor_operation::device::DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2<
NumDimG,
NumDimM,
NumDimN,
NumDimK,
NumDimO,
ADataType,
B0DataType,
B1DataType,
CDataType,
GemmDataType,
ZDataType,
LSEDataType,
Acc0BiasDataType,
Acc1BiasDataType,
AccDataType,
CShuffleDataType,
AElementOp,
B0ElementOp,
Acc0ElementOp,
B1ElementOp,
CElementOp,
GemmSpec,
TensorSpecA,
TensorSpecB0,
TensorSpecB1,
TensorSpecC,
1,
256,
128, // MPerBlock
128, // NPerBlock
32, // KPerBlock
64, // Gemm1NPerBlock
32, // Gemm1KPerBlock
8, // AK1
8, // BK1
2, // B1K1
32, // MPerXDL
32, // NPerXDL
1, // MXdlPerWave
4, // NXdlPerWave
2, // Gemm1NXdlPerWave
1, // DropoutStep
S<4, 64, 1>, // ABlockTransfer
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<4, 64, 1>, // BBlockTransfer
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
1,
S<16, 16, 1>, // B1BlockTransfer
S<0, 2, 1>,
S<0, 2, 1>,
1,
4,
2,
false,
1, // CShuffleMXdlPerWavePerShuffle
2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock
1,
MaskingSpec, // MaskingSpecialization
Deterministic>;
#elif(DIM <= 128)
using DeviceGemmInstance =
ck::tensor_operation::device::DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2<
NumDimG,
NumDimM,
NumDimN,
NumDimK,
NumDimO,
ADataType,
B0DataType,
B1DataType,
CDataType,
GemmDataType,
ZDataType,
LSEDataType,
Acc0BiasDataType,
Acc1BiasDataType,
AccDataType,
CShuffleDataType,
AElementOp,
B0ElementOp,
Acc0ElementOp,
B1ElementOp,
CElementOp,
GemmSpec,
TensorSpecA,
TensorSpecB0,
TensorSpecB1,
TensorSpecC,
1,
256,
128, // MPerBlock
128, // NPerBlock
32, // KPerBlock
128, // Gemm1NPerBlock
32, // Gemm1KPerBlock
8, // AK1
8, // BK1
2, // B1K1
32, // MPerXDL
32, // NPerXDL
1, // MXdlPerWave
4, // NXdlPerWave
4, // Gemm1NXdlPerWave
1, // DropoutStep
S<4, 64, 1>, // ABlockTransfer
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
S<4, 64, 1>, // BBlockTransfer
S<1, 0, 2>,
S<1, 0, 2>,
2,
8,
8,
true,
1,
S<8, 32, 1>, // B1BlockTransfer
S<0, 2, 1>,
S<0, 2, 1>,
1,
4,
2,
false,
1, // CShuffleMXdlPerWavePerShuffle
2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock
1,
MaskingSpec, // MaskingSpecialization
Deterministic>;
#endif
// Ref Gemm0: DataType in, AccDataType out
using ReferenceGemm0Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
B0DataType,
AccDataType,
AccDataType,
AElementOp,
B0ElementOp,
Acc0ElementOp>;
// Ref Softmax: AccDataType in, DataType out
using ReferenceSoftmaxInstance =
ck::tensor_operation::host::ReferenceSoftmax<AccDataType, ADataType, AccDataType>;
// Ref Gemm1: DataType in, DataType out
using ReferenceGemm1Instance = ck::tensor_operation::host::ReferenceBatchedGemm<ADataType,
B1DataType,
CDataType,
AccDataType,
AElementOp,
B1ElementOp,
CElementOp>;
// Ref dropout
using ReferenceDropoutInstance =
ck::tensor_operation::host::ReferenceDropout<ZDataType, ADataType, ADataType>;
#include "run_grouped_multihead_attention_bias_forward.inc"
int main(int argc, char* argv[]) { return run(argc, argv); }
example/52_flash_atten_bias/run_batched_multihead_attention_bias_forward.inc 0 → 100644
View file @ f752739c
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
int run(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = true;
// GEMM shape for A/B0/B1/C
// C_g_m_o = A_g_m_k * B0_g_k_n * B1_g_n_o
ck::index_t M = 1000; // 120
ck::index_t N = 1000; // 1000
ck::index_t K = DIM;
ck::index_t O = DIM;
// Output shape C[G0, M, G1, O]. Batch dim, outer dim, inner dim must match GEMM shape
// C_g0_g1_m_o = reshape(C_g_m_o, [g0, g1, m, o])
// C_g0_m_g1_o = permute(C_g0_g1_m_o, [0, 2, 1, 3])
ck::index_t G0 = 7;
ck::index_t G1 = 13;
bool input_permute = false;
bool output_permute = true;
float p_drop = 0.1;
const unsigned long long seed = 1;
const unsigned long long offset = 0;
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 13)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
K = std::stoi(argv[6]);
O = std::stoi(argv[7]);
G0 = std::stoi(argv[8]);
G1 = std::stoi(argv[9]);
p_drop = std::stof(argv[10]);
input_permute = std::stoi(argv[11]);
output_permute = std::stoi(argv[12]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 11: M, N, K, O, G0, G1\n");
printf("arg10: scale (alpha)\n");
printf("arg11 to 12: input / output permute\n");
exit(0);
}
float p_dropout = 1 - p_drop;
ZDataType p_dropout_in_16bits = ZDataType(std::floor(p_dropout * 65535.0));
float rp_dropout = 1.0 / p_dropout;
float alpha = 1.f / std::sqrt(K);
std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
std::vector<ck::index_t> a_gs_ms_ks_strides =
input_permute
? std::vector<ck::index_t>{M * G1 * K, K, G1 * K, 1} // A layout [G0, M, G1, K]
: std::vector<ck::index_t>{G1 * M * K, M * K, K, 1}; // A layout [G0, G1, M, K]
std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
std::vector<ck::index_t> b0_gs_ns_ks_strides =
input_permute
? std::vector<ck::index_t>{N * G1 * K, K, G1 * K, 1} // B0 layout [G0, N, G1, K]
: std::vector<ck::index_t>{G1 * N * K, N * K, K, 1}; // B0 layout [G0, G1, N, K]
std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
std::vector<ck::index_t> b1_gs_os_ns_strides =
input_permute
? std::vector<ck::index_t>{N * G1 * O, O, 1, G1 * O} // B1 layout [G0, N, G1, O]
: std::vector<ck::index_t>{G1 * N * O, N * O, 1, O}; // B1 layout [G0, G1, N, O]
std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
std::vector<ck::index_t> c_gs_ms_os_strides =
output_permute
? std::vector<ck::index_t>{M * G1 * O, O, G1 * O, 1} // C layout [G0, M, G1, O]
: std::vector<ck::index_t>{G1 * M * O, M * O, O, 1}; // C layout [G0, G1, M, O]
std::vector<ck::index_t> d_gs_ms_ns_lengths{G0, G1, M, N};
std::vector<ck::index_t> d_gs_ms_ns_strides =
input_permute
? std::vector<ck::index_t>{M * G1 * N, N, G1 * N, 1} // D layout [G0, M, G1, N]
: std::vector<ck::index_t>{G1 * M * N, M * N, N, 1}; // D layout [G0, G1, M, N]
std::vector<ck::index_t> z_gs_ms_ns_lengths{G0, G1, M, N};
std::vector<ck::index_t> z_gs_ms_ns_strides =
input_permute
? std::vector<ck::index_t>{M * G1 * N, N, G1 * N, 1} // Z layout [G0, M, G1, N]
: std::vector<ck::index_t>{G1 * M * N, M * N, N, 1}; // Z layout [G0, G1, M, N]
std::vector<ck::index_t> lse_gs_ms_lengths{G0, G1, M};
std::vector<ck::index_t> lse_gs_ms_strides =
std::vector<ck::index_t>{G1 * M, M, 1}; // LSE layout [G0, G1, M]
Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
Tensor<B0DataType> b0_gs_ns_ks(b0_gs_ns_ks_lengths, b0_gs_ns_ks_strides);
Tensor<B1DataType> b1_gs_os_ns(b1_gs_os_ns_lengths, b1_gs_os_ns_strides);
Tensor<CDataType> c_gs_ms_os_host_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
Tensor<CDataType> c_gs_ms_os_device_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
Tensor<DDataType> d_gs_ms_ns(d_gs_ms_ns_lengths, z_gs_ms_ns_strides);
Tensor<ZDataType> z_gs_ms_ns(z_gs_ms_ns_lengths, z_gs_ms_ns_strides);
Tensor<LSEDataType> lse_gs_ms_host_result(lse_gs_ms_lengths, lse_gs_ms_strides);
Tensor<LSEDataType> lse_gs_ms_device_result(lse_gs_ms_lengths, lse_gs_ms_strides);
std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
std::cout << "b0_gs_ns_ks: " << b0_gs_ns_ks.mDesc << std::endl;
std::cout << "b1_gs_os_ns: " << b1_gs_os_ns.mDesc << std::endl;
std::cout << "c_gs_ms_os: " << c_gs_ms_os_host_result.mDesc << std::endl;
std::cout << "z_gs_ms_ns: " << z_gs_ms_ns.mDesc << std::endl;
std::cout << "lse_gs_ms_os: " << lse_gs_ms_host_result.mDesc << std::endl;
z_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<ZDataType>{0});
switch(init_method)
{
case 0: break;
case 1:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_2<DDataType>{-1, 1});
break;
case 2:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_3<DDataType>{-0.5, 0.5});
break;
case 3:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<DDataType>{1});
break;
default:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_Sequential<2>{});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<DDataType>{1});
}
DeviceMem a_device_buf(sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize());
DeviceMem b0_device_buf(sizeof(B0DataType) * b0_gs_ns_ks.mDesc.GetElementSpaceSize());
DeviceMem b1_device_buf(sizeof(B1DataType) * b1_gs_os_ns.mDesc.GetElementSpaceSize());
DeviceMem c_device_buf(sizeof(CDataType) *
c_gs_ms_os_device_result.mDesc.GetElementSpaceSize());
DeviceMem d_device_buf(sizeof(DDataType) * d_gs_ms_ns.mDesc.GetElementSpaceSize());
DeviceMem z_device_buf(sizeof(ZDataType) * z_gs_ms_ns.mDesc.GetElementSpaceSize());
DeviceMem lse_device_buf(sizeof(LSEDataType) *
lse_gs_ms_device_result.mDesc.GetElementSpaceSize());
a_device_buf.ToDevice(a_gs_ms_ks.mData.data());
b0_device_buf.ToDevice(b0_gs_ns_ks.mData.data());
b1_device_buf.ToDevice(b1_gs_os_ns.mData.data());
d_device_buf.ToDevice(d_gs_ms_ns.mData.data());
auto a_element_op = AElementOp{};
auto b0_element_op = B0ElementOp{};
auto acc0_element_op = Acc0ElementOp{alpha};
auto b1_element_op = B1ElementOp{};
auto c_element_op = CElementOp{};
// do GEMM
// TODO ANT: replace array with vector?
