Skip to content

GitLab

Commit 83d13bff authored by letaoqin's avatar letaoqin
Browse files

Merge branch 'mha-train-bias' into mha-train-bias-bwd-type2

parents 5b9b87a1 dccaf0b2
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 438 additions and 4386 deletions
example/32_batched_gemm_scale_softmax_gemm/batched_multihead_attention_forward_v2.cpp
View file @ 83d13bff
......@@ -135,6 +135,7 @@ using DeviceGemmInstance =
8,
8,
true,
4,
S<16, 16, 1>, // B1BlockTransfer
S<0, 2, 1>,
S<0, 2, 1>,
......@@ -146,7 +147,8 @@ using DeviceGemmInstance =
1, // CShuffleNXdlPerWavePerShuffle
S<1, 64, 1, 4>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock
MaskingSpec, // MaskingSpecialization
4,
MaskingSpec, // MaskingSpecialization
Deterministic>;
#elif(DIM <= 64)
using DeviceGemmInstance =
......@@ -206,6 +208,7 @@ using DeviceGemmInstance =
8,
8,
true,
4,
S<16, 16, 1>, // B1BlockTransfer
S<0, 2, 1>,
S<0, 2, 1>,
......@@ -217,7 +220,8 @@ using DeviceGemmInstance =
2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock
MaskingSpec, // MaskingSpecialization
4,
MaskingSpec, // MaskingSpecialization
Deterministic>;
#elif(DIM <= 128)
using DeviceGemmInstance =
......@@ -277,6 +281,7 @@ using DeviceGemmInstance =
8,
8,
true,
4,
S<8, 32, 1>, // B1BlockTransfer
S<0, 2, 1>,
S<0, 2, 1>,
......@@ -288,7 +293,8 @@ using DeviceGemmInstance =
2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock
MaskingSpec, // MaskingSpecialization
4,
MaskingSpec, // MaskingSpecialization
Deterministic>;
#endif
......
example/32_batched_gemm_scale_softmax_gemm/grouped_multihead_attention_forward_v2.cpp
View file @ 83d13bff
......@@ -135,6 +135,7 @@ using DeviceGemmInstance =
8,
8,
true,
1,
S<16, 16, 1>, // B1BlockTransfer
S<0, 2, 1>,
S<0, 2, 1>,
......@@ -146,7 +147,8 @@ using DeviceGemmInstance =
1, // CShuffleNXdlPerWavePerShuffle
S<1, 64, 1, 4>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock
MaskingSpec, // MaskingSpecialization
1,
MaskingSpec, // MaskingSpecialization
Deterministic>;
#elif(DIM <= 64)
using DeviceGemmInstance =
......@@ -206,6 +208,7 @@ using DeviceGemmInstance =
8,
8,
true,
1,
S<16, 16, 1>, // B1BlockTransfer
S<0, 2, 1>,
S<0, 2, 1>,
......@@ -217,7 +220,8 @@ using DeviceGemmInstance =
2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock
MaskingSpec, // MaskingSpecialization
1,
MaskingSpec, // MaskingSpecialization
Deterministic>;
#elif(DIM <= 128)
using DeviceGemmInstance =
......@@ -277,6 +281,7 @@ using DeviceGemmInstance =
8,
8,
true,
1,
S<8, 32, 1>, // B1BlockTransfer
S<0, 2, 1>,
S<0, 2, 1>,
......@@ -288,7 +293,8 @@ using DeviceGemmInstance =
2, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 8>, // CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8, // CShuffleBlockTransferScalarPerVector_NPerBlock
MaskingSpec, // MaskingSpecialization
1,
MaskingSpec, // MaskingSpecialization
Deterministic>;
#endif
......
