Skip to content

GitLab

"include/ck/utility/array.hpp" did not exist on "498e71b09822406b1b050c5eb03edebfe04038a6"
Commit 882b3328 authored by letaoqin's avatar letaoqin
Browse files

fix d0 load desc names

parent c54a1014
Show whitespace changes
Inline Side-by-side
Showing with 53 additions and 53 deletions
example/52_flash_atten_bias/batched_multihead_attention_bias_backward_v2.cpp
View file @ 882b3328
...@@ -113,11 +113,11 @@ using DeviceGemmInstance = ...@@ -113,11 +113,11 @@ using DeviceGemmInstance =
#elif(DIM <= 64) #elif(DIM <= 64)
// clang-format off // clang-format off
using DeviceGemmInstance = using DeviceGemmInstance =
// ########################################################################################| NumDimG| NumDimM| NumDimN| NumDimK| NumDimO| InputDataType| OutputDataType| GemmDataType| ZDataType| LSEDataType| Acc0BiasDataType| Acc1BiasDataType| GemmAcc| CShuffle| A| B| Acc| B1| C| GEMM| ATensorSpec| B0TensorSpec| B1TensorSpec| CTensorSpec| NumGemmK| Block| Gemm01| Gemm0| Gemm0| Gemm1| Gemm1| 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| D0BlockTransfer| CShuffle| CShuffle| CBlockTransferClusterLengths| CShuffleBlockTransferScalarPerVector_NPerBlock| MaskingSpec| Deterministic|
// ########################################################################################| | | | | | | | | | | | | DataType| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| | | | | Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| | | | // ########################################################################################| | | | | | | | | | | | | DataType| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| | | | | Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| SrcScalar| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| | | |
// ########################################################################################| | | | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | | | | | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| | | | // ########################################################################################| | | | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | | | | | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerVector| PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| | | |
// ########################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | // ########################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | Wave| Wave| Wave| Wave| | | | | | | | | | | | | | | | | | | | | |
ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, InputDataType, OutputDataType, GemmDataType, ZDataType, LSEDataType, Acc0BiasDataType, Acc1BiasDataType, AccDataType, ShuffleDataType, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 64, 128, 64, 64, 32, 8, 8, 2, 32, 32, 2, 1, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 2, S<1, 32, 1, 8>, CShuffleBlockTransferScalarPerVector_NPerBlock, MaskingSpec, Deterministic>; ck::tensor_operation::device::DeviceBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1< NumDimG, NumDimM, NumDimN, NumDimK, NumDimO, InputDataType, OutputDataType, GemmDataType, ZDataType, LSEDataType, Acc0BiasDataType, Acc1BiasDataType, AccDataType, ShuffleDataType, QKVElementOp, QKVElementOp, Scale, QKVElementOp, YElementOp, GemmSpec, TensorSpecQ, TensorSpecK, TensorSpecV, TensorSpecY, 1, 256, 64, 128, 64, 64, 32, 8, 8, 2, 32, 32, 2, 1, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 4, 1, 2, S<1, 32, 1, 8>, 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| // ########################################################################################| 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|
// ########################################################################################| | | | | | | | | | | | | DataType| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| | | | | Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| | | | // ########################################################################################| | | | | | | | | | | | | DataType| DataType| Elementwise| Elementwise| Elementwise| Elementwise| Elementwise| Specialization| | | | | Prefetch| Size| MPer| NPer| KPer| NPer| KPer| | | | XDL| XDL| MXdl| NXdl| NXdl| NXdl| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| | | |
// ########################################################################################| | | | | | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | | | | | Stage| | Block| Block| Block| Block| Block| | | | | | Per| Per| Per| Per| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | 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| 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| | | |
......
