Skip to content

GitLab

"docs/source/en/api/pipelines/stable_diffusion/overview.md" did not exist on "2a0c823527694058d410ed6f91b52e7dd9f94ebe"
Commit 01073007 authored by guangzlu's avatar guangzlu
Browse files

add dropout into attn bwd v4

parent cd24392e
Show whitespace changes
Inline Side-by-side
Showing with 209 additions and 92 deletions
include/ck/tensor_operation/gpu/device/impl/device_batched_multihead_attention_backward_xdl_cshuffle_v4.hpp
View file @ 01073007
...@@ -39,7 +39,7 @@ template <typename GridwiseGemm, ...@@ -39,7 +39,7 @@ 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 ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5, typename ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_M4_M5_N3,
typename B1GridDesc_BK0_N_BK1, typename B1GridDesc_BK0_N_BK1,
typename YGridDescriptor_MBlock_MPerBlock_OBlock_OPerBlock, typename YGridDescriptor_MBlock_MPerBlock_OBlock_OPerBlock,
typename LSEGridDescriptor_M, typename LSEGridDescriptor_M,
...@@ -71,8 +71,8 @@ __global__ void ...@@ -71,8 +71,8 @@ __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 ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5 const ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_M4_M5_N3
c_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5, c_grid_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3,
const B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1, const B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1,
const YGridDescriptor_MBlock_MPerBlock_OBlock_OPerBlock const YGridDescriptor_MBlock_MPerBlock_OBlock_OPerBlock
c_grid_desc_mblock_mperblock_nblock_nperblock, c_grid_desc_mblock_mperblock_nblock_nperblock,
...@@ -85,7 +85,9 @@ __global__ void ...@@ -85,7 +85,9 @@ __global__ void
const C0MatrixMask c0_matrix_mask, const C0MatrixMask c0_matrix_mask,
const float p_drop, const float p_drop,
const unsigned long long seed, const unsigned long long seed,
const unsigned long long offset) const unsigned long long offset,
const index_t MRaw,
const index_t NRaw)
{ {
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__)) #if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()]; __shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
...@@ -135,7 +137,7 @@ __global__ void ...@@ -135,7 +137,7 @@ __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,
c_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5, c_grid_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3,
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,
lse_grid_desc_m, lse_grid_desc_m,
...@@ -144,6 +146,9 @@ __global__ void ...@@ -144,6 +146,9 @@ __global__ void
c0_matrix_mask, c0_matrix_mask,
p_drop, p_drop,
ph, ph,
g_idx,
MRaw,
NRaw,
i); i);
} }
} }
...@@ -167,7 +172,7 @@ __global__ void ...@@ -167,7 +172,7 @@ __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,
c_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5, c_grid_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3,
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,
lse_grid_desc_m, lse_grid_desc_m,
...@@ -176,6 +181,9 @@ __global__ void ...@@ -176,6 +181,9 @@ __global__ void
c0_matrix_mask, c0_matrix_mask,
p_drop, p_drop,
ph, ph,
g_idx,
MRaw,
NRaw,
0); 0);
} }
#else #else
...@@ -818,8 +826,8 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -818,8 +826,8 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
seed_ = std::get<0>(seeds); seed_ = std::get<0>(seeds);
offset_ = std::get<1>(seeds); offset_ = std::get<1>(seeds);
c_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_ = c_grid_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3_ =
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_M4_M5_N3(z_grid_desc_m_n_);
// Print(); // Print();
} }