example/52_flash_atten_bias/batched_multihead_attention_bias_forward_v2.cpp
View file @ 83d13bff
......@@ -19,7 +19,7 @@ Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_batched_mha_fwd_xdl_cshuffle_v2r2.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_batched_mha_fwd_xdl_cshuffle_v2.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
......@@ -80,7 +80,7 @@ static constexpr bool Deterministic = false;
#if(DIM <= 32)
using DeviceGemmInstance =
ck::tensor_operation::device::DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2R2<
ck::tensor_operation::device::DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2<
NumDimG,
NumDimM,
NumDimN,
......@@ -153,7 +153,7 @@ using DeviceGemmInstance =
Deterministic>;
#elif(DIM <= 64)
using DeviceGemmInstance =
ck::tensor_operation::device::DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2R2<
ck::tensor_operation::device::DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2<
NumDimG,
NumDimM,
NumDimN,
......@@ -226,7 +226,7 @@ using DeviceGemmInstance =
Deterministic>;
#elif(DIM <= 128)
using DeviceGemmInstance =
ck::tensor_operation::device::DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2R2<
ck::tensor_operation::device::DeviceBatchedMultiheadAttentionForward_Xdl_CShuffle_V2<
NumDimG,
NumDimM,
NumDimN,
......
example/52_flash_atten_bias/grouped_multihead_attention_bias_forward_v2.cpp
View file @ 83d13bff
......@@ -19,7 +19,7 @@ Gemm + Softmax + Gemm fused operation. Computes C_g_m_o = Softmax(A_g_m_k * B0_g
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_mha_fwd_xdl_cshuffle_v2r2.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_mha_fwd_xdl_cshuffle_v2.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp"
......@@ -80,7 +80,7 @@ static constexpr bool Deterministic = false;
#if(DIM <= 32)
using DeviceGemmInstance =
ck::tensor_operation::device::DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2R2<
ck::tensor_operation::device::DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2<
NumDimG,
NumDimM,
NumDimN,
......@@ -153,7 +153,7 @@ using DeviceGemmInstance =
Deterministic>;
#elif(DIM <= 64)
using DeviceGemmInstance =
ck::tensor_operation::device::DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2R2<
ck::tensor_operation::device::DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2<
NumDimG,
NumDimM,
NumDimN,
......@@ -226,7 +226,7 @@ using DeviceGemmInstance =
Deterministic>;
#elif(DIM <= 128)
using DeviceGemmInstance =
ck::tensor_operation::device::DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2R2<
ck::tensor_operation::device::DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2<
NumDimG,
NumDimM,
NumDimN,
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_mha_fwd_xdl_cshuffle_v2.hpp
View file @ 83d13bff
......@@ -100,6 +100,13 @@ __global__ void
const long_index_t lse_batch_offset = __builtin_amdgcn_readfirstlane(static_cast<long_index_t>(
arg_ptr[group_id].compute_base_ptr_of_batch_.GetLSEBasePtr(g_idx)));
typename GridwiseGemm::D0sGridPointer p_d0s_grid = arg_ptr[group_id].p_d0s_grid_;
static_for<0, p_d0s_grid.Size(), 1>{}([&](auto In) {
const long_index_t d0_batch_offset =
__builtin_amdgcn_readfirstlane(static_cast<long_index_t>(
arg_ptr[group_id].compute_base_ptr_of_batch_.GetD0BasePtr(g_idx, In)));
p_d0s_grid(In) = p_d0s_grid(In) + d0_batch_offset;
});
if constexpr(Deterministic)
{
for(index_t i = 0; i < num_blocks_per_batch; i++)
