Skip to content

GitLab

"git@developer.sourcefind.cn:modelzoo/resnet50_tensorflow.git" did not exist on "3bb5dd6c0cba94ee6c81c35e225b0e1df7246436"
Commit 02987ad6 authored by guangzlu's avatar guangzlu
Browse files

add dropout to bwd pt6

parent 610e0ab0
Hide whitespace changes
Inline Side-by-side
Showing with 153 additions and 92 deletions
include/ck/tensor_operation/gpu/device/impl/device_batched_multihead_attention_backward_xdl_cshuffle_v4.hpp
View file @ 02987ad6
...@@ -16,7 +16,7 @@ ...@@ -16,7 +16,7 @@
#include "ck/tensor_operation/gpu/device/masking_specialization.hpp" #include "ck/tensor_operation/gpu/device/masking_specialization.hpp"
#include "ck/tensor_operation/gpu/device/matrix_padder.hpp" #include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_batched_multihead_attention_backward_xdl_cshuffle_pt4.hpp" #include "ck/tensor_operation/gpu/grid/gridwise_batched_multihead_attention_backward_xdl_cshuffle_pt6.hpp"
#include "ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp" #include "ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp"
#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"
...@@ -286,16 +286,6 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -286,16 +286,6 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
// TODO: implement bias combination // TODO: implement bias combination
static_assert(NumAcc0Bias == 0 && NumAcc0Bias == 0, "Bias addition is unimplemented"); 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_V1; using DeviceOp = DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1;
static constexpr auto I0 = Number<0>{}; static constexpr auto I0 = Number<0>{};
...@@ -1002,19 +992,8 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -1002,19 +992,8 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
arg.raw_lengths_mz_nz_kz_gemm1nz_[1]); 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
// to concern Gemm0's loop
#if 1
// if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
// {
// ave_time = launch_kernel(integral_constant<bool, true>{});
// }
// else
// {
// ave_time = launch_kernel(integral_constant<bool, false>{});
// }
ave_time = launch_kernel(integral_constant<bool, false>{}); ave_time = launch_kernel(integral_constant<bool, false>{});
#endif
return ave_time; return ave_time;
} }
...@@ -1034,9 +1013,6 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -1034,9 +1013,6 @@ struct DeviceBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
static bool IsSupportedArgument(const Argument& arg) static bool IsSupportedArgument(const Argument& arg)
{ {
#if 0
arg.Print();
#endif
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a")) if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
{ {
......
include/ck/tensor_operation/gpu/grid/gridwise_batched_multihead_attention_backward_xdl_cshuffle_pt6.hpp
View file @ 02987ad6
...@@ -113,23 +113,24 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -113,23 +113,24 @@ 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 +397,8 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -396,8 +397,8 @@ 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
...@@ -1235,8 +1236,8 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -1235,8 +1236,8 @@ 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,
...@@ -1246,6 +1247,9 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -1246,6 +1247,9 @@ 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);
...@@ -1594,33 +1598,30 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -1594,33 +1598,30 @@ 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, // MGroupNum
m3, // NGroupNum m3, // MInputNum
m4, // NInputNum m4, // registerNum
n2)); // registerNum n2)); // NPerXdl
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);*/
auto z_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>( // z matrix global desc
p_z_grid, z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5.GetElementSpaceSize()); // 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_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
...@@ -1628,13 +1629,13 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -1628,13 +1629,13 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1
auto z_thread_copy_vgpr_to_global = ThreadwiseTensorSliceTransfer_v1r3< auto z_thread_copy_vgpr_to_global = ThreadwiseTensorSliceTransfer_v1r3<
ushort, ushort,
ZDataType, ZDataType,
decltype(z_thread_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_n3_n4_n5), 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
...@@ -1646,17 +1647,17 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -1646,17 +1647,17 @@ 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(num_gemm0_m_block_outer_loop - 1, // 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{}};
// //
...@@ -1943,35 +1944,119 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -1943,35 +1944,119 @@ 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 =
blockwise_dropout.template ApplyDropout<decltype(s_slash_p_thread_buf), s_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2().GetLengths();
decltype(z_tenor_buffer),
true, // 8d block_desc in block scope
decltype(n0), constexpr auto c_block_lengths =
decltype(i)>( s_blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2().GetLengths();
s_slash_p_thread_buf, ph, z_tenor_buffer);
constexpr auto M0 = c_block_lengths[I0];
z_thread_copy_vgpr_to_global.Run( constexpr auto N0 = c_block_lengths[I1];
z_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5, constexpr auto M1 = c_block_lengths[I2];
make_tuple(I0, I0, I0, I0, I0, I0, I0, I0, I0, I0), constexpr auto N1 = c_block_lengths[I3];
z_tenor_buffer, constexpr auto M2 = c_block_lengths[I4];
z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5, constexpr auto M3 = c_block_lengths[I5];
z_grid_buf); constexpr auto M4 = c_block_lengths[I6];
z_thread_copy_vgpr_to_global.MoveDstSliceWindow( constexpr auto N2 = c_block_lengths[I7];
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)); // works like multi-dimension static_for (static_ford), but provides both the linear
}); // index as well as n-d index
z_thread_copy_vgpr_to_global.MoveDstSliceWindow( using Acc0TileIterator = SpaceFillingCurve<
z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5, decltype(c_thread_lengths),
make_multi_index(0, 0, 0, -n0.value, 0, 0, 0, 0, 0, 0)); 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>(
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_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_m4_m5_n3,
z_grid_buf);
} }
else else
{ {
ignore = z_grid_buf; 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 // P_dropped
blockwise_dropout.template ApplyDropout<decltype(s_slash_p_thread_buf), true>( blockwise_dropout
s_slash_p_thread_buf, ph); .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 block_sync_lds(); // wait for gemm1 LDS read
...@@ -2143,8 +2228,8 @@ struct GridwiseBatchedMultiheadAttentionBackward_Xdl_CShuffle_V1 ...@@ -2143,8 +2228,8 @@ 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