Skip to content

GitLab

Commit 56155968 authored by guangzlu's avatar guangzlu
Browse files

added dropout for bwd pt7

parent 02987ad6
Show whitespace changes
Inline Side-by-side
Showing with 153 additions and 97 deletions
include/ck/tensor_operation/gpu/device/impl/device_batched_multihead_attention_backward_xdl_cshuffle_v5.hpp
View file @ 56155968
......@@ -15,7 +15,7 @@
#include "ck/tensor_operation/gpu/device/masking_specialization.hpp"
#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_batched_multihead_attention_backward_xdl_cshuffle_pt5.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_batched_multihead_attention_backward_xdl_cshuffle_pt7.hpp"
#include "ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
......@@ -292,16 +292,6 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V2
// TODO: implement bias combination
static_assert(NumAcc0Bias == 0 && NumAcc0Bias == 0, "Bias addition is unimplemented");
#if 0
// TODO: use alias
static constexpr index_t NumDimGemm0M = NumDimM;
static constexpr index_t NumDimGemm0N = NumDimN;
static constexpr index_t NumDimGemm0K = NumDimK;
static constexpr index_t NumDimGemm1M = NumDimM;
static constexpr index_t NumDimGemm1N = NumDimO;
static constexpr index_t NumDimGemm1K = NumDimN;
#endif
using DeviceOp = DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V2;
static constexpr auto I0 = Number<0>{};
......@@ -1021,7 +1011,7 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V2
// Gemm1_K is split into Gemm1_K0/K1 where K1 is known at compile time, so we only need
// to concern Gemm0's loop
#if 1
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
{
ave_time = launch_kernel(integral_constant<bool, true>{});
......@@ -1030,7 +1020,7 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V2
{
ave_time = launch_kernel(integral_constant<bool, false>{});
}
#endif
return ave_time;
}
......@@ -1050,9 +1040,6 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V2
static bool IsSupportedArgument(const Argument& arg)
{
#if 0
arg.Print();
#endif
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
{
......
include/ck/tensor_operation/gpu/grid/gridwise_batched_multihead_attention_backward_xdl_cshuffle_pt7.hpp
View file @ 56155968
......@@ -127,39 +127,24 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V2
// C desc for source in blockwise copy
__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 N = z_grid_desc_m_n.GetLength(I1);
constexpr auto mfma = MfmaSelector<GemmDataType, MPerXdl, NPerXdl>::selected_mfma;
constexpr auto N3 = mfma.num_groups_per_blk;
constexpr auto N4 = mfma.num_input_blks;
constexpr auto N5 = mfma.group_size;
constexpr auto M3 = mfma.num_groups_per_blk;
constexpr auto M4 = mfma.num_input_blks;
constexpr auto M5 = mfma.group_size;
return transform_tensor_descriptor(
z_grid_desc_m_n,
make_tuple(make_unmerge_transform(
make_tuple(M / MPerBlock, MXdlPerWave, Gemm0MWaves, MPerXdl)),
make_unmerge_transform(
make_tuple(N / NPerBlock, NXdlPerWave, Gemm0NWaves, N3, N4, N5))),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2, 4, 6>{}, Sequence<1, 3, 5, 7, 8, 9>{}));
}
__host__ __device__ static constexpr auto
MakeZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(const index_t M, const index_t N)
{
constexpr auto mfma = MfmaSelector<GemmDataType, MPerXdl, NPerXdl>::selected_mfma;
constexpr auto N3 = mfma.num_groups_per_blk;
constexpr auto N4 = mfma.num_input_blks;
constexpr auto N5 = mfma.group_size;
return transform_tensor_descriptor(
make_naive_tensor_descriptor_packed(make_tuple(M, N)),
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_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, 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()
......@@ -467,8 +452,8 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V2
using DefaultBlock2CTileMap =
remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(KGridDesc_N_K{}))>;
using ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5 = remove_cvref_t<decltype(
MakeCGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(ZGridDesc_M_N{}))>;
using ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_M4_M5_N3 = remove_cvref_t<decltype(
MakeCGridDescriptor_M0_N0_M1_N1_M2_N2_M3_M4_M5_N3(ZGridDesc_M_N{}))>;
// S / dP Gemm (type 1 rcc)
struct Gemm0
......@@ -1183,8 +1168,8 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V2
const CElementwiseOperation& c_element_op,
const QGridDesc_K0_M_K1& q_grid_desc_k0_m_k1,
const KGridDesc_K0_N_K1& k_grid_desc_k0_n_k1,
const ZGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5&
z_grid_desc_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_m4_m5_n3,
const VGridDesc_N0_O_N1& v_grid_desc_n0_o_n1,
const YGridDescriptor_MBlock_MPerBlock_OBlock_OPerBlock&
y_grid_desc_mblock_mperblock_oblock_operblock,
......@@ -1194,6 +1179,9 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V2
const C0MatrixMask& c0_matrix_mask,
const float p_drop,
ck::philox& ph,
const index_t g_idx,
const index_t MRaw,
const index_t NRaw,
const index_t block_idx_n)
{
const FloatGemmAcc p_dropout = type_convert<FloatGemmAcc>(1.0f - p_drop);