......@@ -107,6 +114,7 @@ __global__ void
GridwiseGemm::template Run<HasMainKBlockLoop, IsDropout, IsLseStoring>(
arg_ptr[group_id].p_a_grid_ + a_batch_offset,
arg_ptr[group_id].p_b_grid_ + b_batch_offset,
p_d0s_grid,
arg_ptr[group_id].p_b1_grid_ + b1_batch_offset,
arg_ptr[group_id].p_c_grid_ + c_batch_offset,
arg_ptr[group_id].p_z_grid_ == nullptr
......@@ -124,6 +132,7 @@ __global__ void
c_element_op,
arg_ptr[group_id].a_grid_desc_ak0_m_ak1_,
arg_ptr[group_id].b_grid_desc_bk0_n_bk1_,
arg_ptr[group_id].d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_,
arg_ptr[group_id].b1_grid_desc_bk0_n_bk1_,
arg_ptr[group_id].c_grid_desc_mblock_mperblock_nblock_nperblock_,
arg_ptr[group_id].z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_,
......@@ -144,6 +153,7 @@ __global__ void
GridwiseGemm::template Run<HasMainKBlockLoop, IsDropout, IsLseStoring>(
arg_ptr[group_id].p_a_grid_ + a_batch_offset,
arg_ptr[group_id].p_b_grid_ + b_batch_offset,
p_d0s_grid,
arg_ptr[group_id].p_b1_grid_ + b1_batch_offset,
arg_ptr[group_id].p_c_grid_ + c_batch_offset,
arg_ptr[group_id].p_z_grid_ == nullptr ? nullptr
......@@ -160,6 +170,7 @@ __global__ void
c_element_op,
arg_ptr[group_id].a_grid_desc_ak0_m_ak1_,
arg_ptr[group_id].b_grid_desc_bk0_n_bk1_,
arg_ptr[group_id].d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_,
arg_ptr[group_id].b1_grid_desc_bk0_n_bk1_,
arg_ptr[group_id].c_grid_desc_mblock_mperblock_nblock_nperblock_,
arg_ptr[group_id].z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_,
......@@ -247,6 +258,7 @@ template <index_t NumDimG,
index_t BBlockTransferSrcScalarPerVector,
index_t BBlockTransferDstScalarPerVector_BK1,
bool BBlockLdsExtraN,
index_t Acc0BiasTransferSrcScalarPerVector,
typename B1BlockTransferThreadClusterLengths_BK0_N_BK1,
typename B1BlockTransferThreadClusterArrangeOrder,
typename B1BlockTransferSrcAccessOrder,
......@@ -258,6 +270,7 @@ template <index_t NumDimG,
index_t CShuffleNXdlPerWavePerShuffle,
typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
index_t Acc1BiasTransferSrcScalarPerVector,
MaskingSpecialization MaskingSpec,
bool Deterministic,
LoopScheduler LoopSched = LoopScheduler::Default>
......@@ -285,11 +298,11 @@ struct DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2
static_assert(NumDimG > 0 && NumDimM > 0 && NumDimN > 0 && NumDimK > 0 && NumDimO > 0,
"Number of dimension must be greater than 0");
static constexpr index_t NumAcc0Bias = Acc0BiasDataType::Size();
static constexpr index_t NumAcc1Bias = Acc1BiasDataType::Size();
static constexpr index_t NumD0Tensor = Acc0BiasDataType::Size();
static constexpr index_t NumD1Tensor = Acc1BiasDataType::Size();
// TODO ANT: implement bias combination
static_assert(NumAcc0Bias == 0 && NumAcc0Bias == 0, "Bias addition is unimplemented");
static_assert(NumD1Tensor == 0, "Acc1 Bias addition is unimplemented");
#if 0
// TODO ANT: use alias
......@@ -392,18 +405,44 @@ struct DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2
}
}
static auto MakeD0sGridDescriptor_M_N(
const std::vector<std::vector<ck::index_t>>& acc0_biases_gs_ms_ns_lengths,
const std::vector<std::vector<ck::index_t>>& acc0_biases_gs_ms_ns_strides)
{
return generate_tuple(
[&](auto i) {
return Transform::MakeCGridDescriptor_M_N(acc0_biases_gs_ms_ns_lengths[i],
acc0_biases_gs_ms_ns_strides[i]);
},
Number<NumD0Tensor>{});
}
static auto MakeD0sGridDescriptor_G_M_N(
