Skip to content

GitLab

Unverified Commit c0c52268 authored by Dan Yao's avatar Dan Yao Committed by GitHub
Browse files

Merge pull request #905 from ROCmSoftwarePlatform/mha-train-develop-grad-bias 

flash attention output   bias grad
parents f04ec574 c88d1173
Hide whitespace changes
Inline Side-by-side
Showing with 201 additions and 69 deletions
include/ck/tensor_operation/gpu/grid/gridwise_batched_mha_bwd_xdl_cshuffle_qloop_b2t_v2.hpp
View file @ c0c52268
......@@ -95,6 +95,10 @@ template <typename InputDataType,
PipelineVersion PipelineVer = PipelineVersion::v1>
struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2
{
static_assert(Gemm1NPerBlock % KPerBlock == 0);
static_assert(MPerBlock % Gemm1KPerBlock == 0);
static_assert(NPerBlock % Gemm2KPerBlock == 0);
static_assert(LoopSched == LoopScheduler::Default,
"Non-default loop scheduler is currently not supported");
......@@ -1337,7 +1341,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2
using D0GridDescriptor_M0_N0_M1_M2_N1_M3 =
remove_cvref_t<decltype(MakeD0GridDescriptor_M0_N0_M1_M2_N1_M3(D0GridDesc_M_N{}))>;
struct D0Loader
struct D0Operator
{
template <typename DataType>
struct TypeTransform
......@@ -1357,13 +1361,12 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2
static_assert(NPerXdl == 32);
static_assert(D0BlockTransferSrcScalarPerVector * NThreadClusterLengths <= NPerBlock,
"D0BlockTransferSrcScalarPerVector * NThreadClusterLengths <= NPerBlock");
__host__ __device__ static constexpr auto GetD0BlockWriteDescriptor_M0_N0_M1_M2_N1_M3()
__host__ __device__ static constexpr auto GetD0BlockGlobalDescriptor_M0_N0_M1_M2_N1_M3()
{
// B1 matrix in LDS memory, dst of blockwise copy
return make_naive_tensor_descriptor_packed(
make_tuple(I1, I1, I1, D0M1, Number<NPerBlock>{}, D0M2));
}
__host__ __device__ static constexpr auto GetD0BlockReadDescriptor_N0_N1_M0_M1_M2()
__host__ __device__ static constexpr auto GetD0BlockVgprDescriptor_N0_N1_M0_M1_M2()
{
constexpr auto d0_raw_m0_n_m1 =
make_naive_tensor_descriptor_packed(make_tuple(D0M1, Number<NPerBlock>{}, D0M2));
......@@ -1378,15 +1381,20 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2
make_tuple(Sequence<2, 3>{}, Sequence<0, 1>{}, Sequence<4>{}));
return d0_n0_n1_m0_m1_m2;
}
static constexpr auto d0_block_write_desc_m0_n0_m1_m2_n1_m3 =
GetD0BlockWriteDescriptor_M0_N0_M1_M2_N1_M3();
static constexpr auto d0_block_read_desc_n0_n1_m0_m1_m2 =
GetD0BlockReadDescriptor_N0_N1_M0_M1_M2();
static constexpr auto d0_block_dst_desc_m0_n0_m1_m2_n1_m3 =
GetD0BlockGlobalDescriptor_M0_N0_M1_M2_N1_M3();
static constexpr auto d0_block_src_desc_n0_n1_m0_m1_m2 =
GetD0BlockVgprDescriptor_N0_N1_M0_M1_M2();
static constexpr auto d0_thread_desc_ =
make_naive_tensor_descriptor_packed(make_tuple(I1, I1, I4, I1, D0M2));
using D0BlockwiseCopy = ThreadGroupTensorSliceTransfer_v4r1<
static constexpr auto& d0grad_block_dst_desc_n0_n1_m0_m1_m2 =
d0_block_src_desc_n0_n1_m0_m1_m2;
static constexpr auto& d0grad_block_src_desc_m0_n0_m1_m2_n1_m3 =
d0_block_dst_desc_m0_n0_m1_m2_n1_m3;