include/ck/tensor_operation/gpu/grid/gridwise_batched_mha_bwd_xdl_cshuffle_qloop_b2t_v1.hpp
View file @ 882b3328
...@@ -1271,7 +1271,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1 ...@@ -1271,7 +1271,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1
return make_naive_tensor_descriptor_packed( return make_naive_tensor_descriptor_packed(
make_tuple(I1, I1, I1, D0M1, Number<NPerBlock>{}, D0M2)); make_tuple(I1, I1, I1, D0M1, Number<NPerBlock>{}, D0M2));
} }
__host__ __device__ static constexpr auto GetD0BlockReadDescriptor_N0_N1_M0_M1_M2_M3() __host__ __device__ static constexpr auto GetD0BlockReadDescriptor_N0_N1_M0_M1_M2()
{ {
constexpr auto d0_raw_m0_n_m1 = constexpr auto d0_raw_m0_n_m1 =
make_naive_tensor_descriptor_packed(make_tuple(D0M1, Number<NPerBlock>{}, D0M2)); make_naive_tensor_descriptor_packed(make_tuple(D0M1, Number<NPerBlock>{}, D0M2));
...@@ -1286,10 +1286,10 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1 ...@@ -1286,10 +1286,10 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1
make_tuple(Sequence<2, 3>{}, Sequence<0, 1>{}, Sequence<4>{})); make_tuple(Sequence<2, 3>{}, Sequence<0, 1>{}, Sequence<4>{}));
return d0_n0_n1_m0_m1_m2; return d0_n0_n1_m0_m1_m2;
} }
static constexpr auto d0_block_desc_m0_n0_m1_m2_n1_m3 = static constexpr auto d0_block_write_desc_m0_n0_m1_m2_n1_m3 =
GetD0BlockWriteDescriptor_M0_N0_M1_M2_N1_M3(); GetD0BlockWriteDescriptor_M0_N0_M1_M2_N1_M3();
static constexpr auto d0_block_desc_n0_n1_m0_m1_m2 = static constexpr auto d0_block_read_desc_n0_n1_m0_m1_m2 =
GetD0BlockReadDescriptor_N0_N1_M0_M1_M2_M3(); GetD0BlockReadDescriptor_N0_N1_M0_M1_M2();
static constexpr auto d0_thread_desc_ = static constexpr auto d0_thread_desc_ =
make_naive_tensor_descriptor_packed(make_tuple(I1, I1, I4, I1, D0M2)); make_naive_tensor_descriptor_packed(make_tuple(I1, I1, I4, I1, D0M2));
...@@ -1310,7 +1310,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1 ...@@ -1310,7 +1310,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1
typename TypeTransform<D0DataType>::Type, // SrcData typename TypeTransform<D0DataType>::Type, // SrcData
typename TypeTransform<D0DataType>::Type, // DstData typename TypeTransform<D0DataType>::Type, // DstData
D0GridDescriptor_M0_N0_M1_M2_N1_M3, // SrcDesc D0GridDescriptor_M0_N0_M1_M2_N1_M3, // SrcDesc
decltype(d0_block_desc_m0_n0_m1_m2_n1_m3), // DstDesc decltype(d0_block_write_desc_m0_n0_m1_m2_n1_m3), // DstDesc
Sequence<0, 1, 2, 3, 5, 4>, // SrcDimAccessOrder Sequence<0, 1, 2, 3, 5, 4>, // SrcDimAccessOrder
Sequence<0, 1, 2, 4, 3, 5>, // DstDimAccessOrder Sequence<0, 1, 2, 4, 3, 5>, // DstDimAccessOrder
4, // SrcVectorDim 4, // SrcVectorDim
...@@ -1326,7 +1326,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1 ...@@ -1326,7 +1326,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1
using D0ThreadCopy = using D0ThreadCopy =
ThreadwiseTensorSliceTransfer_v4<typename TypeTransform<D0DataType>::Type, // SrcData ThreadwiseTensorSliceTransfer_v4<typename TypeTransform<D0DataType>::Type, // SrcData
typename TypeTransform<D0DataType>::Type, // DstData typename TypeTransform<D0DataType>::Type, // DstData
decltype(d0_block_desc_n0_n1_m0_m1_m2), // SrcDesc decltype(d0_block_read_desc_n0_n1_m0_m1_m2), // SrcDesc
decltype(d0_thread_desc_), // DstDesc decltype(d0_thread_desc_), // DstDesc
Sequence<1, 1, 4, 1, 4>, // SliceLengths Sequence<1, 1, 4, 1, 4>, // SliceLengths
Sequence<0, 1, 2, 3, 4>, // DimAccessOrder Sequence<0, 1, 2, 3, 4>, // DimAccessOrder
...@@ -1916,7 +1916,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1 ...@@ -1916,7 +1916,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1
d0_grid_desc_m0_n0_m1_m2_n1_m3, d0_grid_desc_m0_n0_m1_m2_n1_m3,