const std::vector<std::vector<ck::index_t>>& acc0_biases_gs_ms_ns_lengths,
const std::vector<std::vector<ck::index_t>>& acc0_biases_gs_ms_ns_strides)
{
return generate_tuple(
[&](auto i) {
return Transform::MakeCGridDescriptor_G_M_N(acc0_biases_gs_ms_ns_lengths[i],
acc0_biases_gs_ms_ns_strides[i]);
},
Number<NumD0Tensor>{});
}
using AGridDesc_AK0_M_AK1 = decltype(MakeAGridDescriptor_AK0_M_AK1({}, {}));
using BGridDesc_BK0_N_BK1 = decltype(MakeBGridDescriptor_BK0_N_BK1({}, {}));
using D0sGridDesc_M_N = decltype(MakeD0sGridDescriptor_M_N({}, {}));
using B1GridDesc_BK0_N_BK1 = decltype(MakeB1GridDescriptor_BK0_N_BK1({}, {}));
using CGridDesc_M_N = decltype(Transform::MakeCGridDescriptor_M_N({}, {}));
using LSEGridDesc_M = decltype(MakeLSEGridDescriptor_M(1));
using ZGridDesc_M_N = decltype(MakeZGridDescriptor_M_N({}, {}));
using AGridDesc_G_M_K = decltype(Transform::MakeAGridDescriptor_G_M_K({}, {}));
using BGridDesc_G_N_K = decltype(Transform::MakeB0GridDescriptor_G_N_K({}, {}));
using B1GridDesc_G_N_K = decltype(Transform::MakeB1GridDescriptor_G_N_K({}, {}));
using CGridDesc_G_M_N = decltype(Transform::MakeCGridDescriptor_G_M_N({}, {}));
using ZGridDesc_G_M_N = decltype(Transform::MakeCGridDescriptor_G_M_N({}, {}));
using AGridDesc_G_M_K = decltype(Transform::MakeAGridDescriptor_G_M_K({}, {}));
using BGridDesc_G_N_K = decltype(Transform::MakeB0GridDescriptor_G_N_K({}, {}));
using D0sGridDesc_G_M_N = decltype(MakeD0sGridDescriptor_G_M_N({}, {}));
using B1GridDesc_G_N_K = decltype(Transform::MakeB1GridDescriptor_G_N_K({}, {}));
using CGridDesc_G_M_N = decltype(Transform::MakeCGridDescriptor_G_M_N({}, {}));
using ZGridDesc_G_M_N = decltype(Transform::MakeCGridDescriptor_G_M_N({}, {}));
constexpr static auto make_MaskOutPredicate()
{
......@@ -426,12 +465,14 @@ struct DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2
{
ComputeBasePtrOfStridedBatch(const AGridDesc_G_M_K& a_grid_desc_g_m_k,
const BGridDesc_G_N_K& b_grid_desc_g_n_k,
const D0sGridDesc_G_M_N& d0s_grid_desc_g_m_n,
const B1GridDesc_G_N_K& b1_grid_desc_g_n_k,
const CGridDesc_G_M_N& c_grid_desc_g_m_n,
const ZGridDesc_G_M_N& z_grid_desc_g_m_n,
index_t BatchStrideLSE)
: a_grid_desc_g_m_k_(a_grid_desc_g_m_k),
b_grid_desc_g_n_k_(b_grid_desc_g_n_k),
d0s_grid_desc_g_m_n_(d0s_grid_desc_g_m_n),
b1_grid_desc_g_n_k_(b1_grid_desc_g_n_k),
c_grid_desc_g_m_n_(c_grid_desc_g_m_n),
z_grid_desc_g_m_n_(z_grid_desc_g_m_n),
......@@ -449,6 +490,13 @@ struct DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2
return b_grid_desc_g_n_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
}
template <index_t I>
__host__ __device__ constexpr long_index_t GetD0BasePtr(index_t g_idx,
Number<I> d0_idx) const
{
return d0s_grid_desc_g_m_n_[d0_idx].CalculateOffset(make_multi_index(g_idx, 0, 0));
}
__host__ __device__ constexpr long_index_t GetB1BasePtr(index_t g_idx) const
{
return b1_grid_desc_g_n_k_.CalculateOffset(make_multi_index(g_idx, 0, 0));
......@@ -472,6 +520,7 @@ struct DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2
private:
AGridDesc_G_M_K a_grid_desc_g_m_k_;
BGridDesc_G_N_K b_grid_desc_g_n_k_;
D0sGridDesc_G_M_N d0s_grid_desc_g_m_n_;
B1GridDesc_G_N_K b1_grid_desc_g_n_k_;
CGridDesc_G_M_N c_grid_desc_g_m_n_;
ZGridDesc_G_M_N z_grid_desc_g_m_n_;
......@@ -482,6 +531,7 @@ struct DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2
// GridwiseGemm