...@@ -879,8 +887,8 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -879,8 +887,8 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
typename GridwiseGemm::YGridDescriptor_MBlock_MPerBlock_OBlock_OPerBlock typename GridwiseGemm::YGridDescriptor_MBlock_MPerBlock_OBlock_OPerBlock
y_grid_desc_mblock_mperblock_oblock_operblock_; y_grid_desc_mblock_mperblock_oblock_operblock_;
typename GridwiseGemm::ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5 typename GridwiseGemm::ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_M4_M5_N3
c_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_; c_grid_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3_;
// block-to-c-tile map // block-to-c-tile map
typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_; typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
...@@ -943,7 +951,7 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -943,7 +951,7 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
CElementwiseOperation, CElementwiseOperation,
DeviceOp::AGridDesc_AK0_M_AK1, DeviceOp::AGridDesc_AK0_M_AK1,
DeviceOp::BGridDesc_BK0_N_BK1, DeviceOp::BGridDesc_BK0_N_BK1,
typename GridwiseGemm::ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5, typename GridwiseGemm::ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_M4_M5_N3,
DeviceOp::B1GridDesc_BK0_N_BK1, DeviceOp::B1GridDesc_BK0_N_BK1,
typename GridwiseGemm::YGridDescriptor_MBlock_MPerBlock_OBlock_OPerBlock, typename GridwiseGemm::YGridDescriptor_MBlock_MPerBlock_OBlock_OPerBlock,
DeviceOp::LSEGridDesc_M, DeviceOp::LSEGridDesc_M,
...@@ -977,7 +985,7 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -977,7 +985,7 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
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.c_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_, arg.c_grid_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3_,
arg.b1_grid_desc_bk0_n_bk1_, arg.b1_grid_desc_bk0_n_bk1_,
arg.y_grid_desc_mblock_mperblock_oblock_operblock_, arg.y_grid_desc_mblock_mperblock_oblock_operblock_,
arg.lse_grid_desc_m_, arg.lse_grid_desc_m_,
...@@ -989,7 +997,9 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -989,7 +997,9 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
arg.c0_matrix_mask_, arg.c0_matrix_mask_,
arg.p_drop_, arg.p_drop_,
arg.seed_, arg.seed_,
arg.offset_); arg.offset_,
arg.raw_lengths_mz_nz_kz_gemm1nz_[0],
arg.raw_lengths_mz_nz_kz_gemm1nz_[1]);
}; };
// Gemm1_K is split into Gemm1_K0/K1 where K1 is known at compile time, so we only need // Gemm1_K is split into Gemm1_K0/K1 where K1 is known at compile time, so we only need
......
include/ck/tensor_operation/gpu/grid/gridwise_batched_multihead_attention_backward_xdl_cshuffle_pt4.hpp
View file @ 01073007
...@@ -113,23 +113,23 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -113,23 +113,23 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
// C desc for source in blockwise copy // C desc for source in blockwise copy
__host__ __device__ static constexpr auto __host__ __device__ static constexpr auto
MakeCGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(const ZGridDesc_M_N& z_grid_desc_m_n) MakeCGridDescriptor_M0_N0_M1_N1_M2_N2_M3_M4_M5_N3(const ZGridDesc_M_N& z_grid_desc_m_n)
{ {
const auto M = z_grid_desc_m_n.GetLength(I0); const auto M = z_grid_desc_m_n.GetLength(I0);
const auto N = z_grid_desc_m_n.GetLength(I1); const auto N = z_grid_desc_m_n.GetLength(I1);
constexpr auto mfma = MfmaSelector<GemmDataType, MPerXdl, NPerXdl>::selected_mfma; constexpr auto mfma = MfmaSelector<GemmDataType, MPerXdl, NPerXdl>::selected_mfma;
constexpr auto N3 = mfma.num_groups_per_blk; constexpr auto M3 = mfma.num_groups_per_blk;