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument(
static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<B0DataType*>(b0_device_buf.GetDeviceBuffer()),
static_cast<B1DataType*>(b1_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
static_cast<ZDataType*>(nullptr),
static_cast<LSEDataType*>(lse_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d_device_buf.GetDeviceBuffer()), //
nullptr,
a_gs_ms_ks_lengths,
a_gs_ms_ks_strides,
b0_gs_ns_ks_lengths,
b0_gs_ns_ks_strides,
b1_gs_os_ns_lengths,
b1_gs_os_ns_strides,
c_gs_ms_os_lengths,
c_gs_ms_os_strides,
z_gs_ms_ns_lengths,
z_gs_ms_ns_strides,
lse_gs_ms_lengths,
d_gs_ms_ns_lengths, // acc0_biases_gs_ms_ns_lengths
d_gs_ms_ns_strides, // acc0_biases_gs_ms_ns_strides
{}, // std::vector<ck::index_t>
{}, // std::vector<ck::index_t>
a_element_op,
b0_element_op,
acc0_element_op,
b1_element_op,
c_element_op,
p_drop, // dropout ratio
{seed, offset}); // dropout random seed and offset, offset should be at least the number of
// elements on a thread
if(!gemm.IsSupportedArgument(argument))
{
std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
return 0;
}
ck::index_t BatchCount = G0 * G1;
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * BatchCount;
std::size_t num_btype = (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O +
sizeof(DDataType) * M * N * std::is_void<DDataType>::value?1:0) *
BatchCount;
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, "
<< gemm.GetTypeString() << std::endl;
if(do_verification)
{
// run for storing z tensor
argument = gemm.MakeArgument(
static_cast<ADataType*>(a_device_buf.GetDeviceBuffer()),
static_cast<B0DataType*>(b0_device_buf.GetDeviceBuffer()),
static_cast<B1DataType*>(b1_device_buf.GetDeviceBuffer()),
static_cast<CDataType*>(c_device_buf.GetDeviceBuffer()),
static_cast<ZDataType*>(z_device_buf.GetDeviceBuffer()),
static_cast<LSEDataType*>(lse_device_buf.GetDeviceBuffer()),
static_cast<DDataType*>(d_device_buf.GetDeviceBuffer()),
nullptr,
a_gs_ms_ks_lengths,
a_gs_ms_ks_strides,
b0_gs_ns_ks_lengths,
b0_gs_ns_ks_strides,
b1_gs_os_ns_lengths,
b1_gs_os_ns_strides,
c_gs_ms_os_lengths,
c_gs_ms_os_strides,
z_gs_ms_ns_lengths,
z_gs_ms_ns_strides,
lse_gs_ms_lengths,
d_gs_ms_ns_lengths,
d_gs_ms_ns_strides,
{},
{},
a_element_op,
b0_element_op,
acc0_element_op,
b1_element_op,
c_element_op,
p_drop, // dropout ratio
{seed, offset}); // dropout random seed and offset, offset should be at least the number
// of elements on a thread
c_device_buf.SetZero();
lse_device_buf.SetZero();
invoker.Run(argument, StreamConfig{nullptr, false});
c_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
z_device_buf.FromDevice(z_gs_ms_ns.mData.data());
lse_device_buf.FromDevice(lse_gs_ms_device_result.mData.data());
Tensor<ADataType> a_g_m_k({BatchCount, M, K});
Tensor<B0DataType> b0_g_k_n({BatchCount, K, N});
Tensor<B1DataType> b1_g_n_o({BatchCount, N, O});
Tensor<AccDataType> acc0_g_m_n({BatchCount, M, N}); // scratch object after gemm0
Tensor<ADataType> a1_g_m_n({BatchCount, M, N}); // scratch object after softmax
Tensor<ADataType> a1_g_m_n_drop({G0 * G1, M, N});
Tensor<LSEDataType> lse_g_m_host_result(
{BatchCount, M}); // scratch object after max + ln(sum)
Tensor<DDataType> d_g_m_n({G0 * G1, M, N});
Tensor<ZDataType> z_g_m_n({G0 * G1, M, N});
Tensor<CDataType> c_g_m_o_host_result({BatchCount, M, O}); // scratch object after gemm1
// permute
a_gs_ms_ks.ForEach([&](auto& self, auto idx) {
a_g_m_k(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
});
b0_gs_ns_ks.ForEach([&](auto& self, auto idx) {
b0_g_k_n(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
});
b1_gs_os_ns.ForEach([&](auto& self, auto idx) {
b1_g_n_o(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
});
d_gs_ms_ns.ForEach([&](auto& self, auto idx) {
d_g_m_n(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
});
z_gs_ms_ns.ForEach([&](auto& self, auto idx) {
z_g_m_n(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
});
// gemm 0
auto ref_gemm0 = ReferenceGemm0Instance{};
auto ref_gemm0_invoker = ref_gemm0.MakeInvoker();
auto ref_gemm0_argument = ref_gemm0.MakeArgument(
a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, acc0_element_op);
ref_gemm0_invoker.Run(ref_gemm0_argument);
// bias
acc0_g_m_n.ForEach([&](auto& self, auto idx) { self(idx) += d_g_m_n(idx); });
// masking
const auto mask = DeviceGemmInstance::C0MatrixMask(M, N);
acc0_g_m_n.ForEach([&](auto& self, auto idx) {
if(mask.IsMaskedElement(idx[1], idx[2]))
self(idx) = -ck::NumericLimits<float>::Infinity();
});
// softmax
auto ref_softmax = ReferenceSoftmaxInstance{};
auto ref_softmax_invoker = ref_softmax.MakeInvoker();
auto ref_softmax_argument =
ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2}, &lse_g_m_host_result);
ref_softmax_invoker.Run(ref_softmax_argument);
// dropout after softmax
auto ref_dropout = ReferenceDropoutInstance{};
auto ref_dropout_invoker = ref_dropout.MakeInvoker();
auto ref_dropout_argment = ref_dropout.MakeArgument(
z_g_m_n, a1_g_m_n, a1_g_m_n_drop, p_dropout_in_16bits, rp_dropout);
ref_dropout_invoker.Run(ref_dropout_argment);
// gemm1
auto ref_gemm1 = ReferenceGemm1Instance{};
auto ref_gemm1_invoker = ref_gemm1.MakeInvoker();
auto ref_gemm1_argument = ref_gemm1.MakeArgument(a1_g_m_n_drop,
b1_g_n_o,
c_g_m_o_host_result,
PassThrough{},
b1_element_op,
c_element_op);
ref_gemm1_invoker.Run(ref_gemm1_argument);
// permute
c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
const size_t& g0 = idx[0];
const size_t& g1 = idx[1];
const size_t g = g0 * G1 + g1;
self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
});
lse_gs_ms_host_result.ForEach([&](auto& self, auto idx) {
const size_t& g0 = idx[0];
const size_t& g1 = idx[1];
const size_t g = g0 * G1 + g1;
self(idx) = lse_g_m_host_result(g, idx[2]);
});
// default absolute error and relative error is 0.001
double rtol = 1e-3;
double atol = 1e-3;
// when BF16 is taken, set absolute error and relative error to 0.01
if(std::is_same_v<DataType, ck::bhalf_t> || std::is_same_v<GemmDataType, ck::bhalf_t>)
{
rtol = 1e-2;
atol = 1e-2;
}
return ck::utils::check_err(c_gs_ms_os_device_result.mData,
c_gs_ms_os_host_result.mData,
"Error: Incorrect results c!",
rtol,
atol) &&
ck::utils::check_err(lse_gs_ms_device_result.mData,
lse_gs_ms_host_result.mData,
"Error: Incorrect results lse!",
rtol,
atol)
? 0
: 1;
}
return 0;
}
example/52_flash_atten_bias/run_grouped_multihead_attention_bias_forward.inc 0 → 100644
View file @ f752739c
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2023, Advanced Micro Devices, Inc. All rights reserved.
int run(int argc, char* argv[])
{
bool do_verification = true;
int init_method = 1;
bool time_kernel = true;
bool input_permute = false;
bool output_permute = true;
float p_drop = 0.2;
const unsigned long long seed = 1;
const unsigned long long offset = 0;
if(argc == 1)
{
// use default case
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 7)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
p_drop = std::stoi(argv[4]);
input_permute = std::stoi(argv[5]);
output_permute = std::stoi(argv[6]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 5: input / output permute\n");
exit(0);
}
float p_dropout = 1 - p_drop;
uint16_t p_dropout_in_16bits = uint16_t(std::floor(p_dropout * 65535.0));
float rp_dropout = 1.0 / p_dropout;
float alpha = 1; // scaling after 1st gemm
std::size_t group_count = 8;
// Problem descs
std::vector<DeviceGemmInstance::ProblemDesc> problem_descs;
std::vector<const void*> p_a;
std::vector<const void*> p_b0;
std::vector<const void*> p_b1;
std::vector<void*> p_c;
std::vector<const void*> p_d;
std::vector<void*> p_z; // for result verification
std::vector<void*> p_z_nullptr; // for time test
std::vector<void*> p_lse;
std::vector<std::vector<int>> g0_g1_m_n_k_o;
std::vector<Tensor<ADataType>> a_tensors;
std::vector<Tensor<B0DataType>> b0_tensors;
std::vector<Tensor<B1DataType>> b1_tensors;
std::vector<Tensor<CDataType>> c_tensors;
std::vector<Tensor<DDataType>> d_tensors;
std::vector<Tensor<ZDataType>> z_tensors;
std::vector<Tensor<LSEDataType>> lse_tensors;
using DeviceMemPtr = std::unique_ptr<DeviceMem>;
std::vector<DeviceMemPtr> a_tensors_device;
std::vector<DeviceMemPtr> b0_tensors_device;
std::vector<DeviceMemPtr> b1_tensors_device;
std::vector<DeviceMemPtr> c_tensors_device;
std::vector<DeviceMemPtr> d_tensors_device;
std::vector<DeviceMemPtr> z_tensors_device;
std::vector<DeviceMemPtr> lse_tensors_device;
std::size_t flop = 0, num_byte = 0;
// std::cout << "group count " << group_count << ". printing first 4 groups\n";
for(std::size_t i = 0; i < group_count; i++)
{
int M = 128 * (rand() % 8) + (rand() % 128);
int N = 128 * (rand() % 8) + (rand() % 128);
int K = DIM;
int O = DIM;
int G0 = rand() % 3 + 1;
int G1 = rand() % 5 + 1;
g0_g1_m_n_k_o.push_back({G0, G1, M, N, K, O});
std::vector<ck::index_t> a_gs_ms_ks_lengths{G0, G1, M, K};
std::vector<ck::index_t> a_gs_ms_ks_strides =
input_permute
? std::vector<ck::index_t>{M * G1 * K, K, G1 * K, 1} // A layout [G0, M, G1, K]
: std::vector<ck::index_t>{G1 * M * K, M * K, K, 1}; // A layout [G0, G1, M, K]
std::vector<ck::index_t> b0_gs_ns_ks_lengths{G0, G1, N, K};
std::vector<ck::index_t> b0_gs_ns_ks_strides =