using GridwiseGemm = GridwiseBatchedMultiheadAttentionForward_Xdl_CShuffle_V2<
ADataType, // TODO: distinguish A/B datatype
Acc0BiasDataType,
ZDataType,
GemmDataType,
GemmAccDataType,
......@@ -496,6 +546,7 @@ struct DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2
InMemoryDataOperationEnum::Set,
AGridDesc_AK0_M_AK1,
BGridDesc_BK0_N_BK1,
D0sGridDesc_M_N,
B1GridDesc_BK0_N_BK1,
CGridDesc_M_N,
ZGridDesc_M_N,
......@@ -531,6 +582,7 @@ struct DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2
BBlockTransferDstScalarPerVector_BK1,
true,
BBlockLdsExtraN,
Acc0BiasTransferSrcScalarPerVector,
B1BlockTransferThreadClusterLengths_BK0_N_BK1,
B1BlockTransferThreadClusterArrangeOrder,
B1BlockTransferSrcAccessOrder,
......@@ -543,6 +595,7 @@ struct DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2
CShuffleNXdlPerWavePerShuffle,
CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
CShuffleBlockTransferScalarPerVector_NPerBlock,
Acc1BiasTransferSrcScalarPerVector,
LoopSched,
Transform::matrix_padder.PadN,
MaskingSpec != MaskingSpecialization::MaskDisabled,
......@@ -555,6 +608,7 @@ struct DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2
// pointers
const ADataType* p_a_grid_;
const BDataType* p_b_grid_;
typename GridwiseGemm::D0sGridPointer p_d0s_grid_;
const B1DataType* p_b1_grid_;
CDataType* p_c_grid_;
ZDataType* p_z_grid_;
......@@ -563,6 +617,8 @@ struct DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2
// tensor descriptors for block/thread-wise copy
AGridDesc_AK0_M_AK1 a_grid_desc_ak0_m_ak1_;
BGridDesc_BK0_N_BK1 b_grid_desc_bk0_n_bk1_;
typename GridwiseGemm::D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5
d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5_;
B1GridDesc_BK0_N_BK1 b1_grid_desc_bk0_n_bk1_;
typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
c_grid_desc_mblock_mperblock_nblock_nperblock_;
......@@ -600,6 +656,9 @@ struct DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2
// for gridwise gemm check
CGridDesc_M_N c_grid_desc_m_n_;
// raw data
std::array<std::vector<ck::index_t>, NumD0Tensor> d0s_n_length_stride_;
};
// Argument
......@@ -628,28 +687,40 @@ struct DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2
b1_element_op_{b1_element_op},
c_element_op_{c_element_op}
{
ignore = p_acc1_biases_vec;
// TODO ANT: implement bias addition
group_count_ = problem_desc_vec.size();
if(!(group_count_ == p_a_vec.size() && group_count_ == p_b_vec.size() &&
group_count_ == p_b1_vec.size() && group_count_ == p_c_vec.size()))
group_count_ == p_b1_vec.size() && group_count_ == p_c_vec.size() &&
(group_count_ == p_acc0_biases_vec.size() || p_acc0_biases_vec.size() == 0)))
{
throw std::runtime_error("wrong! group_count_ != a/b/b1/c_vec.size");
}
if(!(p_acc0_biases_vec.size() == p_acc1_biases_vec.size()))
{
throw std::runtime_error("wrong! acc0_bias_vec.size != acc1_bias_vec.size");
}
grid_size_ = 0;
index_t z_random_matrix_offset = 0;
for(std::size_t i = 0; i < group_count_; i++)
{
const auto p_a_grid = static_cast<const ADataType*>(p_a_vec[i]);
const auto p_b_grid = static_cast<const BDataType*>(p_b_vec[i]);
const auto p_a_grid = static_cast<const ADataType*>(p_a_vec[i]);
const auto p_b_grid = static_cast<const BDataType*>(p_b_vec[i]);
const auto& problem_desc = problem_desc_vec[i];
std::array<std::vector<ck::index_t>, NumD0Tensor> d0s_n_length_stride;
typename GridwiseGemm::D0sGridPointer p_d0s_grid;