constexpr auto N4 = mfma.num_input_blks; constexpr auto M4 = mfma.num_input_blks;
constexpr auto N5 = mfma.group_size; constexpr auto M5 = mfma.group_size;
return transform_tensor_descriptor( return transform_tensor_descriptor(
z_grid_desc_m_n, z_grid_desc_m_n,
make_tuple(make_unmerge_transform( make_tuple(make_unmerge_transform(
make_tuple(M / MPerBlock, MXdlPerWave, Gemm0MWaves, MPerXdl)), make_tuple(M / MPerBlock, MXdlPerWave, Gemm0MWaves, M3, M4, M5)),
make_unmerge_transform( make_unmerge_transform(
make_tuple(N / NPerBlock, NXdlPerWave, Gemm0NWaves, N3, N4, N5))), make_tuple(N / NPerBlock, NXdlPerWave, Gemm0NWaves, NPerXdl))),
make_tuple(Sequence<0>{}, Sequence<1>{}), make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2, 4, 6>{}, Sequence<1, 3, 5, 7, 8, 9>{})); make_tuple(Sequence<0, 2, 4, 6, 7, 8>{}, Sequence<1, 3, 5, 9>{}));
} }
__device__ static auto GetGemm0WaveIdx() __device__ static auto GetGemm0WaveIdx()
...@@ -396,8 +396,8 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -396,8 +396,8 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
using DefaultBlock2CTileMap = using DefaultBlock2CTileMap =
remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(KGridDesc_N_K{}))>; remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(KGridDesc_N_K{}))>;
using ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5 = remove_cvref_t<decltype( using ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_M4_M5_N3 = remove_cvref_t<decltype(
MakeCGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(ZGridDesc_M_N{}))>; MakeCGridDescriptor_M0_N0_M1_N1_M2_N2_M3_M4_M5_N3(ZGridDesc_M_N{}))>;
// Q / K / V / dY // Q / K / V / dY
struct GemmBlockwiseCopy struct GemmBlockwiseCopy
...@@ -1234,8 +1234,8 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -1234,8 +1234,8 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
const CElementwiseOperation& c_element_op, const CElementwiseOperation& c_element_op,
const QGridDesc_K0_M_K1& q_grid_desc_k0_m_k1, const QGridDesc_K0_M_K1& q_grid_desc_k0_m_k1,
const KGridDesc_K0_N_K1& k_grid_desc_k0_n_k1, const KGridDesc_K0_N_K1& k_grid_desc_k0_n_k1,
const ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5& const ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_M4_M5_N3&
z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5, z_grid_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3,
const VGridDesc_O0_N_O1& v_grid_desc_o0_n_o1, const VGridDesc_O0_N_O1& v_grid_desc_o0_n_o1,
const YGridDescriptor_MBlock_MPerBlock_OBlock_OPerBlock& const YGridDescriptor_MBlock_MPerBlock_OBlock_OPerBlock&
y_grid_desc_mblock_mperblock_oblock_operblock, y_grid_desc_mblock_mperblock_oblock_operblock,
...@@ -1245,6 +1245,9 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -1245,6 +1245,9 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
const C0MatrixMask& c0_matrix_mask, const C0MatrixMask& c0_matrix_mask,
const float p_drop, const float p_drop,
ck::philox& ph, ck::philox& ph,
const index_t g_idx,
const index_t MRaw,
const index_t NRaw,
const index_t block_idx_n) const index_t block_idx_n)
{ {
const FloatGemmAcc p_dropout = type_convert<FloatGemmAcc>(1.0f - p_drop); const FloatGemmAcc p_dropout = type_convert<FloatGemmAcc>(1.0f - p_drop);
...@@ -1578,11 +1581,11 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -1578,11 +1581,11 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
// z vgpr copy to global // z vgpr copy to global
// //
// z matrix threadwise desc // z matrix threadwise desc
constexpr auto z_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5 = constexpr auto z_thread_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3 =
make_naive_tensor_descriptor_packed(make_tuple(I1, // MBlockId make_naive_tensor_descriptor_packed(make_tuple(I1, // MBlockId