using D0BlockwiseCopyGlobalToLds = ThreadGroupTensorSliceTransfer_v4r1<
ThisThreadBlock,
tensor_operation::element_wise::PassThrough,
tensor_operation::element_wise::PassThrough,
......@@ -1397,34 +1405,77 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2
1,
BlockSize / NThreadClusterLengths,
NThreadClusterLengths,
1>, // ThreadClusterLengths
Sequence<0, 1, 2, 3, 5, 4>, // ThreadClusterArrangeOrder
typename TypeTransform<D0DataType>::Type, // SrcData
typename TypeTransform<D0DataType>::Type, // DstData
D0GridDescriptor_M0_N0_M1_M2_N1_M3, // SrcDesc
decltype(d0_block_write_desc_m0_n0_m1_m2_n1_m3), // DstDesc
Sequence<0, 1, 2, 3, 5, 4>, // SrcDimAccessOrder
Sequence<0, 1, 2, 4, 3, 5>, // DstDimAccessOrder
4, // SrcVectorDim
5, // DstVectorDim
D0BlockTransferSrcScalarPerVector, // SrcScalarPerVector
4, // DstScalarPerVector
1>, // ThreadClusterLengths
Sequence<0, 1, 2, 3, 5, 4>, // ThreadClusterArrangeOrder
typename TypeTransform<D0DataType>::Type, // SrcData
typename TypeTransform<D0DataType>::Type, // DstData
D0GridDescriptor_M0_N0_M1_M2_N1_M3, // SrcDesc
decltype(d0_block_dst_desc_m0_n0_m1_m2_n1_m3), // DstDesc
Sequence<0, 1, 2, 3, 5, 4>, // SrcDimAccessOrder
Sequence<0, 1, 2, 4, 3, 5>, // DstDimAccessOrder
4, // SrcVectorDim
5, // DstVectorDim
4, // SrcScalarPerVector
4, // DstScalarPerVector
1,
1,
true,
true, // DstResetCoord
1>;
using D0ThreadWiseCopy =
ThreadwiseTensorSliceTransfer_v4<typename TypeTransform<D0DataType>::Type, // SrcData
typename TypeTransform<D0DataType>::Type, // DstData
decltype(d0_block_read_desc_n0_n1_m0_m1_m2), // SrcDesc
decltype(d0_thread_desc_), // DstDesc
using D0ThreadwiseCopyLdsToVgpr =
ThreadwiseTensorSliceTransfer_v4<typename TypeTransform<D0DataType>::Type, // SrcData
typename TypeTransform<D0DataType>::Type, // DstData
decltype(d0_block_src_desc_n0_n1_m0_m1_m2), // SrcDesc
decltype(d0_thread_desc_), // DstDesc
Sequence<1, 1, 4, 1, 4>, // SliceLengths
Sequence<0, 1, 2, 3, 4>, // DimAccessOrder
4, // SrcVectorDim
2, // SrcScalarPerVector
4, // SrcScalarPerVector
2>;
using D0GradThreadwiseCopyVgprToLds = ThreadwiseTensorSliceTransfer_v1r3<
FloatGemmAcc,
typename TypeTransform<D0DataType>::Type,
decltype(d0_thread_desc_),
decltype(d0grad_block_dst_desc_n0_n1_m0_m1_m2),
tensor_operation::element_wise::Scale, // CElementwiseOperation
Sequence<1, 1, 4, 1, 4>, // SliceLengths
Sequence<0, 1, 2, 3, 4>, // AccessOrder
4, // VectorDim
4, // ScalarPerVector
InMemoryDataOperationEnum::Set, // GlobalMemoryDataOperation
1, // DstScalarStrideInVector
true>;
using D0GradBlockwiseCopyLdsToGlobal = ThreadGroupTensorSliceTransfer_v4r1<
ThisThreadBlock,
tensor_operation::element_wise::PassThrough,
tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum::Set,
Sequence<I1, I1, I1, D0M1, NPerBlock, D0M2>, // BlockSliceLengths
Sequence<1,
1,
1,
BlockSize / NThreadClusterLengths,
NThreadClusterLengths,
1>, // ThreadClusterLengths
Sequence<0, 1, 2, 3, 5, 4>, // ThreadClusterArrangeOrder
typename TypeTransform<D0DataType>::Type, // SrcData
typename TypeTransform<D0DataType>::Type, // DstData