static_for<0, NumD0Tensor, 1>{}([&](auto j) {
using D0DataType = remove_cvref_t<tuple_element_t<j.value, Acc0BiasDataType>>;
// D0 pointer
p_d0s_grid(j) = static_cast<const D0DataType*>(p_acc0_biases_vec[i][j]);
// for check
d0s_n_length_stride[j].push_back(
problem_desc.acc0_biases_gs_ms_ns_lengths[j][NumDimG + NumDimM]);
d0s_n_length_stride[j].push_back(
problem_desc.acc0_biases_gs_ms_ns_strides[j][NumDimG + NumDimM]);
});
const auto p_b1_grid = static_cast<const B1DataType*>(p_b1_vec[i]);
const auto p_c_grid = static_cast<CDataType*>(p_c_vec[i]);
const auto p_z_grid = static_cast<ZDataType*>(p_z_vec[i]);
......@@ -660,12 +731,16 @@ struct DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2
is_lse_storing_ = false;
}
const auto& problem_desc = problem_desc_vec[i];
const auto a_grid_desc_ak0_m_ak1 = MakeAGridDescriptor_AK0_M_AK1(
problem_desc.a_gs_ms_ks_lengths, problem_desc.a_gs_ms_ks_strides);
const auto b_grid_desc_bk0_n_bk1 = MakeBGridDescriptor_BK0_N_BK1(
problem_desc.b0_gs_ns_ks_lengths, problem_desc.b0_gs_ns_ks_strides);
const D0sGridDesc_M_N d0s_grid_desc_m_n{
DeviceOp::MakeD0sGridDescriptor_M_N(problem_desc.acc0_biases_gs_ms_ns_lengths,
problem_desc.acc0_biases_gs_ms_ns_strides)};
const auto d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5 =
GridwiseGemm::MakeD0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(
d0s_grid_desc_m_n);
const auto b1_grid_desc_bk0_n_bk1 = MakeB1GridDescriptor_BK0_N_BK1(
problem_desc.b1_gs_os_ns_lengths, problem_desc.b1_gs_os_ns_strides);
const auto c_grid_desc_m_n = Transform::MakeCGridDescriptor_M_N(
......@@ -679,6 +754,9 @@ struct DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2
problem_desc.a_gs_ms_ks_lengths, problem_desc.a_gs_ms_ks_strides);
const auto b_grid_desc_g_n_k = Transform::MakeB0GridDescriptor_G_N_K(
problem_desc.b0_gs_ns_ks_lengths, problem_desc.b0_gs_ns_ks_strides);
const auto d0s_grid_desc_g_m_n = DeviceOp::MakeD0sGridDescriptor_G_M_N(
problem_desc.acc0_biases_gs_ms_ns_lengths,
problem_desc.acc0_biases_gs_ms_ns_strides);
const auto b1_grid_desc_g_n_k = Transform::MakeB1GridDescriptor_G_N_K(
problem_desc.b1_gs_os_ns_lengths, problem_desc.b1_gs_os_ns_strides);
const auto c_grid_desc_g_m_n = Transform::MakeCGridDescriptor_G_M_N(
......@@ -710,6 +788,7 @@ struct DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2
const auto compute_base_ptr_of_batch = ComputeBasePtrOfStridedBatch(
a_grid_desc_g_m_k,
b_grid_desc_g_n_k,
d0s_grid_desc_g_m_n,
b1_grid_desc_g_n_k,
c_grid_desc_g_m_n,
z_grid_desc_g_m_n,
......@@ -721,12 +800,12 @@ struct DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2
grid_size_ += grid_size_grp;
// for each group, make sure acc0_biases_gs_ms_ns_lengths.size() == NumAcc0Bias and
// for each group, make sure acc0_biases_gs_ms_ns_lengths.size() == NumD0Tensor and
// so on
if(!(problem_desc.acc0_biases_gs_ms_ns_lengths.size() == NumAcc0Bias &&
problem_desc.acc0_biases_gs_ms_ns_strides.size() == NumAcc0Bias &&
problem_desc.acc1_biases_gs_ms_os_lengths.size() == NumAcc1Bias &&
problem_desc.acc1_biases_gs_ms_os_strides.size() == NumAcc1Bias))
if(!(problem_desc.acc0_biases_gs_ms_ns_lengths.size() == NumD0Tensor &&
problem_desc.acc0_biases_gs_ms_ns_strides.size() == NumD0Tensor &&
problem_desc.acc1_biases_gs_ms_os_lengths.size() == NumD1Tensor &&
problem_desc.acc1_biases_gs_ms_os_strides.size() == NumD1Tensor))
{
throw std::runtime_error(
"wrong! number of biases in function argument does not "