input_permute
? std::vector<ck::index_t>{N * G1 * K, K, G1 * K, 1} // B0 layout [G0, N, G1, K]
: std::vector<ck::index_t>{G1 * N * K, N * K, K, 1}; // B0 layout [G0, G1, N, K]
std::vector<ck::index_t> b1_gs_os_ns_lengths{G0, G1, O, N};
std::vector<ck::index_t> b1_gs_os_ns_strides =
input_permute
? std::vector<ck::index_t>{N * G1 * O, O, 1, G1 * O} // B1 layout [G0, N, G1, O]
: std::vector<ck::index_t>{G1 * N * O, N * O, 1, O}; // B1 layout [G0, G1, N, O]
std::vector<ck::index_t> c_gs_ms_os_lengths{G0, G1, M, O};
std::vector<ck::index_t> c_gs_ms_os_strides =
output_permute
? std::vector<ck::index_t>{M * G1 * O, O, G1 * O, 1} // C layout [G0, M, G1, O]
: std::vector<ck::index_t>{G1 * M * O, M * O, O, 1}; // C layout [G0, G1, M, O]
std::vector<ck::index_t> d_gs_ms_ns_lengths{G0, G1, M, N};
std::vector<ck::index_t> d_gs_ms_ns_strides =
input_permute
? std::vector<ck::index_t>{M * G1 * N, N, G1 * N, 1} // D layout [G0, M, G1, N]
: std::vector<ck::index_t>{G1 * M * N, M * N, N, 1}; // D layout [G0, G1, M, N]
std::vector<ck::index_t> z_gs_ms_ns_lengths{G0, G1, M, N};
std::vector<ck::index_t> z_gs_ms_ns_strides =
input_permute
? std::vector<ck::index_t>{M * G1 * N, N, G1 * N, 1} // Z layout [G0, M, G1, N]
: std::vector<ck::index_t>{G1 * M * N, M * N, N, 1}; // Z layout [G0, G1, M, N]
std::vector<ck::index_t> lse_gs_ms_lengths{G0, G1, M};
std::vector<ck::index_t> lse_gs_ms_strides =
std::vector<ck::index_t>{G1 * M, M, 1}; // LSE layout [G0, G1, M]
problem_descs.push_back({a_gs_ms_ks_lengths,
a_gs_ms_ks_strides,
b0_gs_ns_ks_lengths,
b0_gs_ns_ks_strides,
b1_gs_os_ns_lengths,
b1_gs_os_ns_strides,
c_gs_ms_os_lengths,
c_gs_ms_os_strides,
z_gs_ms_ns_lengths,
z_gs_ms_ns_strides,
lse_gs_ms_lengths,
lse_gs_ms_strides,
d_gs_ms_ns_lengths, // acc0_biases_gs_ms_ns_lengths
d_gs_ms_ns_strides, // acc0_biases_gs_ms_ns_strides
{}, // acc1_biases_gs_ms_os_lengths
{}}); // acc1_biases_gs_ms_os_strides
// C_m_o = A_m_k * B0_k_n * B1_n_o
Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
Tensor<B0DataType> b0_gs_ns_ks(b0_gs_ns_ks_lengths, b0_gs_ns_ks_strides);
Tensor<B1DataType> b1_gs_os_ns(b1_gs_os_ns_lengths, b1_gs_os_ns_strides);
Tensor<CDataType> c_gs_ms_os_device_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
Tensor<DDataType> d_gs_ms_ns(d_gs_ms_ns_lengths, d_gs_ms_ns_strides);
Tensor<ZDataType> z_gs_ms_ns(z_gs_ms_ns_lengths, z_gs_ms_ns_strides);
Tensor<LSEDataType> lse_gs_ms_device_result(lse_gs_ms_lengths, lse_gs_ms_strides);
int Batch = G0 * G1;
flop += (size_t(M) * N * K * 2 + size_t(M) * N * O * 2) * Batch;
num_byte += (sizeof(ADataType) * M * K + sizeof(B0DataType) * K * N +
sizeof(B1DataType) * N * O + sizeof(CDataType) * M * O +
sizeof(DDataType) * M * N * (std::is_void<Acc0BiasDataType>::value ? 0 : 1)) *
Batch;
if(i < 4)
{
std::cout << "a_gs_ms_ks[" << i << "]: " << a_gs_ms_ks.mDesc << ", "
<< "b0_gs_ns_ks[" << i << "]: " << b0_gs_ns_ks.mDesc << ", "
<< "b1_gs_os_ns[" << i << "]: " << b1_gs_os_ns.mDesc << ", "
<< "c_gs_ms_os[" << i << "]: " << c_gs_ms_os_device_result.mDesc << ", "
<< "d_gs_ms_ns[" << i << "]: " << d_gs_ms_ns.mDesc << ", "
<< "z_gs_ms_ns[" << i << "]: " << z_gs_ms_ns.mDesc << ", "
<< "lse_gs_ms_os[" << i << "]: " << lse_gs_ms_device_result.mDesc
<< std::endl;
}
z_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<ZDataType>{0});
switch(init_method)
{
case 0: break;
case 1:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_2<B0DataType>{-2, 2});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_2<B1DataType>{-2, 2});
d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_2<DDataType>{-1, 1});
break;
case 2:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_3<B0DataType>{0.0, 1.0});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_3<B1DataType>{-0.5, 0.5});
d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_3<DDataType>{-0.5, 0.5});
break;
case 3:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_2<ADataType>{-2, 2});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Diagonal<B0DataType>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<DDataType>{1});
break;
default:
a_gs_ms_ks.GenerateTensorValue(GeneratorTensor_1<ADataType>{1});
b0_gs_ns_ks.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
b1_gs_os_ns.GenerateTensorValue(GeneratorTensor_Diagonal<B1DataType>{});
d_gs_ms_ns.GenerateTensorValue(GeneratorTensor_1<DDataType>{1});
}
a_tensors.push_back(a_gs_ms_ks);
b0_tensors.push_back(b0_gs_ns_ks);
b1_tensors.push_back(b1_gs_os_ns);
c_tensors.push_back(c_gs_ms_os_device_result);
d_tensors.push_back(d_gs_ms_ns);
z_tensors.push_back(z_gs_ms_ns);
lse_tensors.push_back(lse_gs_ms_device_result);
a_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(ADataType) * a_gs_ms_ks.mDesc.GetElementSpaceSize()));
b0_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(B0DataType) * b0_gs_ns_ks.mDesc.GetElementSpaceSize()));
b1_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(B1DataType) * b1_gs_os_ns.mDesc.GetElementSpaceSize()));
c_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(CDataType) * c_gs_ms_os_device_result.mDesc.GetElementSpaceSize()));
d_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(DDataType) * d_gs_ms_ns.mDesc.GetElementSpaceSize()));
z_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(ZDataType) * z_gs_ms_ns.mDesc.GetElementSpaceSize()));
lse_tensors_device.emplace_back(std::make_unique<DeviceMem>(
sizeof(LSEDataType) * lse_gs_ms_device_result.mDesc.GetElementSpaceSize()));
a_tensors_device[i]->ToDevice(a_gs_ms_ks.mData.data());
b0_tensors_device[i]->ToDevice(b0_gs_ns_ks.mData.data());
b1_tensors_device[i]->ToDevice(b1_gs_os_ns.mData.data());
d_tensors_device[i]->ToDevice(d_gs_ms_ns.mData.data());
p_a.push_back(a_tensors_device[i]->GetDeviceBuffer());
p_b0.push_back(b0_tensors_device[i]->GetDeviceBuffer());
p_b1.push_back(b1_tensors_device[i]->GetDeviceBuffer());
p_c.push_back(c_tensors_device[i]->GetDeviceBuffer());
p_d.push_back(d_tensors_device[i]->GetDeviceBuffer());
p_z.push_back(z_tensors_device[i]->GetDeviceBuffer());
p_z_nullptr.push_back(nullptr);
p_lse.push_back(lse_tensors_device[i]->GetDeviceBuffer());
}
auto a_element_op = AElementOp{};
auto b0_element_op = B0ElementOp{};
auto acc0_element_op = Acc0ElementOp{alpha};
auto b1_element_op = B1ElementOp{};
auto c_element_op = CElementOp{};
// do GEMM
auto gemm = DeviceGemmInstance{};
auto invoker = gemm.MakeInvoker();
auto argument =
gemm.MakeArgument(p_a,
p_b0,
p_b1,
p_c,
p_z_nullptr,
p_lse,
p_d, // p_acc0_biases
{}, // p_acc1_biases
problem_descs,
a_element_op,
b0_element_op,
acc0_element_op,
b1_element_op,
c_element_op,
p_drop, // dropout ratio
{seed, offset}); // dropout random seed and offset, offset should be
// at least the number of elements on a thread
// specify workspace for problem_desc
DeviceMem problem_desc_workspace(gemm.GetWorkSpaceSize(&argument));
gemm.SetWorkSpacePointer(&argument, problem_desc_workspace.GetDeviceBuffer());
if(!gemm.IsSupportedArgument(argument))
{
std::cout << gemm.GetTypeString() << " does not support this problem" << std::endl;
return 0;
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_byte / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< gemm.GetTypeString() << std::endl;
bool pass = true;
if(do_verification)
{
argument =
gemm.MakeArgument(p_a,
p_b0,
p_b1,
p_c,
p_z,
p_lse,
p_d, // p_acc0_biases
{}, // p_acc1_biases
problem_descs,
a_element_op,
b0_element_op,
acc0_element_op,
b1_element_op,
c_element_op,
p_drop, // dropout ratio
{seed, offset}); // dropout random seed and offset, offset should be
// at least the number of elements on a thread
// specify workspace for problem_desc
DeviceMem problem_desc_workspace_verify(gemm.GetWorkSpaceSize(&argument));
gemm.SetWorkSpacePointer(&argument, problem_desc_workspace_verify.GetDeviceBuffer());
invoker.Run(argument, StreamConfig{nullptr, false});
for(std::size_t i = 0; i < group_count; i++)
{
const int& G0 = g0_g1_m_n_k_o[i][0];
const int& G1 = g0_g1_m_n_k_o[i][1];
const int& M = g0_g1_m_n_k_o[i][2];
const int& N = g0_g1_m_n_k_o[i][3];
const int& K = g0_g1_m_n_k_o[i][4];
const int& O = g0_g1_m_n_k_o[i][5];
const auto& c_gs_ms_os_lengths = problem_descs[i].c_gs_ms_os_lengths;
const auto& c_gs_ms_os_strides = problem_descs[i].c_gs_ms_os_strides;
const auto& lse_gs_ms_lengths = problem_descs[i].lse_gs_ms_lengths;
const auto& lse_gs_ms_strides = problem_descs[i].lse_gs_ms_strides;
const auto& a_gs_ms_ks = a_tensors[i];
const auto& b0_gs_ns_ks = b0_tensors[i];
const auto& b1_gs_os_ns = b1_tensors[i];
const auto& d_gs_ms_ns = d_tensors[i];
auto& c_gs_ms_os_device_result = c_tensors[i];
auto& z_gs_ms_ns_device_result = z_tensors[i];
auto& lse_gs_ms_device_result = lse_tensors[i];
auto& c_gs_ms_os_device_buf = *c_tensors_device[i];
auto& z_gs_ms_ns_device_buf = *z_tensors_device[i];
auto& lse_gs_ms_device_buf = *lse_tensors_device[i];
c_gs_ms_os_device_buf.FromDevice(c_gs_ms_os_device_result.mData.data());
z_gs_ms_ns_device_buf.FromDevice(z_gs_ms_ns_device_result.mData.data());
lse_gs_ms_device_buf.FromDevice(lse_gs_ms_device_result.mData.data());
Tensor<ADataType> a_g_m_k({G0 * G1, M, K});
Tensor<B0DataType> b0_g_k_n({G0 * G1, K, N});
Tensor<B1DataType> b1_g_n_o({G0 * G1, N, O});