I1, // NBlockID I1, // NBlockID
m0, // MRepeat m0, // MRepeat
I1, // NRepeat n0, // NRepeat
m1, // MWaveId m1, // MWaveId
n1, // NWaveId n1, // NWaveId
m2, // MPerXdl m2, // MPerXdl
...@@ -1592,33 +1595,29 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -1592,33 +1595,29 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
StaticBuffer<AddressSpaceEnum::Vgpr, StaticBuffer<AddressSpaceEnum::Vgpr,
unsigned short, unsigned short,
z_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5.GetElementSpaceSize(), z_thread_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3.GetElementSpaceSize(),
true> true>
z_tenor_buffer; z_tenor_buffer;
z_tenor_buffer.Clear(); z_tenor_buffer.Clear();
// z matrix global desc
/*const auto M = q_grid_desc_k0_m_k1.GetLength(I1);
const auto N = k_grid_desc_k0_n_k1.GetLength(I1);
auto z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5 =
MakeZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(M, N);*/
// z matrix global desc
auto z_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>( auto z_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_z_grid, z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5.GetElementSpaceSize()); p_z_grid, z_grid_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3.GetElementSpaceSize());
const auto wave_id = GetGemm0WaveIdx(); const auto wave_id = GetGemm0WaveIdx();
const auto wave_m_n_id = GetGemm0WaveMNIdx(wave_id[I2]); // I2: 0~63 const auto wave_m_n_id = GetGemm0WaveMNIdx(wave_id[I2]); // I2: 0~63
auto z_thread_copy_vgpr_to_global = ThreadwiseTensorSliceTransfer_v1r3< auto z_thread_copy_vgpr_to_global =
ushort, ThreadwiseTensorSliceTransfer_v1r3<ushort,
ZDataType, ZDataType,
decltype(z_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5), decltype(
decltype(z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5), z_thread_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3),
decltype(z_grid_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3),
tensor_operation::element_wise::PassThrough, tensor_operation::element_wise::PassThrough,
Sequence<I1, // MBlockId Sequence<I1, // MBlockId
I1, // NBlockID I1, // NBlockID
m0, // MRepeat m0, // MRepeat
I1, // NRepeat n0, // NRepeat
m1, // MWaveId m1, // MWaveId
n1, // NWaveId n1, // NWaveId
m2, // MPerXdl m2, // MPerXdl
...@@ -1630,17 +1629,17 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -1630,17 +1629,17 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
1, // DstScalarPerVector 1, // DstScalarPerVector
InMemoryDataOperationEnum::Set, InMemoryDataOperationEnum::Set,
1, // DstScalarStrideInVector 1, // DstScalarStrideInVector
true>{z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5, true>{z_grid_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3,
make_multi_index(block_work_idx_n, // MBlockId make_multi_index(0, // MBlockId
0, // NBlockId block_work_idx_n, // NBlockId
0, // mrepeat 0, // mrepeat
0, // nrepeat 0, // nrepeat
wave_id[I0], // MWaveId wave_id[I0], // MWaveId
wave_id[I1], // NWaveId wave_id[I1], // NWaveId
wave_m_n_id[I1], // MPerXdl 0, // MPerXdl
0, // group wave_m_n_id[I0], // group
wave_m_n_id[I0], // NInputIndex 0, // NInputIndex
0), wave_m_n_id[I1]),
tensor_operation::element_wise::PassThrough{}}; tensor_operation::element_wise::PassThrough{}};
// //
...@@ -1950,35 +1949,143 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -1950,35 +1949,143 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
// save z to global // save z to global
if(p_z_grid) if(p_z_grid)
{ {
// P_dropped // 8d thread_desc in thread scope
static_for<0, n0, 1>{}([&](auto i) { constexpr auto c_thread_lengths =
s_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