......@@ -740,12 +819,14 @@ struct DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2
group_kernel_args_.push_back({p_a_grid,
p_b_grid,
p_d0s_grid,
p_b1_grid,
p_c_grid,
p_z_grid,
p_lse_grid,
a_grid_desc_ak0_m_ak1,
b_grid_desc_bk0_n_bk1,
d0s_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
b1_grid_desc_bk0_n_bk1,
c_grid_desc_mblock_mperblock_nblock_nperblock,
z_grid_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
......@@ -777,7 +858,8 @@ struct DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2
problem_desc.b1_gs_os_ns_strides[NumDimG + NumDimO + NumDimN - 1]},
{problem_desc.c_gs_ms_os_strides[NumDimG + NumDimM - 1],
problem_desc.c_gs_ms_os_strides[NumDimG + NumDimM + NumDimO - 1]},
c_grid_desc_m_n});
c_grid_desc_m_n,
d0s_n_length_stride});
}
is_dropout_ = p_dropout > 0.0; //
......@@ -997,6 +1079,19 @@ struct DeviceGroupedMultiheadAttentionForward_Xdl_CShuffle_V2
return false;
}
for(int In = 0; In < NumD0Tensor; In++)
{
if(device_arg.d0s_n_length_stride_[In][1] == 1 &&
device_arg.d0s_n_length_stride_[In][0] % Acc0BiasTransferSrcScalarPerVector != 0)
{
return false;
}
if(device_arg.d0s_n_length_stride_[In][1] != 1 &&
Acc0BiasTransferSrcScalarPerVector != 1)
{
return false;
}
}
// Check if having main loop
const auto K = kernel_arg.a_grid_desc_ak0_m_ak1_.GetLength(I0) *
kernel_arg.a_grid_desc_ak0_m_ak1_.GetLength(I2);
......
include/ck/tensor_operation/gpu/grid/gridwise_batched_mha_fwd_xdl_cshuffle_v2.hpp
View file @ 83d13bff
......@@ -25,6 +25,7 @@ namespace ck {
*
*/
template <typename FloatAB,
typename D0sDataType,
typename ZDataType,
typename FloatGemm,
typename FloatGemmAcc,
......@@ -39,6 +40,7 @@ template <typename FloatAB,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
typename AGridDesc_AK0_M_AK1,
typename BGridDesc_BK0_N_BK1,
typename D0sGridDesc_M_N,
typename B1GridDesc_BK0_N_BK1,
typename CGridDesc_M_N,
typename ZGridDesc_M_N,
......@@ -74,6 +76,7 @@ template <typename FloatAB,
index_t BBlockTransferDstScalarPerVector_BK1,
bool BThreadTransferSrcResetCoordinateAfterRun, // ignored
index_t BBlockLdsExtraN,
index_t D0BlockTransferSrcScalarPerVector,
typename B1BlockTransferThreadClusterLengths_BK0_N_BK1,
typename B1BlockTransferThreadClusterArrangeOrder,
typename B1BlockTransferSrcAccessOrder,
......@@ -86,6 +89,7 @@ template <typename FloatAB,
index_t CShuffleNXdlPerWavePerShuffle,
typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
index_t D1BlockTransferSrcScalarPerVector,
LoopScheduler LoopSched,
bool PadN,
bool MaskOutUpperTriangle,
......@@ -93,6 +97,11 @@ template <typename FloatAB,
PipelineVersion PipelineVer = PipelineVersion::v1>
struct GridwiseBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
{
static_assert(D0BlockTransferSrcScalarPerVector == 1 ||
D0BlockTransferSrcScalarPerVector == 2 ||
D0BlockTransferSrcScalarPerVector == 4,
"D0BlockTransferSrcScalarPerVector must be 1 or 2 or 4");
static constexpr index_t NumD0Tensor = D0sDataType::Size();
static_assert(LoopSched == LoopScheduler::Default,
"Non-default loop scheduler is currently not supported");
......@@ -407,6 +416,52 @@ struct GridwiseBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
c_grid_desc_m_n);
}
static constexpr auto MakeD0sGridPointer()
{
return generate_tuple(
[&](auto i) {
using D0DataType = remove_cvref_t<tuple_element_t<i.value, D0sDataType>>;
return static_cast<const D0DataType*>(nullptr);