Tensor<AccDataType> acc0_g_m_n({G0 * G1, M, N}); // scratch object after gemm0
Tensor<AccDataType> d_g_m_n({G0 * G1, M, N});
Tensor<ADataType> a1_g_m_n({G0 * G1, M, N}); // scratch object after softmax
Tensor<ADataType> a1_g_m_n_drop({G0 * G1, M, N}); // scratch object after softmax
Tensor<CDataType> c_g_m_o_host_result({G0 * G1, M, O}); // scratch object after gemm1
Tensor<CDataType> c_gs_ms_os_host_result(c_gs_ms_os_lengths, c_gs_ms_os_strides);
Tensor<ZDataType> z_g_m_n({G0 * G1, M, N});
Tensor<LSEDataType> lse_g_m_host_result({G0 * G1, M}); // scratch object after gemm1
Tensor<LSEDataType> lse_gs_ms_host_result(lse_gs_ms_lengths, lse_gs_ms_strides);
// permute
a_gs_ms_ks.ForEach([&](auto& self, auto idx) {
a_g_m_k(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
});
b0_gs_ns_ks.ForEach([&](auto& self, auto idx) {
b0_g_k_n(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
});
b1_gs_os_ns.ForEach([&](auto& self, auto idx) {
b1_g_n_o(idx[0] * G1 + idx[1], idx[3], idx[2]) = self(idx);
});
d_gs_ms_ns.ForEach([&](auto& self, auto idx) {
d_g_m_n(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
});
z_gs_ms_ns_device_result.ForEach([&](auto& self, auto idx) {
z_g_m_n(idx[0] * G1 + idx[1], idx[2], idx[3]) = self(idx);
});
// gemm 0
auto ref_gemm0 = ReferenceGemm0Instance{};
auto ref_gemm0_invoker = ref_gemm0.MakeInvoker();
auto ref_gemm0_argument = ref_gemm0.MakeArgument(
a_g_m_k, b0_g_k_n, acc0_g_m_n, a_element_op, b0_element_op, acc0_element_op);
ref_gemm0_invoker.Run(ref_gemm0_argument);
// bias
acc0_g_m_n.ForEach([&](auto& self, auto idx) { self(idx) += d_g_m_n(idx); });
// masking
const auto mask = DeviceGemmInstance::C0MatrixMask(M, N);
acc0_g_m_n.ForEach([&](auto& self, auto idx) {
if(mask.IsMaskedElement(idx[1], idx[2]))
self(idx) = -ck::NumericLimits<float>::Infinity();
});
// softmax
auto ref_softmax = ReferenceSoftmaxInstance{};
auto ref_softmax_invoker = ref_softmax.MakeInvoker();
auto ref_softmax_argument =
ref_softmax.MakeArgument(acc0_g_m_n, a1_g_m_n, 1, 0, {2}, &lse_g_m_host_result);
ref_softmax_invoker.Run(ref_softmax_argument);
// printf("print z_g_m_n \n");
// z_g_m_n.ForEach([&](auto& self, auto idx) {printf("%u ", self(idx));});
// dropout after softmax
auto ref_dropout = ReferenceDropoutInstance{};
auto ref_dropout_invoker = ref_dropout.MakeInvoker();
auto ref_dropout_argment = ref_dropout.MakeArgument(
z_g_m_n, a1_g_m_n, a1_g_m_n_drop, p_dropout_in_16bits, rp_dropout);
ref_dropout_invoker.Run(ref_dropout_argment);
// gemm 1
auto ref_gemm1 = ReferenceGemm1Instance{};
auto ref_gemm1_invoker = ref_gemm1.MakeInvoker();
auto ref_gemm1_argument = ref_gemm1.MakeArgument(a1_g_m_n_drop,
b1_g_n_o,
c_g_m_o_host_result,
PassThrough{},
b1_element_op,
c_element_op);
ref_gemm1_invoker.Run(ref_gemm1_argument);
// permute
c_gs_ms_os_host_result.ForEach([&](auto& self, auto idx) {
const size_t& g0 = idx[0];
const size_t& g1 = idx[1];
const size_t g = g0 * G1 + g1;
self(idx) = c_g_m_o_host_result(g, idx[2], idx[3]);
});
lse_gs_ms_host_result.ForEach([&](auto& self, auto idx) {
const size_t& g0 = idx[0];
const size_t& g1 = idx[1];
const size_t g = g0 * G1 + g1;
self(idx) = lse_g_m_host_result(g, idx[2]);
});
// default absolute error and relative error is 0.001
double rtol = 1e-3;
double atol = 1e-3;
// when BF16 is taken, set absolute error and relative error to 0.01
if(std::is_same_v<DataType, ck::bhalf_t> || std::is_same_v<GemmDataType, ck::bhalf_t>)
{
rtol = 1e-2;
atol = 1e-2;
}
// bool pass_ =
// ck::utils::check_err(c_gs_ms_os_device_result.mData,
// c_gs_ms_os_host_result.mData);
bool pass_ = ck::utils::check_err(c_gs_ms_os_device_result.mData,
c_gs_ms_os_host_result.mData,
"Error: Incorrect results c!",
rtol,
atol) &&
ck::utils::check_err(lse_gs_ms_device_result.mData,
lse_gs_ms_host_result.mData,
"Error: Incorrect results lse!",
rtol,
atol);
if(!pass_)
{
std::cout << "from group: " << i << std::endl;
}
pass &= pass_;
}
if(pass)
{
std::cout << "Verification passed." << std::endl;
}
}
return pass ? 0 : 1;
}
include/ck/tensor_operation/gpu/block/blockwise_dropout.hpp
View file @ f752739c
...@@ -138,12 +138,12 @@ struct BlockwiseDropout ...@@ -138,12 +138,12 @@ struct BlockwiseDropout
constexpr int tmp_size = MRepeat * KRepeat; constexpr int tmp_size = MRepeat * KRepeat;
int philox_calls = tmp_size / 4; int philox_calls = tmp_size / 8;
ushort tmp[tmp_size]; ushort tmp[tmp_size];
for(int i = 0; i < philox_calls; i++) for(int i = 0; i < philox_calls; i++)
{ {
ph.get_random_4x16((tmp + i * 4), element_global_1d_id + i * Offset{} * MRaw); ph.get_random_8x16((tmp + i * 8), element_global_1d_id + i * Offset{} * MRaw);
} }
block_sync_lds(); block_sync_lds();
...@@ -179,12 +179,12 @@ struct BlockwiseDropout ...@@ -179,12 +179,12 @@ struct BlockwiseDropout
constexpr int tmp_size = MRepeat * KRepeat; constexpr int tmp_size = MRepeat * KRepeat;
int philox_calls = tmp_size / 4; int philox_calls = tmp_size / 8;
ushort tmp[tmp_size]; ushort tmp[tmp_size];
for(int i = 0; i < philox_calls; i++) for(int i = 0; i < philox_calls; i++)
{ {
ph.get_random_4x16((tmp + i * 4), element_global_1d_id + i * Offset{} * MRaw); ph.get_random_8x16((tmp + i * 8), element_global_1d_id + i * Offset{} * MRaw);
} }
block_sync_lds(); block_sync_lds();
...@@ -218,21 +218,19 @@ struct BlockwiseDropout ...@@ -218,21 +218,19 @@ struct BlockwiseDropout
} }
// get raw z matrix with random number for shuffle // get raw z matrix with random number for shuffle
template <typename ZThreadBuffer, template <typename ZThreadBuffer, typename Step, typename Offset>
typename Step,
typename Offset> // N3*N4=8
__host__ __device__ void GenerateZMatrixAttnFwd(ck::philox& ph, __host__ __device__ void GenerateZMatrixAttnFwd(ck::philox& ph,
index_t element_global_1d_id, index_t element_global_1d_id,
ZThreadBuffer& z_thread_buf) ZThreadBuffer& z_thread_buf)
{ {
constexpr int tmp_size = MRepeat * KRepeat / Step{}.value; constexpr int tmp_size = MRepeat * KRepeat / Step{}.value;
int philox_calls = tmp_size / 4; int philox_calls = tmp_size / 8;
ushort tmp[tmp_size]; ushort tmp[tmp_size];
for(int i = 0; i < philox_calls; i++) for(int i = 0; i < philox_calls; i++)
{ {
ph.get_random_4x16((tmp + i * 4), element_global_1d_id + i * Offset{}); ph.get_random_8x16((tmp + i * 8), element_global_1d_id + i * Offset{});
} }
static_for<0, tmp_size, 1>{}([&](auto i) { z_thread_buf(i) = tmp[i.value]; }); static_for<0, tmp_size, 1>{}([&](auto i) { z_thread_buf(i) = tmp[i.value]; });
......
include/ck/tensor_operation/gpu/device/device_batched_gemm_softmax_gemm_permute.hpp
View file @ f752739c
...@@ -87,9 +87,6 @@ template <index_t NumDimG, ...@@ -87,9 +87,6 @@ template <index_t NumDimG,
MaskingSpecialization MaskingSpec> MaskingSpecialization MaskingSpec>
struct DeviceBatchedMultiheadAttentionForward : public BaseOperator struct DeviceBatchedMultiheadAttentionForward : public BaseOperator
{ {
static constexpr index_t NumAcc0Bias = Acc0BiasDataType::Size();
static constexpr index_t NumAcc1Bias = Acc1BiasDataType::Size();
virtual std::unique_ptr<BaseArgument> MakeArgumentPointer( virtual std::unique_ptr<BaseArgument> MakeArgumentPointer(
const void* p_a, const void* p_a,
const void* p_b0, const void* p_b0,
...@@ -97,8 +94,8 @@ struct DeviceBatchedMultiheadAttentionForward : public BaseOperator ...@@ -97,8 +94,8 @@ struct DeviceBatchedMultiheadAttentionForward : public BaseOperator
void* p_c, void* p_c,
void* p_z, void* p_z,
void* p_lse, void* p_lse,
const std::array<void*, NumAcc0Bias> p_acc0_biases, const void* p_acc0_biases,
const std::array<void*, NumAcc1Bias> p_acc1_biases, const void* p_acc1_biases,
const std::vector<index_t>& a_gs_ms_ks_lengths, const std::vector<index_t>& a_gs_ms_ks_lengths,
const std::vector<index_t>& a_gs_ms_ks_strides, const std::vector<index_t>& a_gs_ms_ks_strides,
const std::vector<index_t>& b_gs_ns_ks_lengths, const std::vector<index_t>& b_gs_ns_ks_lengths,
...@@ -110,11 +107,11 @@ struct DeviceBatchedMultiheadAttentionForward : public BaseOperator ...@@ -110,11 +107,11 @@ struct DeviceBatchedMultiheadAttentionForward : public BaseOperator
const std::vector<index_t>& z_gs_ms_ns_lengths, // z_gs_ms_os_lengths const std::vector<index_t>& z_gs_ms_ns_lengths, // z_gs_ms_os_lengths
const std::vector<index_t>& z_gs_ms_ns_strides, // z_gs_ms_os_strides const std::vector<index_t>& z_gs_ms_ns_strides, // z_gs_ms_os_strides
const std::vector<index_t>& lse_gs_ms_lengths, // lse_gs_ms_lengths const std::vector<index_t>& lse_gs_ms_lengths, // lse_gs_ms_lengths
const std::array<std::vector<index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_lengths, const std::vector<index_t>& acc0_biases_gs_ms_ns_lengths,
const std::array<std::vector<index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_strides, const std::vector<index_t>& acc0_biases_gs_ms_ns_strides,
const std::array<std::vector<index_t>, NumAcc1Bias> const std::vector<index_t>&
acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths
const std::array<std::vector<index_t>, NumAcc1Bias> const std::vector<index_t>&
acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
AElementwiseOperation a_element_op, AElementwiseOperation a_element_op,
B0ElementwiseOperation b0_element_op, B0ElementwiseOperation b0_element_op,
......