decltype(d0grad_block_src_desc_m0_n0_m1_m2_n1_m3), // SrcDesc
D0GridDescriptor_M0_N0_M1_M2_N1_M3, // DstDesc
Sequence<0, 1, 2, 4, 3, 5>, // SrcDimAccessOrder
Sequence<0, 1, 2, 3, 5, 4>, // DstDimAccessOrder
5, // SrcVectorDim
4, // DstVectorDim
4, // SrcScalarPerVector
D0BlockTransferSrcScalarPerVector, // DstScalarPerVector
1,
1,
true,
true, // DstResetCoord
1>;
};
struct SharedMemTrait
......@@ -1466,10 +1517,10 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2
sizeof(GemmDataType) / sizeof(FloatGemmAcc);
static constexpr auto d0_block_space_size_aligned = math::integer_least_multiple(
D0Loader::d0_block_write_desc_m0_n0_m1_m2_n1_m3.GetElementSpaceSize(), max_lds_align);
D0Operator::d0_block_dst_desc_m0_n0_m1_m2_n1_m3.GetElementSpaceSize(), max_lds_align);
static constexpr auto d0_block_space_offset =
k_block_space_size_aligned.value * sizeof(GemmDataType) /
D0Loader::template TypeTransform<D0DataType>::Size;
D0Operator::template TypeTransform<D0DataType>::Size;
// LDS allocation for C shuffle in LDS
static constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
......@@ -1497,7 +1548,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2
sizeof(FloatGemmAcc);
const index_t d0_bytes_end =
(SharedMemTrait::d0_block_space_offset + SharedMemTrait::d0_block_space_size_aligned) *
D0Loader::template TypeTransform<D0DataType>::Size0;
D0Operator::template TypeTransform<D0DataType>::Size0;
const index_t c_block_bytes_end =
SharedMemTrait::c_block_space_size * sizeof(FloatCShuffle);
......@@ -1526,6 +1577,7 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2
const InputDataType* __restrict__ p_ygrad_grid,
OutputDataType* __restrict__ p_qgrad_grid,
OutputDataType* __restrict__ p_kgrad_grid,
D0DataType* __restrict__ p_d0grad_grid,
OutputDataType* __restrict__ p_vgrad_grid,
void* __restrict__ p_shared,
const AElementwiseOperation& a_element_op,
......@@ -2080,17 +2132,31 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2
index_t gemm0_m_block_outer_index = num_gemm0_m_block_outer_loop - 1;
// D0
auto d0_block_copy_global_to_lds = typename D0Loader::D0BlockwiseCopy(
auto d0_block_copy_global_to_lds = typename D0Operator::D0BlockwiseCopyGlobalToLds(
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),
tensor_operation::element_wise::PassThrough{},
D0Loader::d0_block_write_desc_m0_n0_m1_m2_n1_m3,
D0Operator::d0_block_dst_desc_m0_n0_m1_m2_n1_m3,
make_multi_index(0, 0, 0, 0, 0, 0),
tensor_operation::element_wise::PassThrough{});
auto d0_thread_copy_lds_to_vgpr = typename D0Loader::D0ThreadWiseCopy(
auto d0_thread_copy_lds_to_vgpr = typename D0Operator::D0ThreadwiseCopyLdsToVgpr(
make_tuple(wave_id[I1], wave_m_n_id[I1], 0, wave_m_n_id[I0], 0));
auto& d0grad_grid_desc_m0_n0_m1_m2_n1_m3 = d0_grid_desc_m0_n0_m1_m2_n1_m3;
auto d0grad_thread_copy_vgpr_to_lds = typename D0Operator::D0GradThreadwiseCopyVgprToLds(
D0Operator::d0grad_block_dst_desc_n0_n1_m0_m1_m2,
make_tuple(wave_id[I1], wave_m_n_id[I1], 0, wave_m_n_id[I0], 0),
tensor_operation::element_wise::Scale{rp_dropout});
auto d0grad_block_copy_lds_to_global = typename D0Operator::D0GradBlockwiseCopyLdsToGlobal(