// 8d block_desc in block scope
constexpr auto c_block_lengths =
s_blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
constexpr auto M0 = c_block_lengths[I0];
constexpr auto N0 = c_block_lengths[I1];
constexpr auto M1 = c_block_lengths[I2];
constexpr auto N1 = c_block_lengths[I3];
constexpr auto M2 = c_block_lengths[I4];
constexpr auto M3 = c_block_lengths[I6];
constexpr auto M4 = c_block_lengths[I5];
constexpr auto N2 = c_block_lengths[I7];
// works like multi-dimension static_for (static_ford), but provides both the linear
// index as well as n-d index
using Acc0TileIterator = SpaceFillingCurve<
decltype(c_thread_lengths),
typename arithmetic_sequence_gen<0, c_thread_lengths.Size(), 1>::type,
typename uniform_sequence_gen<c_thread_lengths.Size(), 1>::type,
false>; // SnakeCurved
constexpr auto block_idx_to_m_n_adaptor = make_single_stage_tensor_adaptor(
make_tuple(make_unmerge_transform(make_tuple(M0, M1, M2, M3, M4)),
make_unmerge_transform(make_tuple(N0, N1, N2))),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2, 4, 5, 6>{}, Sequence<1, 3, 7>{}));
auto acc0_thread_idx = Acc0TileIterator::GetIndex(I0) + acc0_thread_origin;
auto m_local = block_idx_to_m_n_adaptor.CalculateBottomIndex(acc0_thread_idx)[I0];
auto n_local = block_idx_to_m_n_adaptor.CalculateBottomIndex(acc0_thread_idx)[I1];
auto m_global = m_local + m_block_data_idx_on_grid;
auto n_global = n_local + n_block_data_idx_on_grid;
auto global_elem_id =
MRaw * NRaw * g_idx + m_global * NRaw + n_global; // unique element global 1d id
// if(get_block_1d_id() == 0 && get_thread_local_1d_id()==64){
// printf("global_elem_id is %d \n", global_elem_id);
//}
// index_t id_step = Acc0TileIterator::GetNumOfAccess() / n0.value;
// if(get_thread_global_1d_id() == 0){
// printf("Acc0TileIterator::GetNumOfAccess() is %d \n",
// Acc0TileIterator::GetNumOfAccess()); printf("n0.value is %d \n", n0.value);
// printf("id_step is %d \n", id_step);
//}
blockwise_dropout.template ApplyDropout<decltype(s_slash_p_thread_buf), blockwise_dropout.template ApplyDropout<decltype(s_slash_p_thread_buf),
decltype(z_tenor_buffer), decltype(z_tenor_buffer),
true, true>(
decltype(n0), s_slash_p_thread_buf, ph, global_elem_id, z_tenor_buffer);
decltype(i)>(
s_slash_p_thread_buf, ph, z_tenor_buffer);
z_thread_copy_vgpr_to_global.Run( z_thread_copy_vgpr_to_global.Run(z_thread_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3,
z_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
make_tuple(I0, I0, I0, I0, I0, I0, I0, I0, I0, I0), make_tuple(I0, I0, I0, I0, I0, I0, I0, I0, I0, I0),
z_tenor_buffer, z_tenor_buffer,
z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5, z_grid_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3,
z_grid_buf); z_grid_buf);
z_thread_copy_vgpr_to_global.MoveDstSliceWindow( //// P_dropped
z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5, // static_for<0, n0, 1>{}([&](auto i) {
make_multi_index(0, 0, 0, 1, 0, 0, 0, 0, 0, 0)); // blockwise_dropout.template ApplyDropout<decltype(s_slash_p_thread_buf),
}); // decltype(z_tenor_buffer),
z_thread_copy_vgpr_to_global.MoveDstSliceWindow( // true,
z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5, // decltype(n0),
make_multi_index(0, 0, 0, -n0.value, 0, 0, 0, 0, 0, 0)); // decltype(i)>(
// s_slash_p_thread_buf, ph, z_tenor_buffer);
//
// z_thread_copy_vgpr_to_global.Run(
// z_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
// make_tuple(I0, I0, I0, I0, I0, I0, I0, I0, I0, I0),
// z_tenor_buffer,
// z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
// z_grid_buf);
// z_thread_copy_vgpr_to_global.MoveDstSliceWindow(
// z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