......@@ -1571,33 +1559,30 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V2
// z vgpr copy to global
//
// 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
I1, // NBlockID
m0, // MRepeat
I1, // NRepeat
n0, // NRepeat
m1, // MWaveId
n1, // NWaveId
m2, // MPerXdl
m3, // NGroupNum
m4, // NInputNum
n2)); // registerNum
m2, // MGroupNum
m3, // MInputNum
m4, // registerNum
n2)); // NPerXdl
StaticBuffer<AddressSpaceEnum::Vgpr,
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>
z_tenor_buffer;
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);*/
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());
// z matrix global desc
// ignore = p_z_tmp_grid;
auto z_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>( // tmp buffer for shuffle
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_m_n_id = GetGemm0WaveMNIdx(wave_id[I2]); // I2: 0~63
......@@ -1605,13 +1590,13 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V2
auto z_thread_copy_vgpr_to_global = ThreadwiseTensorSliceTransfer_v1r3<
ushort,
ZDataType,
decltype(z_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5),
decltype(z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5),
decltype(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,
Sequence<I1, // MBlockId
I1, // NBlockID
m0, // MRepeat
I1, // NRepeat
n0, // NRepeat
m1, // MWaveId
n1, // NWaveId
m2, // MPerXdl
......@@ -1623,17 +1608,17 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V2
1, // DstScalarPerVector
InMemoryDataOperationEnum::Set,
1, // DstScalarStrideInVector
true>{z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
make_multi_index(block_work_idx_n, // MBlockId
0, // NBlockId
true>{z_grid_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3,
make_multi_index(num_gemm0_m_block_outer_loop - 1, // MBlockId
block_work_idx_n, // NBlockId
0, // mrepeat
0, // nrepeat
wave_id[I0], // MWaveId
wave_id[I1], // NWaveId
wave_m_n_id[I1], // MPerXdl
0, // group
wave_m_n_id[I0], // NInputIndex
0),
0, // MPerXdl
wave_m_n_id[I0], // group
0, // NInputIndex
wave_m_n_id[I1]),
tensor_operation::element_wise::PassThrough{}};
//
......@@ -1876,34 +1861,118 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V2
if(p_z_grid)
{
// P_dropped
static_for<0, n0, 1>{}([&](auto i) {
blockwise_dropout.template ApplyDropout<decltype(s_slash_p_thread_buf),
constexpr auto c_thread_lengths =
s_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2().GetLengths();
// 8d block_desc in block scope
constexpr auto c_block_lengths =
s_blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2().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_raw =
MRaw * NRaw * g_idx + m_global * NRaw + n_global; // unique element global 1d id
auto global_elem_id =
(global_elem_id_raw % 4) * MRaw + int(global_elem_id_raw / 4) * 4;
blockwise_dropout.template ApplyDropoutAttnBwdSaveZ<decltype(s_slash_p_thread_buf),
decltype(z_tenor_buffer),
true,
decltype(n0),
decltype(i)>(
s_slash_p_thread_buf, ph, z_tenor_buffer);
true>(
s_slash_p_thread_buf, ph, global_elem_id, z_tenor_buffer, MRaw);
z_thread_copy_vgpr_to_global.Run(
z_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
z_thread_copy_vgpr_to_global.Run(z_thread_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3,
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_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3,
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
{
ignore = z_grid_buf;
// 8d thread_desc in thread scope
constexpr auto c_thread_lengths =
s_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2().GetLengths();
// 8d block_desc in block scope
constexpr auto c_block_lengths =
s_blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2().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>{}));
// if(get_thread_global_1d_id()==0){
// printf("tid 0 m_global & n_global is %d & %d \n", m_global , n_global);
//}
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;
// if(get_thread_global_1d_id()==0){
// printf("tid 0 m_global & n_global is %d & %d \n", m_global , n_global);
// }
// if(get_thread_global_1d_id()==32){
// printf("tid 32 m_global & n_global is %d & %d \n", m_global , n_global);
// }
auto global_elem_id_raw =
MRaw * NRaw * g_idx + m_global * NRaw + n_global; // unique element global 1d id
auto global_elem_id =
(global_elem_id_raw % 4) * MRaw + int(global_elem_id_raw / 4) * 4;
// P_dropped
blockwise_dropout.template ApplyDropout<decltype(s_slash_p_thread_buf), true>(
s_slash_p_thread_buf, ph);
blockwise_dropout
.template ApplyDropoutAttnBwd<decltype(s_slash_p_thread_buf), true>(
s_slash_p_thread_buf, ph, global_elem_id, MRaw);
}
block_sync_lds(); // wait for gemm1 LDS read
......@@ -2160,8 +2229,8 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V2
make_multi_index(-1, 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, 1, 0, 0, 0, 0, 0, 0, 0, 0));
z_grid_desc_m0_n0_m1_n1_m2_n2_m3_m4_m5_n3,
make_multi_index(-1, 0, 0, 0, 0, 0, 0, 0, 0, 0));
} while(0 < gemm0_m_block_outer_index--); // end j loop
......
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