D0Operator::d0grad_block_src_desc_m0_n0_m1_m2_n1_m3,
make_multi_index(0, 0, 0, 0, 0, 0),
tensor_operation::element_wise::PassThrough{},
d0grad_grid_desc_m0_n0_m1_m2_n1_m3,
make_multi_index(gemm0_m_block_outer_index, block_work_idx_n, 0, 0, 0, 0),
tensor_operation::element_wise::PassThrough{});
if constexpr(Deterministic)
{
block_sync_lds();
......@@ -2295,50 +2361,53 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2
// add bias
if constexpr(!is_same<D0DataType, void>::value)
{
static constexpr auto& c_thread_desc = s_blockwise_gemm.GetCThreadDesc();
if(p_d0_grid != nullptr)
{
static constexpr auto& c_thread_desc = s_blockwise_gemm.GetCThreadDesc();
const auto d0_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_d0_grid, d0_grid_desc_m0_n0_m1_m2_n1_m3.GetElementSpaceSize());
const auto d0_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_d0_grid, d0_grid_desc_m0_n0_m1_m2_n1_m3.GetElementSpaceSize());
auto d0_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<D0DataType*>(p_shared) + SharedMemTrait::d0_block_space_offset,
D0Loader::d0_block_write_desc_m0_n0_m1_m2_n1_m3.GetElementSpaceSize());
auto d0_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<D0DataType*>(p_shared) + SharedMemTrait::d0_block_space_offset,
D0Operator::d0_block_dst_desc_m0_n0_m1_m2_n1_m3.GetElementSpaceSize());
auto d0_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, D0DataType>(
D0Loader::d0_thread_desc_.GetElementSpaceSize());
ignore = d0_thread_buf;
auto d0_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, D0DataType>(
D0Operator::d0_thread_desc_.GetElementSpaceSize());
static_for<0, D0M0, 1>{}([&](auto mr) {
// load data to lds
d0_block_copy_global_to_lds.RunRead(d0_grid_desc_m0_n0_m1_m2_n1_m3,
d0_grid_buf);
static_for<0, D0M0, 1>{}([&](auto mr) {
// load data to lds
d0_block_copy_global_to_lds.RunRead(d0_grid_desc_m0_n0_m1_m2_n1_m3,
d0_grid_buf);
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_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_block_copy_global_to_lds.RunWrite(
D0Loader::d0_block_write_desc_m0_n0_m1_m2_n1_m3, d0_block_buf);
block_sync_lds();
// read data form lds
d0_thread_copy_lds_to_vgpr.Run(D0Loader::d0_block_read_desc_n0_n1_m0_m1_m2,
make_tuple(I0, I0, I0, I0, I0),
d0_block_buf,
D0Loader::d0_thread_desc_,
make_tuple(I0, I0, I0, I0, I0),
d0_thread_buf);
// bias add
static_for<0, d0_thread_buf.Size(), 1>{}([&](auto i) {
constexpr index_t c_offset =
c_thread_desc.CalculateOffset(make_tuple(mr, I0, i));
s_slash_p_thread_buf(Number<c_offset>{}) +=
ck::type_convert<FloatGemmAcc>(d0_thread_buf[i]);
d0_block_copy_global_to_lds.RunWrite(
D0Operator::d0_block_dst_desc_m0_n0_m1_m2_n1_m3, d0_block_buf);
block_sync_lds();
// read data form lds
d0_thread_copy_lds_to_vgpr.Run(D0Operator::d0_block_src_desc_n0_n1_m0_m1_m2,
make_tuple(I0, I0, I0, I0, I0),
d0_block_buf,
D0Operator::d0_thread_desc_,
make_tuple(I0, I0, I0, I0, I0),
d0_thread_buf);
// bias add
static_for<0, d0_thread_buf.Size(), 1>{}([&](auto i) {
constexpr index_t c_offset =