// make_multi_index(0, 0, 0, 1, 0, 0, 0, 0, 0, 0));
//});
// z_thread_copy_vgpr_to_global.MoveDstSliceWindow(
// z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
// make_multi_index(0, 0, 0, -n0.value, 0, 0, 0, 0, 0, 0));
} }
else else
{ {
// 8d thread_desc in thread scope
constexpr auto c_thread_lengths =
s_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
// 8d block_desc in block scope
constexpr auto c_block_lengths =
s_blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
constexpr auto M0 = c_block_lengths[I0];
constexpr auto N0 = c_block_lengths[I1];
constexpr auto M1 = c_block_lengths[I2];
constexpr auto N1 = c_block_lengths[I3];
constexpr auto M2 = c_block_lengths[I4];
constexpr auto M3 = c_block_lengths[I5];
constexpr auto M4 = c_block_lengths[I6];
constexpr auto N2 = c_block_lengths[I7];
// works like multi-dimension static_for (static_ford), but provides both the linear
// index as well as n-d index
using Acc0TileIterator = SpaceFillingCurve<
decltype(c_thread_lengths),
typename arithmetic_sequence_gen<0, c_thread_lengths.Size(), 1>::type,
typename uniform_sequence_gen<c_thread_lengths.Size(), 1>::type,
false>; // SnakeCurved
constexpr auto block_idx_to_m_n_adaptor = make_single_stage_tensor_adaptor(
make_tuple(make_unmerge_transform(make_tuple(M0, M1, M2, M3, M4)),
make_unmerge_transform(make_tuple(N0, N1, N2))),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2, 4, 5, 6>{}, Sequence<1, 3, 7>{}));
auto acc0_thread_idx = Acc0TileIterator::GetIndex(I0) + acc0_thread_origin;
auto m_local = block_idx_to_m_n_adaptor.CalculateBottomIndex(acc0_thread_idx)[I0];
auto n_local = block_idx_to_m_n_adaptor.CalculateBottomIndex(acc0_thread_idx)[I1];
auto m_global = m_local + m_block_data_idx_on_grid;
auto n_global = n_local + n_block_data_idx_on_grid;
auto global_elem_id =
MRaw * NRaw * g_idx + m_global * NRaw + n_global; // unique element global 1d id
// index_t id_step = Acc0TileIterator::GetNumOfAccess() / n0.value;
ignore = z_grid_buf; ignore = z_grid_buf;
// P_dropped // P_dropped
blockwise_dropout.template ApplyDropout<decltype(s_slash_p_thread_buf), true>( blockwise_dropout.template ApplyDropout<decltype(s_slash_p_thread_buf), true>(
s_slash_p_thread_buf, ph); s_slash_p_thread_buf, ph, global_elem_id);
// ignore = z_grid_buf;
//// P_dropped
// blockwise_dropout.template ApplyDropout<decltype(s_slash_p_thread_buf), true>(
// s_slash_p_thread_buf, ph);
} }
block_sync_lds(); // wait for gemm1 LDS read block_sync_lds(); // wait for gemm1 LDS read
...@@ -2160,8 +2267,8 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -2160,8 +2267,8 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
qgrad_thread_copy_vgpr_to_global.MoveDstSliceWindow( qgrad_thread_copy_vgpr_to_global.MoveDstSliceWindow(
qgrad_grid_desc_m0_o0_m1_o1_m2_o2_o3_o4, Gemm2::c_block_slice_copy_step); // step M qgrad_grid_desc_m0_o0_m1_o1_m2_o2_o3_o4, Gemm2::c_block_slice_copy_step); // step M
z_thread_copy_vgpr_to_global.MoveDstSliceWindow( z_thread_copy_vgpr_to_global.MoveDstSliceWindow(
z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5, z_grid_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3,
make_multi_index(0, 1, 0, 0, 0, 0, 0, 0, 0, 0)); make_multi_index(1, 0, 0, 0, 0, 0, 0, 0, 0, 0));
lse_thread_copy_global_to_vgpr.MoveSrcSliceWindow(lse_grid_desc_mb_m0_m1_m2_m3_m4, lse_thread_copy_global_to_vgpr.MoveSrcSliceWindow(lse_grid_desc_mb_m0_m1_m2_m3_m4,
make_multi_index(1, 0, 0, 0, 0, 0)); make_multi_index(1, 0, 0, 0, 0, 0));
y_threadwise_copy.MoveSrcSliceWindow(y_grid_desc_mblock_mperblock_oblock_operblock, y_threadwise_copy.MoveSrcSliceWindow(y_grid_desc_mblock_mperblock_oblock_operblock,
......
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