include/ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute.hpp
View file @ f752739c
...@@ -111,11 +111,11 @@ struct DeviceGroupedMultiheadAttentionForward : public BaseOperator ...@@ -111,11 +111,11 @@ struct DeviceGroupedMultiheadAttentionForward : public BaseOperator
std::vector<index_t> lse_gs_ms_lengths; std::vector<index_t> lse_gs_ms_lengths;
std::vector<index_t> lse_gs_ms_strides; std::vector<index_t> lse_gs_ms_strides;
std::vector<std::vector<index_t>> acc0_biases_gs_ms_ns_lengths; std::vector<index_t> acc0_biases_gs_ms_ns_lengths;
std::vector<std::vector<index_t>> acc0_biases_gs_ms_ns_strides; std::vector<index_t> acc0_biases_gs_ms_ns_strides;
std::vector<std::vector<index_t>> acc1_biases_gs_ms_os_lengths; std::vector<index_t> acc1_biases_gs_ms_os_lengths;
std::vector<std::vector<index_t>> acc1_biases_gs_ms_os_strides; std::vector<index_t> acc1_biases_gs_ms_os_strides;
}; };
virtual std::unique_ptr<BaseArgument> virtual std::unique_ptr<BaseArgument>
...@@ -125,9 +125,9 @@ struct DeviceGroupedMultiheadAttentionForward : public BaseOperator ...@@ -125,9 +125,9 @@ struct DeviceGroupedMultiheadAttentionForward : public BaseOperator
std::vector<void*> p_c_vec, std::vector<void*> p_c_vec,
std::vector<void*> p_z_vec, std::vector<void*> p_z_vec,
std::vector<void*> p_lse_vec, std::vector<void*> p_lse_vec,
std::vector<std::vector<const void*>> p_acc0_biases_vec, std::vector<const void*> p_acc0_biases_vec,
std::vector<std::vector<const void*>> p_acc1_biases_vec, std::vector<const void*> p_acc1_biases_vec,
std::vector<ProblemDesc> problem_desc_vec, std::vector<ProblemDesc>& problem_desc_vec,
AElementwiseOperation a_element_op, AElementwiseOperation a_element_op,
B0ElementwiseOperation b0_element_op, B0ElementwiseOperation b0_element_op,
Acc0ElementwiseOperation acc0_element_op, Acc0ElementwiseOperation acc0_element_op,
......
include/ck/tensor_operation/gpu/device/impl/device_batched_mha_bwd_xdl_cshuffle_kloop_v1.hpp
View file @ f752739c
...@@ -21,8 +21,6 @@ ...@@ -21,8 +21,6 @@
#include "ck/host_utility/device_prop.hpp" #include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp" #include "ck/host_utility/kernel_launch.hpp"
#include "ck/library/utility/host_tensor.hpp"
namespace ck { namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
......
include/ck/tensor_operation/gpu/device/impl/device_batched_mha_bwd_xdl_cshuffle_kloop_v2.hpp
View file @ f752739c
...@@ -20,8 +20,6 @@ ...@@ -20,8 +20,6 @@
#include "ck/host_utility/device_prop.hpp" #include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp" #include "ck/host_utility/kernel_launch.hpp"
#include "ck/library/utility/host_tensor.hpp"
namespace ck { namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
......
include/ck/tensor_operation/gpu/device/impl/device_batched_mha_bwd_xdl_cshuffle_qloop_light_v1.hpp
View file @ f752739c
...@@ -10,7 +10,6 @@ ...@@ -10,7 +10,6 @@
#include "ck/utility/philox_rand.hpp" #include "ck/utility/philox_rand.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp" #include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp" #include "ck/tensor_description/tensor_descriptor_helper.hpp"
// #include "ck/tensor_operation/gpu/device/device_batched_multihead_attention_backward.hpp" // TODO
#include "ck/tensor_operation/gpu/device/device_base.hpp" #include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/masking_specialization.hpp" #include "ck/tensor_operation/gpu/device/masking_specialization.hpp"
...@@ -22,8 +21,6 @@ ...@@ -22,8 +21,6 @@
#include "ck/host_utility/device_prop.hpp" #include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp" #include "ck/host_utility/kernel_launch.hpp"
#include "ck/library/utility/host_tensor.hpp"
namespace ck { namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
......
include/ck/tensor_operation/gpu/device/impl/device_batched_mha_bwd_xdl_cshuffle_qloop_light_v2.hpp
View file @ f752739c
...@@ -21,8 +21,6 @@ ...@@ -21,8 +21,6 @@
#include "ck/host_utility/device_prop.hpp" #include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp" #include "ck/host_utility/kernel_launch.hpp"
#include "ck/library/utility/host_tensor.hpp"
namespace ck { namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
......
include/ck/tensor_operation/gpu/device/impl/device_batched_mha_bwd_xdl_cshuffle_qloop_v1.hpp
View file @ f752739c
...@@ -10,7 +10,6 @@ ...@@ -10,7 +10,6 @@
#include "ck/utility/philox_rand.hpp" #include "ck/utility/philox_rand.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp" #include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp" #include "ck/tensor_description/tensor_descriptor_helper.hpp"
// #include "ck/tensor_operation/gpu/device/device_batched_multihead_attention_backward.hpp" // TODO
#include "ck/tensor_operation/gpu/device/device_base.hpp" #include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/masking_specialization.hpp" #include "ck/tensor_operation/gpu/device/masking_specialization.hpp"
...@@ -21,8 +20,6 @@ ...@@ -21,8 +20,6 @@
#include "ck/host_utility/device_prop.hpp" #include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp" #include "ck/host_utility/kernel_launch.hpp"
#include "ck/library/utility/host_tensor.hpp"
namespace ck { namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
......
include/ck/tensor_operation/gpu/device/impl/device_batched_mha_bwd_xdl_cshuffle_qloop_v2.hpp
View file @ f752739c
...@@ -20,8 +20,6 @@ ...@@ -20,8 +20,6 @@
#include "ck/host_utility/device_prop.hpp" #include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp" #include "ck/host_utility/kernel_launch.hpp"
#include "ck/library/utility/host_tensor.hpp"
namespace ck { namespace ck {
namespace tensor_operation { namespace tensor_operation {
namespace device { namespace device {
......
include/ck/tensor_operation/gpu/device/impl/device_batched_mha_fwd_xdl_cshuffle_v2.hpp
View file @ f752739c
...@@ -25,6 +25,7 @@ namespace device { ...@@ -25,6 +25,7 @@ namespace device {
template <typename GridwiseGemm, template <typename GridwiseGemm,
typename FloatAB, typename FloatAB,
typename D0DataType,
typename FloatC, typename FloatC,
typename ZDataType, typename ZDataType,
typename FloatLSE, typename FloatLSE,
...@@ -36,9 +37,10 @@ template <typename GridwiseGemm, ...@@ -36,9 +37,10 @@ template <typename GridwiseGemm,
typename CElementwiseOperation, typename CElementwiseOperation,
typename AGridDesc_AK0_M_AK1, typename AGridDesc_AK0_M_AK1,
typename BGridDesc_BK0_N_BK1, typename BGridDesc_BK0_N_BK1,
typename D0GridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5,
typename B1GridDesc_BK0_N_BK1, typename B1GridDesc_BK0_N_BK1,
typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock, typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
typename ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5, typename ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_M4_N4_N5_N6,
typename LSEGridDescriptor_M, typename LSEGridDescriptor_M,
typename Block2CTileMap, typename Block2CTileMap,
typename ComputeBasePtrOfStridedBatch, typename ComputeBasePtrOfStridedBatch,
...@@ -54,6 +56,7 @@ __global__ void ...@@ -54,6 +56,7 @@ __global__ void
kernel_batched_multiheadattention_forward_xdl_cshuffle_v2( kernel_batched_multiheadattention_forward_xdl_cshuffle_v2(
const FloatAB* __restrict__ p_a_grid, const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid, const FloatAB* __restrict__ p_b_grid,
const D0DataType* __restrict__ p_d0_grid,
const FloatAB* __restrict__ p_b1_grid, const FloatAB* __restrict__ p_b1_grid,
FloatC* __restrict__ p_c_grid, FloatC* __restrict__ p_c_grid,
ZDataType* __restrict__ p_z_grid, ZDataType* __restrict__ p_z_grid,
...@@ -65,11 +68,13 @@ __global__ void ...@@ -65,11 +68,13 @@ __global__ void
const CElementwiseOperation c_element_op, const CElementwiseOperation c_element_op,
const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1, const AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1,
const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1, const BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1,
const D0GridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5
d0_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
const B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1, const B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1,
const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
c_grid_desc_mblock_mperblock_nblock_nperblock, c_grid_desc_mblock_mperblock_nblock_nperblock,
const ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5 const ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_M4_N4_N5_N6
z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5, z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_m4_n4_n5_n6,
const LSEGridDescriptor_M lse_grid_desc_m, const LSEGridDescriptor_M lse_grid_desc_m,
const Block2CTileMap block_2_ctile_map, const Block2CTileMap block_2_ctile_map,
const index_t batch_count, const index_t batch_count,
...@@ -102,6 +107,15 @@ __global__ void ...@@ -102,6 +107,15 @@ __global__ void
static_cast<long_index_t>(compute_base_ptr_of_batch.GetZBasePtr(g_idx))); static_cast<long_index_t>(compute_base_ptr_of_batch.GetZBasePtr(g_idx)));
const long_index_t lse_batch_offset = __builtin_amdgcn_readfirstlane( const long_index_t lse_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_base_ptr_of_batch.GetLSEBasePtr(g_idx))); static_cast<long_index_t>(compute_base_ptr_of_batch.GetLSEBasePtr(g_idx)));
const long_index_t d0_batch_offset = __builtin_amdgcn_readfirstlane(
static_cast<long_index_t>(compute_base_ptr_of_batch.GetD0BasePtr(g_idx)));
const D0DataType* tmp_p_d0_grid = nullptr;
if constexpr(!is_same<D0DataType, void>::value)
{
tmp_p_d0_grid = p_d0_grid + d0_batch_offset;
}
// const index_t global_thread_id = get_thread_global_1d_id(); // const index_t global_thread_id = get_thread_global_1d_id();
ck::philox ph(seed, 0, offset); ck::philox ph(seed, 0, offset);
...@@ -115,6 +129,7 @@ __global__ void ...@@ -115,6 +129,7 @@ __global__ void
GridwiseGemm::template Run<HasMainKBlockLoop, IsDropout, IsLseStoring>( GridwiseGemm::template Run<HasMainKBlockLoop, IsDropout, IsLseStoring>(
p_a_grid + a_batch_offset, p_a_grid + a_batch_offset,
p_b_grid + b_batch_offset, p_b_grid + b_batch_offset,
tmp_p_d0_grid,
p_b1_grid + b1_batch_offset, p_b1_grid + b1_batch_offset,