c_thread_desc.CalculateOffset(make_tuple(mr, I0, i));
s_slash_p_thread_buf(Number<c_offset>{}) +=
ck::type_convert<FloatGemmAcc>(d0_thread_buf[i]);
});
});
});
d0_block_copy_global_to_lds.MoveSrcSliceWindow(
d0_grid_desc_m0_n0_m1_m2_n1_m3, make_multi_index(-1, 0, -D0M0.value, 0, 0, 0));
d0_block_copy_global_to_lds.MoveSrcSliceWindow(
d0_grid_desc_m0_n0_m1_m2_n1_m3,
make_multi_index(-1, 0, -D0M0.value, 0, 0, 0));
}
}
// P_i: = softmax(scalar * S_i:)
......@@ -2545,6 +2614,46 @@ struct GridwiseBatchedMultiheadAttentionBackward_Qloop_Xdl_CShuffle_V2
: y_dot_ygrad_thread_buf[Number<m>{}]);
});
// output bias grad
if constexpr(!is_same<D0DataType, void>::value)
{
if(p_d0grad_grid != nullptr)
{
auto d0grad_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_d0grad_grid, d0grad_grid_desc_m0_n0_m1_m2_n1_m3.GetElementSpaceSize());
auto d0grad_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<D0DataType*>(p_shared) + SharedMemTrait::d0_block_space_offset,
D0Operator::d0grad_block_src_desc_m0_n0_m1_m2_n1_m3.GetElementSpaceSize());
static_for<0, D0M0, 1>{}([&](auto mr) {
d0grad_thread_copy_vgpr_to_lds.Run(
D0Operator::d0_thread_desc_,
make_tuple(mr, I0, I0, I0, I0),
sgrad_thread_buf,
D0Operator::d0grad_block_dst_desc_n0_n1_m0_m1_m2,
d0grad_block_buf);
block_sync_lds();
// write data from lds to global
d0grad_block_copy_lds_to_global.Run(
D0Operator::d0grad_block_src_desc_m0_n0_m1_m2_n1_m3,
d0grad_block_buf,
d0grad_grid_desc_m0_n0_m1_m2_n1_m3,
d0grad_grid_buf,
I0);
d0grad_block_copy_lds_to_global.MoveDstSliceWindow(
d0grad_grid_desc_m0_n0_m1_m2_n1_m3, make_multi_index(0, 0, 1, 0, 0, 0));
});
d0grad_block_copy_lds_to_global.MoveDstSliceWindow(
d0grad_grid_desc_m0_n0_m1_m2_n1_m3,
make_multi_index(-1, 0, -D0M0.value, 0, 0, 0));
}
}
SubThreadBlock<BlockSize> gemm2_a_copy_subgroup(s_blockwise_gemm.GetWaveIdx()[I0],
s_blockwise_gemm.GetWaveIdx()[I1]);
......
include/ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp
View file @ c0c52268
......@@ -282,6 +282,29 @@ struct TransformBatchedContractionContractionToBatchedGemmGemm
return matrix_padder.PadCDescriptor_M_N(
MakeCGridDescriptorPair(c_gs_ms_os_lengths_vec, c_gs_ms_os_strides_vec).second);
}
//
// C0
//
static auto MakeC0GridDescriptorPair(const std::vector<index_t>& c_gs_ms_ns_lengths_vec,
const std::vector<index_t>& c_gs_ms_ns_strides_vec)
{
return MakeGridDescriptorPair<NumDimG, NumDimM, NumDimN, CSpec>(c_gs_ms_ns_lengths_vec,
c_gs_ms_ns_strides_vec);
}
// TODO: rename to G_MRaw_NRaw
static auto MakeC0GridDescriptor_G_M_N(const std::vector<index_t>& c_gs_ms_ns_lengths_vec,
const std::vector<index_t>& c_gs_ms_ns_strides_vec)
{
return MakeC0GridDescriptorPair(c_gs_ms_ns_lengths_vec, c_gs_ms_ns_strides_vec).first;
}
static auto MakeC0GridDescriptor_M_N(const std::vector<index_t>& c_gs_ms_ns_lengths_vec,
const std::vector<index_t>& c_gs_ms_ns_strides_vec)
{
return matrix_padder.PadC0Descriptor_M_N(
MakeC0GridDescriptorPair(c_gs_ms_ns_lengths_vec, c_gs_ms_ns_strides_vec).second);
}
};
} // namespace tensor_operation
......
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