make_multi_index(gemm0_m_block_outer_index, block_work_idx_n, 0, 0, 0, 0), make_multi_index(gemm0_m_block_outer_index, block_work_idx_n, 0, 0, 0, 0),
tensor_operation::element_wise::PassThrough{}, tensor_operation::element_wise::PassThrough{},
D0Loader::d0_block_desc_m0_n0_m1_m2_n1_m3, D0Loader::d0_block_write_desc_m0_n0_m1_m2_n1_m3,
make_multi_index(0, 0, 0, 0, 0, 0), make_multi_index(0, 0, 0, 0, 0, 0),
tensor_operation::element_wise::PassThrough{}); tensor_operation::element_wise::PassThrough{});
...@@ -2080,7 +2080,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1 ...@@ -2080,7 +2080,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1
auto d0_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>( auto d0_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<D0DataType*>(p_shared) + SharedMemTrait::k_block_space_offset, static_cast<D0DataType*>(p_shared) + SharedMemTrait::k_block_space_offset,
D0Loader::d0_block_desc_m0_n0_m1_m2_n1_m3.GetElementSpaceSize()); D0Loader::d0_block_write_desc_m0_n0_m1_m2_n1_m3.GetElementSpaceSize());
auto d0_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, D0DataType>( auto d0_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, D0DataType>(
D0Loader::d0_thread_desc_.GetElementSpaceSize()); D0Loader::d0_thread_desc_.GetElementSpaceSize());
...@@ -2093,11 +2093,11 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1 ...@@ -2093,11 +2093,11 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V1
d0_block_copy_global_to_lds.MoveSrcSliceWindow( d0_block_copy_global_to_lds.MoveSrcSliceWindow(
d0_grid_desc_m0_n0_m1_m2_n1_m3, make_multi_index(0, 0, 1, 0, 0, 0)); d0_grid_desc_m0_n0_m1_m2_n1_m3, make_multi_index(0, 0, 1, 0, 0, 0));
d0_block_copy_global_to_lds.RunWrite(D0Loader::d0_block_desc_m0_n0_m1_m2_n1_m3, d0_block_copy_global_to_lds.RunWrite(
d0_block_buf); D0Loader::d0_block_write_desc_m0_n0_m1_m2_n1_m3, d0_block_buf);
block_sync_lds(); block_sync_lds();
// read data form lds // read data form lds
d0_thread_copy_lds_to_vgpr.Run(D0Loader::d0_block_desc_n0_n1_m0_m1_m2, d0_thread_copy_lds_to_vgpr.Run(D0Loader::d0_block_read_desc_n0_n1_m0_m1_m2,
make_tuple(I0, I0, I0, I0, I0), make_tuple(I0, I0, I0, I0, I0),
d0_block_buf, d0_block_buf,
D0Loader::d0_thread_desc_, D0Loader::d0_thread_desc_,
......
include/ck/tensor_operation/gpu/grid/gridwise_batched_mha_bwd_xdl_cshuffle_qloop_b2t_v2.hpp
View file @ 882b3328
...@@ -1217,9 +1217,9 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2 ...@@ -1217,9 +1217,9 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2
make_tuple(Sequence<2, 3>{}, Sequence<0, 1>{}, Sequence<4>{})); make_tuple(Sequence<2, 3>{}, Sequence<0, 1>{}, Sequence<4>{}));
return d0_n0_n1_m0_m1_m2; return d0_n0_n1_m0_m1_m2;
} }
static constexpr auto d0_block_desc_m0_n0_m1_m2_n1_m3 = static constexpr auto d0_block_write_desc_m0_n0_m1_m2_n1_m3 =
GetD0BlockWriteDescriptor_M0_N0_M1_M2_N1_M3(); GetD0BlockWriteDescriptor_M0_N0_M1_M2_N1_M3();
static constexpr auto d0_block_desc_n0_n1_m0_m1_m2 = static constexpr auto d0_block_read_desc_n0_n1_m0_m1_m2 =
GetD0BlockReadDescriptor_N0_N1_M0_M1_M2(); GetD0BlockReadDescriptor_N0_N1_M0_M1_M2();
static constexpr auto d0_thread_desc_ = static constexpr auto d0_thread_desc_ =
...@@ -1241,7 +1241,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2 ...@@ -1241,7 +1241,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2
typename TypeTransform<D0DataType>::Type, // SrcData typename TypeTransform<D0DataType>::Type, // SrcData
typename TypeTransform<D0DataType>::Type, // DstData typename TypeTransform<D0DataType>::Type, // DstData
D0GridDescriptor_M0_N0_M1_M2_N1_M3, // SrcDesc D0GridDescriptor_M0_N0_M1_M2_N1_M3, // SrcDesc