p_c_grid + c_batch_offset, p_c_grid + c_batch_offset,
p_z_grid == nullptr ? nullptr : p_z_grid + z_batch_offset, p_z_grid == nullptr ? nullptr : p_z_grid + z_batch_offset,
...@@ -127,9 +142,10 @@ __global__ void ...@@ -127,9 +142,10 @@ __global__ void
c_element_op, c_element_op,
a_grid_desc_ak0_m_ak1, a_grid_desc_ak0_m_ak1,
b_grid_desc_bk0_n_bk1, b_grid_desc_bk0_n_bk1,
d0_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
b1_grid_desc_bk0_n_bk1, b1_grid_desc_bk0_n_bk1,
c_grid_desc_mblock_mperblock_nblock_nperblock, c_grid_desc_mblock_mperblock_nblock_nperblock,
z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5, z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_m4_n4_n5_n6,
lse_grid_desc_m, lse_grid_desc_m,
block_2_ctile_map, block_2_ctile_map,
c0_matrix_mask, c0_matrix_mask,
...@@ -146,6 +162,7 @@ __global__ void ...@@ -146,6 +162,7 @@ __global__ void
GridwiseGemm::template Run<HasMainKBlockLoop, IsDropout, IsLseStoring>( GridwiseGemm::template Run<HasMainKBlockLoop, IsDropout, IsLseStoring>(
p_a_grid + a_batch_offset, p_a_grid + a_batch_offset,
p_b_grid + b_batch_offset, p_b_grid + b_batch_offset,
tmp_p_d0_grid,
p_b1_grid + b1_batch_offset, p_b1_grid + b1_batch_offset,
p_c_grid + c_batch_offset, p_c_grid + c_batch_offset,
p_z_grid == nullptr ? nullptr : p_z_grid + z_batch_offset, p_z_grid == nullptr ? nullptr : p_z_grid + z_batch_offset,
...@@ -158,9 +175,10 @@ __global__ void ...@@ -158,9 +175,10 @@ __global__ void
c_element_op, c_element_op,
a_grid_desc_ak0_m_ak1, a_grid_desc_ak0_m_ak1,
b_grid_desc_bk0_n_bk1, b_grid_desc_bk0_n_bk1,
d0_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
b1_grid_desc_bk0_n_bk1, b1_grid_desc_bk0_n_bk1,
c_grid_desc_mblock_mperblock_nblock_nperblock, c_grid_desc_mblock_mperblock_nblock_nperblock,
z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5, z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_m4_n4_n5_n6,
lse_grid_desc_m, lse_grid_desc_m,
block_2_ctile_map, block_2_ctile_map,
c0_matrix_mask, c0_matrix_mask,
...@@ -174,6 +192,7 @@ __global__ void ...@@ -174,6 +192,7 @@ __global__ void
#else #else
ignore = p_a_grid; ignore = p_a_grid;
ignore = p_b_grid; ignore = p_b_grid;
ignore = p_d0_grid;
ignore = p_b1_grid; ignore = p_b1_grid;
ignore = p_c_grid; ignore = p_c_grid;
ignore = p_z_grid; ignore = p_z_grid;
...@@ -185,9 +204,10 @@ __global__ void ...@@ -185,9 +204,10 @@ __global__ void
ignore = c_element_op; ignore = c_element_op;
ignore = a_grid_desc_ak0_m_ak1; ignore = a_grid_desc_ak0_m_ak1;
ignore = b_grid_desc_bk0_n_bk1; ignore = b_grid_desc_bk0_n_bk1;
ignore = d0_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5;
ignore = b1_grid_desc_bk0_n_bk1; ignore = b1_grid_desc_bk0_n_bk1;
ignore = c_grid_desc_mblock_mperblock_nblock_nperblock; ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
ignore = z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5; ignore = z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_m4_n4_n5_n6;
ignore = lse_grid_desc_m; ignore = lse_grid_desc_m;
ignore = block_2_ctile_map; ignore = block_2_ctile_map;
ignore = batch_count; ignore = batch_count;
...@@ -247,6 +267,7 @@ template <index_t NumDimG, ...@@ -247,6 +267,7 @@ template <index_t NumDimG,
index_t MXdlPerWave, index_t MXdlPerWave,
index_t NXdlPerWave, index_t NXdlPerWave,
index_t Gemm1NXdlPerWave, index_t Gemm1NXdlPerWave,
index_t DropoutStep,
typename ABlockTransferThreadClusterLengths_AK0_M_AK1, typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
typename ABlockTransferThreadClusterArrangeOrder, typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder, typename ABlockTransferSrcAccessOrder,
...@@ -261,6 +282,7 @@ template <index_t NumDimG, ...@@ -261,6 +282,7 @@ template <index_t NumDimG,
index_t BBlockTransferSrcScalarPerVector, index_t BBlockTransferSrcScalarPerVector,
index_t BBlockTransferDstScalarPerVector_BK1, index_t BBlockTransferDstScalarPerVector_BK1,
bool BBlockLdsExtraN, bool BBlockLdsExtraN,
index_t Acc0BiasTransferSrcScalarPerVector,
typename B1BlockTransferThreadClusterLengths_BK0_N_BK1, typename B1BlockTransferThreadClusterLengths_BK0_N_BK1,
typename B1BlockTransferThreadClusterArrangeOrder, typename B1BlockTransferThreadClusterArrangeOrder,
typename B1BlockTransferSrcAccessOrder, typename B1BlockTransferSrcAccessOrder,
...@@ -272,6 +294,7 @@ template <index_t NumDimG, ...@@ -272,6 +294,7 @@ template <index_t NumDimG,
index_t CShuffleNXdlPerWavePerShuffle, index_t CShuffleNXdlPerWavePerShuffle,
typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CShuffleBlockTransferScalarPerVector_NPerBlock, index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
index_t Acc1BiasTransferSrcScalarPerVector,
MaskingSpecialization MaskingSpec, MaskingSpecialization MaskingSpec,
bool Deterministic, bool Deterministic,
LoopScheduler LoopSched = LoopScheduler::Default> LoopScheduler LoopSched = LoopScheduler::Default>
...@@ -299,11 +322,10 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2 ...@@ -299,11 +322,10 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
static_assert(NumDimG > 0 && NumDimM > 0 && NumDimN > 0 && NumDimK > 0 && NumDimO > 0, static_assert(NumDimG > 0 && NumDimM > 0 && NumDimN > 0 && NumDimK > 0 && NumDimO > 0,
"Number of dimension must be greater than 0"); "Number of dimension must be greater than 0");
static constexpr index_t NumAcc0Bias = Acc0BiasDataType::Size();
static constexpr index_t NumAcc1Bias = Acc1BiasDataType::Size();
// TODO ANT: implement bias combination // TODO ANT: implement bias combination
static_assert(NumAcc0Bias == 0 && NumAcc0Bias == 0, "Bias addition is unimplemented"); static_assert(std::is_void<Acc1BiasDataType>::value, "Acc1 Bias addition is unimplemented");
using D0DataType = Acc0BiasDataType;
using D1DataType = Acc1BiasDataType;
#if 0 #if 0
// TODO ANT: use alias // TODO ANT: use alias
...@@ -389,12 +411,14 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2 ...@@ -389,12 +411,14 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
using AGridDesc_AK0_M_AK1 = decltype(MakeAGridDescriptor_AK0_M_AK1({}, {})); using AGridDesc_AK0_M_AK1 = decltype(MakeAGridDescriptor_AK0_M_AK1({}, {}));
using BGridDesc_BK0_N_BK1 = decltype(MakeBGridDescriptor_BK0_N_BK1({}, {})); using BGridDesc_BK0_N_BK1 = decltype(MakeBGridDescriptor_BK0_N_BK1({}, {}));
using D0GridDesc_M_N = decltype(Transform::MakeCGridDescriptor_M_N({}, {}));
using B1GridDesc_BK0_N_BK1 = decltype(MakeB1GridDescriptor_BK0_N_BK1({}, {})); using B1GridDesc_BK0_N_BK1 = decltype(MakeB1GridDescriptor_BK0_N_BK1({}, {}));
using CGridDesc_M_N = decltype(Transform::MakeCGridDescriptor_M_N({}, {})); using CGridDesc_M_N = decltype(Transform::MakeCGridDescriptor_M_N({}, {}));
using ZGridDesc_M_N = decltype(MakeZGridDescriptor_M_N({}, {})); using ZGridDesc_M_N = decltype(MakeZGridDescriptor_M_N({}, {}));
using LSEGridDesc_M = decltype(MakeLSEGridDescriptor_M(1)); using LSEGridDesc_M = decltype(MakeLSEGridDescriptor_M(1));
using AGridDesc_G_M_K = decltype(Transform::MakeAGridDescriptor_G_M_K({}, {})); using AGridDesc_G_M_K = decltype(Transform::MakeAGridDescriptor_G_M_K({}, {}));
using BGridDesc_G_N_K = decltype(Transform::MakeB0GridDescriptor_G_N_K({}, {})); using BGridDesc_G_N_K = decltype(Transform::MakeB0GridDescriptor_G_N_K({}, {}));
using D0GridDesc_G_M_N = decltype(Transform::MakeCGridDescriptor_G_M_N({}, {}));
using B1GridDesc_G_N_K = decltype(Transform::MakeB1GridDescriptor_G_N_K({}, {})); using B1GridDesc_G_N_K = decltype(Transform::MakeB1GridDescriptor_G_N_K({}, {}));
using CGridDesc_G_M_N = decltype(Transform::MakeCGridDescriptor_G_M_N({}, {})); using CGridDesc_G_M_N = decltype(Transform::MakeCGridDescriptor_G_M_N({}, {}));
using ZGridDesc_G_M_N = decltype(Transform::MakeCGridDescriptor_G_M_N({}, {})); using ZGridDesc_G_M_N = decltype(Transform::MakeCGridDescriptor_G_M_N({}, {}));
...@@ -420,12 +444,14 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2 ...@@ -420,12 +444,14 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
{ {
ComputeBasePtrOfStridedBatch(const AGridDesc_G_M_K& a_grid_desc_g_m_k, ComputeBasePtrOfStridedBatch(const AGridDesc_G_M_K& a_grid_desc_g_m_k,
const BGridDesc_G_N_K& b_grid_desc_g_n_k, const BGridDesc_G_N_K& b_grid_desc_g_n_k,
const D0GridDesc_G_M_N& d0_grid_desc_g_m_n,
const B1GridDesc_G_N_K& b1_grid_desc_g_n_k, const B1GridDesc_G_N_K& b1_grid_desc_g_n_k,
const CGridDesc_G_M_N& c_grid_desc_g_m_n, const CGridDesc_G_M_N& c_grid_desc_g_m_n,
const ZGridDesc_G_M_N& z_grid_desc_g_m_n, const ZGridDesc_G_M_N& z_grid_desc_g_m_n,
index_t BatchStrideLSE) index_t BatchStrideLSE)
: a_grid_desc_g_m_k_(a_grid_desc_g_m_k), : a_grid_desc_g_m_k_(a_grid_desc_g_m_k),
b_grid_desc_g_n_k_(b_grid_desc_g_n_k), b_grid_desc_g_n_k_(b_grid_desc_g_n_k),
d0_grid_desc_g_m_n_(d0_grid_desc_g_m_n),
b1_grid_desc_g_n_k_(b1_grid_desc_g_n_k), b1_grid_desc_g_n_k_(b1_grid_desc_g_n_k),
c_grid_desc_g_m_n_(c_grid_desc_g_m_n), c_grid_desc_g_m_n_(c_grid_desc_g_m_n),
z_grid_desc_g_m_n_(z_grid_desc_g_m_n), z_grid_desc_g_m_n_(z_grid_desc_g_m_n),
...@@ -443,6 +469,11 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2 ...@@ -443,6 +469,11 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
return b_grid_desc_g_n_k_.CalculateOffset(make_multi_index(g_idx, 0, 0)); return b_grid_desc_g_n_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
} }
__host__ __device__ constexpr long_index_t GetD0BasePtr(index_t g_idx) const
{
return d0_grid_desc_g_m_n_.CalculateOffset(make_multi_index(g_idx, 0, 0));
}
__host__ __device__ constexpr long_index_t GetB1BasePtr(index_t g_idx) const __host__ __device__ constexpr long_index_t GetB1BasePtr(index_t g_idx) const
{ {
return b1_grid_desc_g_n_k_.CalculateOffset(make_multi_index(g_idx, 0, 0)); return b1_grid_desc_g_n_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
...@@ -466,6 +497,7 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2 ...@@ -466,6 +497,7 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
private: private:
AGridDesc_G_M_K a_grid_desc_g_m_k_; AGridDesc_G_M_K a_grid_desc_g_m_k_;
BGridDesc_G_N_K b_grid_desc_g_n_k_; BGridDesc_G_N_K b_grid_desc_g_n_k_;
D0GridDesc_G_M_N d0_grid_desc_g_m_n_;
B1GridDesc_G_N_K b1_grid_desc_g_n_k_; B1GridDesc_G_N_K b1_grid_desc_g_n_k_;
CGridDesc_G_M_N c_grid_desc_g_m_n_; CGridDesc_G_M_N c_grid_desc_g_m_n_;
ZGridDesc_G_M_N z_grid_desc_g_m_n_; ZGridDesc_G_M_N z_grid_desc_g_m_n_;
...@@ -475,6 +507,7 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2 ...@@ -475,6 +507,7 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
// GridwiseGemm // GridwiseGemm
using GridwiseGemm = GridwiseBatchedMultiheadAttentionForward_Xdl_CShuffle_V2< using GridwiseGemm = GridwiseBatchedMultiheadAttentionForward_Xdl_CShuffle_V2<
ADataType, // TODO: distinguish A/B datatype ADataType, // TODO: distinguish A/B datatype
Acc0BiasDataType,
ZDataType, ZDataType,
GemmDataType, GemmDataType,
GemmAccDataType, GemmAccDataType,
...@@ -489,6 +522,7 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2 ...@@ -489,6 +522,7 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
InMemoryDataOperationEnum::Set, InMemoryDataOperationEnum::Set,
AGridDesc_AK0_M_AK1, AGridDesc_AK0_M_AK1,
BGridDesc_BK0_N_BK1, BGridDesc_BK0_N_BK1,
D0GridDesc_M_N,
B1GridDesc_BK0_N_BK1, B1GridDesc_BK0_N_BK1,
CGridDesc_M_N, CGridDesc_M_N,
ZGridDesc_M_N, ZGridDesc_M_N,
...@@ -508,6 +542,7 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2 ...@@ -508,6 +542,7 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
MXdlPerWave, MXdlPerWave,
NXdlPerWave, NXdlPerWave,
Gemm1NXdlPerWave, Gemm1NXdlPerWave,
DropoutStep,
ABlockTransferThreadClusterLengths_AK0_M_AK1, ABlockTransferThreadClusterLengths_AK0_M_AK1,
ABlockTransferThreadClusterArrangeOrder, ABlockTransferThreadClusterArrangeOrder,
ABlockTransferSrcAccessOrder, ABlockTransferSrcAccessOrder,
...@@ -524,6 +559,7 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2 ...@@ -524,6 +559,7 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
BBlockTransferDstScalarPerVector_BK1, BBlockTransferDstScalarPerVector_BK1,
true, true,
BBlockLdsExtraN, BBlockLdsExtraN,
Acc0BiasTransferSrcScalarPerVector,
B1BlockTransferThreadClusterLengths_BK0_N_BK1, B1BlockTransferThreadClusterLengths_BK0_N_BK1,
B1BlockTransferThreadClusterArrangeOrder, B1BlockTransferThreadClusterArrangeOrder,
B1BlockTransferSrcAccessOrder, B1BlockTransferSrcAccessOrder,
...@@ -536,6 +572,7 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2 ...@@ -536,6 +572,7 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
CShuffleNXdlPerWavePerShuffle, CShuffleNXdlPerWavePerShuffle,
CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
CShuffleBlockTransferScalarPerVector_NPerBlock, CShuffleBlockTransferScalarPerVector_NPerBlock,
Acc1BiasTransferSrcScalarPerVector,
LoopSched, LoopSched,
Transform::matrix_padder.PadN, Transform::matrix_padder.PadN,
MaskingSpec != MaskingSpecialization::MaskDisabled, MaskingSpec != MaskingSpecialization::MaskDisabled,
...@@ -545,15 +582,14 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2 ...@@ -545,15 +582,14 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
// FIXME: constness // FIXME: constness
struct Argument : public BaseArgument struct Argument : public BaseArgument
{ {
Argument( Argument(const ADataType* p_a_grid,
const ADataType* p_a_grid,
const BDataType* p_b_grid, const BDataType* p_b_grid,
const B1DataType* p_b1_grid, const B1DataType* p_b1_grid,
CDataType* p_c_grid, CDataType* p_c_grid,
ZDataType* p_z_grid, ZDataType* p_z_grid,
LSEDataType* p_lse_grid, LSEDataType* p_lse_grid,
const std::array<void*, NumAcc0Bias> p_acc0_biases, const D0DataType* p_acc0_biases,
const std::array<void*, NumAcc1Bias> p_acc1_biases, const D1DataType* p_acc1_biases,
const std::vector<index_t>& a_gs_ms_ks_lengths, const std::vector<index_t>& a_gs_ms_ks_lengths,
const std::vector<index_t>& a_gs_ms_ks_strides, const std::vector<index_t>& a_gs_ms_ks_strides,
const std::vector<index_t>& b_gs_ns_ks_lengths, const std::vector<index_t>& b_gs_ns_ks_lengths,
...@@ -565,11 +601,11 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2 ...@@ -565,11 +601,11 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
const std::vector<index_t>& z_gs_ms_ns_lengths, const std::vector<index_t>& z_gs_ms_ns_lengths,
const std::vector<index_t>& z_gs_ms_ns_strides, const std::vector<index_t>& z_gs_ms_ns_strides,
const std::vector<index_t>& lse_gs_ms_lengths, const std::vector<index_t>& lse_gs_ms_lengths,
const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_lengths, const std::vector<index_t>& acc0_biases_gs_ms_ns_lengths,
const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_strides, const std::vector<index_t>& acc0_biases_gs_ms_ns_strides,
const std::array<std::vector<ck::index_t>, NumAcc1Bias> const std::vector<index_t>&
acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths
const std::array<std::vector<ck::index_t>, NumAcc1Bias> const std::vector<index_t>&
acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
AElementwiseOperation a_element_op, AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op, BElementwiseOperation b_element_op,
...@@ -580,6 +616,7 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2 ...@@ -580,6 +616,7 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
std::tuple<unsigned long long, unsigned long long> seeds) std::tuple<unsigned long long, unsigned long long> seeds)
: p_a_grid_{p_a_grid}, : p_a_grid_{p_a_grid},
p_b_grid_{p_b_grid}, p_b_grid_{p_b_grid},
p_d0_grid_{p_acc0_biases},
p_b1_grid_{p_b1_grid}, p_b1_grid_{p_b1_grid},
p_c_grid_{p_c_grid}, p_c_grid_{p_c_grid},
p_z_grid_{p_z_grid}, p_z_grid_{p_z_grid},
...@@ -628,16 +665,14 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2 ...@@ -628,16 +665,14 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
compute_base_ptr_of_batch_{ compute_base_ptr_of_batch_{
a_grid_desc_g_m_k_, a_grid_desc_g_m_k_,
b_grid_desc_g_n_k_, b_grid_desc_g_n_k_,
d0_grid_desc_g_m_n_,
b1_grid_desc_g_n_k_, b1_grid_desc_g_n_k_,
c_grid_desc_g_m_n_, c_grid_desc_g_m_n_,
z_grid_desc_g_m_n_, z_grid_desc_g_m_n_,
type_convert<index_t>(lse_grid_desc_m_.GetElementSpaceSize())} type_convert<index_t>(lse_grid_desc_m_.GetElementSpaceSize())}
{ {
// TODO ANT: implement bias addition // TODO ANT: implement bias addition
ignore = p_acc0_biases;
ignore = p_acc1_biases; ignore = p_acc1_biases;
ignore = acc0_biases_gs_ms_ns_lengths;
ignore = acc0_biases_gs_ms_ns_strides;
ignore = acc1_biases_gs_ms_gemm1ns_lengths; ignore = acc1_biases_gs_ms_gemm1ns_lengths;
ignore = acc1_biases_gs_ms_gemm1ns_strides; ignore = acc1_biases_gs_ms_gemm1ns_strides;
...@@ -650,6 +685,21 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2 ...@@ -650,6 +685,21 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
c_grid_desc_mblock_mperblock_nblock_nperblock_ = c_grid_desc_mblock_mperblock_nblock_nperblock_ =
GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock( GridwiseGemm::MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
c_grid_desc_m_n_); c_grid_desc_m_n_);
if constexpr(!is_same<D0DataType, void>::value)
{
d0_grid_desc_m_n_ = Transform::MakeCGridDescriptor_M_N(
acc0_biases_gs_ms_ns_lengths, acc0_biases_gs_ms_ns_strides);
d0_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_ =
GridwiseGemm::MakeD0GridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(
d0_grid_desc_m_n_);
d0_grid_desc_g_m_n_ = Transform::MakeCGridDescriptor_G_M_N(
acc0_biases_gs_ms_ns_lengths, acc0_biases_gs_ms_ns_strides);
d0_n_length_stride_.push_back(acc0_biases_gs_ms_ns_lengths[NumDimG + NumDimM]);
d0_n_length_stride_.push_back(acc0_biases_gs_ms_ns_strides[NumDimG + NumDimM]);
}
} }
is_dropout_ = p_dropout > 0.0; // is_dropout_ = p_dropout > 0.0; //
...@@ -661,8 +711,9 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2 ...@@ -661,8 +711,9 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
seed_ = std::get<0>(seeds); seed_ = std::get<0>(seeds);
offset_ = std::get<1>(seeds); offset_ = std::get<1>(seeds);
z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_ = z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_m4_n4_n5_n6_ =
GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(z_grid_desc_m_n_); GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_M4_N4_N5_N6(
z_grid_desc_m_n_);
m_raw_padded_ = GridwiseGemm::GetPaddedSize(raw_lengths_mz_nz_kz_gemm1nz_[0]); m_raw_padded_ = GridwiseGemm::GetPaddedSize(raw_lengths_mz_nz_kz_gemm1nz_[0]);
n_raw_padded_ = GridwiseGemm::GetPaddedSize(raw_lengths_mz_nz_kz_gemm1nz_[1]); n_raw_padded_ = GridwiseGemm::GetPaddedSize(raw_lengths_mz_nz_kz_gemm1nz_[1]);
...@@ -681,6 +732,13 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2 ...@@ -681,6 +732,13 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
std::cout << "b_grid_desc_g_n_k_: " << b_grid_desc_g_n_k_.GetLength(I0) << ", " std::cout << "b_grid_desc_g_n_k_: " << b_grid_desc_g_n_k_.GetLength(I0) << ", "
<< b_grid_desc_g_n_k_.GetLength(I1) << ", " << b_grid_desc_g_n_k_.GetLength(I1) << ", "
<< b_grid_desc_g_n_k_.GetLength(I2) << '\n'; << b_grid_desc_g_n_k_.GetLength(I2) << '\n';
std::cout << "d0_grid_desc_g_m_n_: " << d0_grid_desc_g_m_n_.GetLength(I0) << ", "
<< d0_grid_desc_g_m_n_.GetLength(I1) << ", "
<< d0_grid_desc_g_m_n_.GetLength(I2) << '\n';
std::cout << "d0_grid_desc_m_n_: " << d0_grid_desc_m_n_.GetLength(I0) << ", "
<< d0_grid_desc_m_n_.GetLength(I1) << '\n';
std::cout << "b1_grid_desc_g_n_k_: " << b1_grid_desc_g_n_k_.GetLength(I0) << ", " std::cout << "b1_grid_desc_g_n_k_: " << b1_grid_desc_g_n_k_.GetLength(I0) << ", "
<< b1_grid_desc_g_n_k_.GetLength(I1) << ", " << b1_grid_desc_g_n_k_.GetLength(I1) << ", "
<< b1_grid_desc_g_n_k_.GetLength(I2) << '\n'; << b1_grid_desc_g_n_k_.GetLength(I2) << '\n';
...@@ -692,6 +750,7 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2 ...@@ -692,6 +750,7 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
// pointers // pointers
const ADataType* p_a_grid_; const ADataType* p_a_grid_;
const BDataType* p_b_grid_; const BDataType* p_b_grid_;
const D0DataType* p_d0_grid_;
const B1DataType* p_b1_grid_; const B1DataType* p_b1_grid_;
CDataType* p_c_grid_; CDataType* p_c_grid_;
ZDataType* p_z_grid_; ZDataType* p_z_grid_;
...@@ -700,6 +759,9 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2 ...@@ -700,6 +759,9 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
// tensor descriptor // tensor descriptor
AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_; AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_; BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
D0GridDesc_M_N d0_grid_desc_m_n_;
typename GridwiseGemm::D0GridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5
d0_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_;