decltype(d0_block_desc_m0_n0_m1_m2_n1_m3), // DstDesc decltype(d0_block_write_desc_m0_n0_m1_m2_n1_m3), // DstDesc
Sequence<0, 1, 2, 3, 5, 4>, // SrcDimAccessOrder Sequence<0, 1, 2, 3, 5, 4>, // SrcDimAccessOrder
Sequence<0, 1, 2, 4, 3, 5>, // DstDimAccessOrder Sequence<0, 1, 2, 4, 3, 5>, // DstDimAccessOrder
4, // SrcVectorDim 4, // SrcVectorDim
...@@ -1257,7 +1257,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2 ...@@ -1257,7 +1257,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2
using D0ThreadCopy = using D0ThreadCopy =
ThreadwiseTensorSliceTransfer_v4<typename TypeTransform<D0DataType>::Type, // SrcData ThreadwiseTensorSliceTransfer_v4<typename TypeTransform<D0DataType>::Type, // SrcData
typename TypeTransform<D0DataType>::Type, // DstData typename TypeTransform<D0DataType>::Type, // DstData
decltype(d0_block_desc_n0_n1_m0_m1_m2), // SrcDesc decltype(d0_block_read_desc_n0_n1_m0_m1_m2), // SrcDesc
decltype(d0_thread_desc_), // DstDesc decltype(d0_thread_desc_), // DstDesc
Sequence<1, 1, 4, 1, 4>, // SliceLengths Sequence<1, 1, 4, 1, 4>, // SliceLengths
Sequence<0, 1, 2, 3, 4>, // DimAccessOrder Sequence<0, 1, 2, 3, 4>, // DimAccessOrder
...@@ -1837,7 +1837,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2 ...@@ -1837,7 +1837,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2
d0_grid_desc_m0_n0_m1_m2_n1_m3, d0_grid_desc_m0_n0_m1_m2_n1_m3,
make_multi_index(gemm0_m_block_outer_index, block_work_idx_n, 0, 0, 0, 0), make_multi_index(gemm0_m_block_outer_index, block_work_idx_n, 0, 0, 0, 0),
tensor_operation::element_wise::PassThrough{}, tensor_operation::element_wise::PassThrough{},
D0Loader::d0_block_desc_m0_n0_m1_m2_n1_m3, D0Loader::d0_block_write_desc_m0_n0_m1_m2_n1_m3,
make_multi_index(0, 0, 0, 0, 0, 0), make_multi_index(0, 0, 0, 0, 0, 0),
tensor_operation::element_wise::PassThrough{}); tensor_operation::element_wise::PassThrough{});
...@@ -1994,7 +1994,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2 ...@@ -1994,7 +1994,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2
auto d0_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>( auto d0_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<D0DataType*>(p_shared) + SharedMemTrait::a_block_space_offset, static_cast<D0DataType*>(p_shared) + SharedMemTrait::a_block_space_offset,
D0Loader::d0_block_desc_m0_n0_m1_m2_n1_m3.GetElementSpaceSize()); D0Loader::d0_block_write_desc_m0_n0_m1_m2_n1_m3.GetElementSpaceSize());
auto d0_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, D0DataType>( auto d0_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, D0DataType>(
D0Loader::d0_thread_desc_.GetElementSpaceSize()); D0Loader::d0_thread_desc_.GetElementSpaceSize());
...@@ -2008,11 +2008,11 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2 ...@@ -2008,11 +2008,11 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2
d0_block_copy_global_to_lds.MoveSrcSliceWindow( d0_block_copy_global_to_lds.MoveSrcSliceWindow(
d0_grid_desc_m0_n0_m1_m2_n1_m3, make_multi_index(0, 0, 1, 0, 0, 0)); d0_grid_desc_m0_n0_m1_m2_n1_m3, make_multi_index(0, 0, 1, 0, 0, 0));
d0_block_copy_global_to_lds.RunWrite(D0Loader::d0_block_desc_m0_n0_m1_m2_n1_m3, d0_block_copy_global_to_lds.RunWrite(
d0_block_buf); D0Loader::d0_block_write_desc_m0_n0_m1_m2_n1_m3, d0_block_buf);
block_sync_lds(); block_sync_lds();
// read data form lds // read data form lds
d0_thread_copy_lds_to_vgpr.Run(D0Loader::d0_block_desc_n0_n1_m0_m1_m2, d0_thread_copy_lds_to_vgpr.Run(D0Loader::d0_block_read_desc_n0_n1_m0_m1_m2,
make_tuple(I0, I0, I0, I0, I0), make_tuple(I0, I0, I0, I0, I0),
d0_block_buf, d0_block_buf,
D0Loader::d0_thread_desc_, D0Loader::d0_thread_desc_,
......
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