B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1_; B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1_;
CGridDesc_M_N c_grid_desc_m_n_; CGridDesc_M_N c_grid_desc_m_n_;
ZGridDesc_M_N z_grid_desc_m_n_; ZGridDesc_M_N z_grid_desc_m_n_;
...@@ -707,14 +769,16 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2 ...@@ -707,14 +769,16 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
AGridDesc_G_M_K a_grid_desc_g_m_k_; AGridDesc_G_M_K a_grid_desc_g_m_k_;
BGridDesc_G_N_K b_grid_desc_g_n_k_; BGridDesc_G_N_K b_grid_desc_g_n_k_;
D0GridDesc_G_M_N d0_grid_desc_g_m_n_;
B1GridDesc_G_N_K b1_grid_desc_g_n_k_; B1GridDesc_G_N_K b1_grid_desc_g_n_k_;
CGridDesc_G_M_N c_grid_desc_g_m_n_; CGridDesc_G_M_N c_grid_desc_g_m_n_;
ZGridDesc_G_M_N z_grid_desc_g_m_n_; ZGridDesc_G_M_N z_grid_desc_g_m_n_;
typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
c_grid_desc_mblock_mperblock_nblock_nperblock_; c_grid_desc_mblock_mperblock_nblock_nperblock_;
typename GridwiseGemm::ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5 typename GridwiseGemm::ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_M4_N4_N5_N6
z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_; z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_m4_n4_n5_n6_;
// block-to-c-tile map // block-to-c-tile map
typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_; typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
...@@ -750,6 +814,9 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2 ...@@ -750,6 +814,9 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
index_t m_raw_padded_; index_t m_raw_padded_;
index_t n_raw_padded_; index_t n_raw_padded_;
// raw data
std::vector<ck::index_t> d0_n_length_stride_;
}; };
// Invoker // Invoker
...@@ -780,6 +847,7 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2 ...@@ -780,6 +847,7 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
const auto kernel = kernel_batched_multiheadattention_forward_xdl_cshuffle_v2< const auto kernel = kernel_batched_multiheadattention_forward_xdl_cshuffle_v2<
GridwiseGemm, GridwiseGemm,
ADataType, // TODO: distiguish A/B datatype ADataType, // TODO: distiguish A/B datatype
D0DataType,
CDataType, CDataType,
ZDataType, ZDataType,
LSEDataType, LSEDataType,
...@@ -791,9 +859,10 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2 ...@@ -791,9 +859,10 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
CElementwiseOperation, CElementwiseOperation,
DeviceOp::AGridDesc_AK0_M_AK1, DeviceOp::AGridDesc_AK0_M_AK1,
DeviceOp::BGridDesc_BK0_N_BK1, DeviceOp::BGridDesc_BK0_N_BK1,
typename GridwiseGemm::D0GridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5,
DeviceOp::B1GridDesc_BK0_N_BK1, DeviceOp::B1GridDesc_BK0_N_BK1,
typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock, typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
typename GridwiseGemm::ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5, typename GridwiseGemm::ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_M4_N4_N5_N6,
DeviceOp::LSEGridDesc_M, DeviceOp::LSEGridDesc_M,
typename GridwiseGemm::DefaultBlock2CTileMap, typename GridwiseGemm::DefaultBlock2CTileMap,
ComputeBasePtrOfStridedBatch, ComputeBasePtrOfStridedBatch,
...@@ -811,6 +880,7 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2 ...@@ -811,6 +880,7 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
0, 0,
arg.p_a_grid_, arg.p_a_grid_,
arg.p_b_grid_, arg.p_b_grid_,
arg.p_d0_grid_,
arg.p_b1_grid_, arg.p_b1_grid_,
arg.p_c_grid_, arg.p_c_grid_,
arg.p_z_grid_, arg.p_z_grid_,
...@@ -822,9 +892,10 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2 ...@@ -822,9 +892,10 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
arg.c_element_op_, arg.c_element_op_,
arg.a_grid_desc_ak0_m_ak1_, arg.a_grid_desc_ak0_m_ak1_,
arg.b_grid_desc_bk0_n_bk1_, arg.b_grid_desc_bk0_n_bk1_,
arg.d0_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_,
arg.b1_grid_desc_bk0_n_bk1_, arg.b1_grid_desc_bk0_n_bk1_,
arg.c_grid_desc_mblock_mperblock_nblock_nperblock_, arg.c_grid_desc_mblock_mperblock_nblock_nperblock_,
arg.z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_, arg.z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_m4_n4_n5_n6_,
arg.lse_grid_desc_m_, arg.lse_grid_desc_m_,
arg.block_2_ctile_map_, arg.block_2_ctile_map_,
arg.batch_count_, arg.batch_count_,
...@@ -952,6 +1023,19 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2 ...@@ -952,6 +1023,19 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
return false; return false;
} }
if constexpr(!is_same<D0DataType, void>::value)
{
if(arg.d0_n_length_stride_[1] == 1 &&
arg.d0_n_length_stride_[0] % Acc0BiasTransferSrcScalarPerVector != 0)
{
return false;
}
if(arg.d0_n_length_stride_[1] != 1 && Acc0BiasTransferSrcScalarPerVector != 1)
{
return false;
}
}
// Note: we need raw lengths since threadwise copy can not handle vector load when part of // Note: we need raw lengths since threadwise copy can not handle vector load when part of
// vector is out of bounds // vector is out of bounds
// Note: need lowest dim in Ms/Ns/Ks/Os, not merged M/N/K/O // Note: need lowest dim in Ms/Ns/Ks/Os, not merged M/N/K/O
...@@ -1003,15 +1087,15 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2 ...@@ -1003,15 +1087,15 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg)); return IsSupportedArgument(*dynamic_cast<const Argument*>(p_arg));
} }
static auto MakeArgument( static auto
const ADataType* p_a, MakeArgument(const ADataType* p_a,
const BDataType* p_b, const BDataType* p_b,
const B1DataType* p_b1, const B1DataType* p_b1,
CDataType* p_c, CDataType* p_c,
ZDataType* p_z, ZDataType* p_z,
LSEDataType* p_lse, LSEDataType* p_lse,
const std::array<void*, NumAcc0Bias> p_acc0_biases, const D0DataType* p_acc0_biases,
const std::array<void*, NumAcc1Bias> p_acc1_biases, const D1DataType* p_acc1_biases,
const std::vector<index_t>& a_gs_ms_ks_lengths, const std::vector<index_t>& a_gs_ms_ks_lengths,
const std::vector<index_t>& a_gs_ms_ks_strides, const std::vector<index_t>& a_gs_ms_ks_strides,
const std::vector<index_t>& b_gs_ns_ks_lengths, const std::vector<index_t>& b_gs_ns_ks_lengths,
...@@ -1023,11 +1107,11 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2 ...@@ -1023,11 +1107,11 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
const std::vector<index_t>& z_gs_ms_ns_lengths, const std::vector<index_t>& z_gs_ms_ns_lengths,
const std::vector<index_t>& z_gs_ms_ns_strides, const std::vector<index_t>& z_gs_ms_ns_strides,
const std::vector<index_t>& lse_gs_ms_lengths, const std::vector<index_t>& lse_gs_ms_lengths,
const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_lengths, const std::vector<index_t>& acc0_biases_gs_ms_ns_lengths,
const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_strides, const std::vector<index_t>& acc0_biases_gs_ms_ns_strides,
const std::array<std::vector<ck::index_t>, NumAcc1Bias> const std::vector<index_t>&
acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths
const std::array<std::vector<ck::index_t>, NumAcc1Bias> const std::vector<index_t>&
acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
AElementwiseOperation a_element_op, AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op, BElementwiseOperation b_element_op,
...@@ -1080,8 +1164,8 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2 ...@@ -1080,8 +1164,8 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
void* p_c, void* p_c,
void* p_z, void* p_z,
void* p_lse, void* p_lse,
const std::array<void*, NumAcc0Bias> p_acc0_biases, const void* p_acc0_biases,
const std::array<void*, NumAcc1Bias> p_acc1_biases, const void* p_acc1_biases,
const std::vector<index_t>& a_gs_ms_ks_lengths, const std::vector<index_t>& a_gs_ms_ks_lengths,
const std::vector<index_t>& a_gs_ms_ks_strides, const std::vector<index_t>& a_gs_ms_ks_strides,
const std::vector<index_t>& b_gs_ns_ks_lengths, const std::vector<index_t>& b_gs_ns_ks_lengths,
...@@ -1093,11 +1177,11 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2 ...@@ -1093,11 +1177,11 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
const std::vector<index_t>& z_gs_ms_ns_lengths, const std::vector<index_t>& z_gs_ms_ns_lengths,
const std::vector<index_t>& z_gs_ms_ns_strides, const std::vector<index_t>& z_gs_ms_ns_strides,
const std::vector<index_t>& lse_gs_ms_lengths, const std::vector<index_t>& lse_gs_ms_lengths,
const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_lengths, const std::vector<index_t>& acc0_biases_gs_ms_ns_lengths,
const std::array<std::vector<ck::index_t>, NumAcc0Bias> acc0_biases_gs_ms_ns_strides, const std::vector<index_t>& acc0_biases_gs_ms_ns_strides,
const std::array<std::vector<ck::index_t>, NumAcc1Bias> const std::vector<index_t>&
acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths acc1_biases_gs_ms_gemm1ns_lengths, // acc1_biases_gs_ms_os_lengths
const std::array<std::vector<ck::index_t>, NumAcc1Bias> const std::vector<index_t>&
acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides acc1_biases_gs_ms_gemm1ns_strides, // acc1_biases_gs_ms_os_strides
AElementwiseOperation a_element_op, AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op, BElementwiseOperation b_element_op,
...@@ -1107,14 +1191,15 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2 ...@@ -1107,14 +1191,15 @@ struct DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
float p_dropout, float p_dropout,
std::tuple<unsigned long long, unsigned long long> seeds) override std::tuple<unsigned long long, unsigned long long> seeds) override
{ {
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a), return std::make_unique<Argument>(
static_cast<const ADataType*>(p_a),
static_cast<const BDataType*>(p_b), static_cast<const BDataType*>(p_b),
static_cast<const B1DataType*>(p_b1), static_cast<const B1DataType*>(p_b1),
static_cast<CDataType*>(p_c), static_cast<CDataType*>(p_c),
static_cast<ZDataType*>(p_z), static_cast<ZDataType*>(p_z),
static_cast<LSEDataType*>(p_lse), static_cast<LSEDataType*>(p_lse),
p_acc0_biases, // cast in struct Argument static_cast<const D0DataType*>(p_acc0_biases), // cast in struct Argument
p_acc1_biases, // cast in struct Argument static_cast<const D1DataType*>(p_acc1_biases), // cast in struct Argument
a_gs_ms_ks_lengths, a_gs_ms_ks_lengths,
a_gs_ms_ks_strides, a_gs_ms_ks_strides,
b_gs_ns_ks_lengths, b_gs_ns_ks_lengths,
......
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