},
Number<NumD0Tensor>{});
}
// D0 desc for source in blockwise copy
template <typename D0GridDesc_M_N>
__host__ __device__ static constexpr auto
MakeGemm0D0GridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(const D0GridDesc_M_N& d0_grid_desc_m_n)
{
const auto M = d0_grid_desc_m_n.GetLength(I0);
const auto N = d0_grid_desc_m_n.GetLength(I1);
constexpr auto mfma = MfmaSelector<FloatAB, 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(
d0_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>{}));
}
// D0s desc for source in blockwise copy
__host__ __device__ static constexpr auto
MakeD0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(const D0sGridDesc_M_N& ds_grid_desc_m_n)
{
return generate_tuple(
[&](auto i) {
return MakeGemm0D0GridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(ds_grid_desc_m_n[i]);
},
Number<NumD0Tensor>{});
}
using D0sGridPointer = decltype(MakeD0sGridPointer());
using D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5 = remove_cvref_t<decltype(
MakeD0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5(D0sGridDesc_M_N{}))>;
using CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(CGridDesc_M_N{}))>;
......@@ -465,6 +520,7 @@ struct GridwiseBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
typename C0MatrixMask>
__device__ static void Run(const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
D0sGridPointer p_d0s_grid,
const FloatAB* __restrict__ p_b1_grid,
FloatC* __restrict__ p_c_grid,
ZDataType* __restrict__ p_z_grid,
......@@ -477,6 +533,8 @@ struct GridwiseBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
const CElementwiseOperation& c_element_op,
const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
const D0sGridDescriptor_M0_N0_M1_N1_M2_N2_M3_N3_N4_N5&
d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
const B1GridDesc_BK0_N_BK1& b1_grid_desc_bk0_n_bk1,
const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
c_grid_desc_mblock_mperblock_nblock_nperblock,
......@@ -891,6 +949,65 @@ struct GridwiseBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
// gemm1 K loop
index_t gemm1_k_block_outer_index = 0;
const auto wave_id = GetGemm0WaveIdx();
const auto wave_m_n_id = GetGemm0WaveMNIdx(wave_id[I2]); // I2: 0~63
// bias (d matrix)
constexpr auto d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5 =
make_naive_tensor_descriptor_packed(make_tuple(I1, // MBlockId
I1, // NBlockId
m0, // MRepeat
n0, // NRepeat
m1, // MWaveId
n1, // NWaveId
m2, // MPerXdl
n2, // NGroupNum
n3, // NInputNum
n4)); // RegisterNum
auto d0s_threadwise_copy = generate_tuple(
[&](auto i) {
using D0DataType = remove_cvref_t<tuple_element_t<i.value, D0sDataType>>;
return ThreadwiseTensorSliceTransfer_v2<
D0DataType,
D0DataType,
decltype(d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i]),
decltype(d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5),
Sequence<I1, // MBlockId
I1, // NBlockID
m0, // MRepeat
n0, // NRepeat
m1, // MWaveId
n1, // NWaveId
m2, // MPerXdl
n2, // NGroupNum
n3, // NInputNum
n4>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7, 8, 9>,
9,
D0BlockTransferSrcScalarPerVector,
1,
false>(d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
make_multi_index(block_work_idx_m, // MBlockId
0, // 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)); // register number
},
Number<NumD0Tensor>{});
const auto d0s_grid_buf = generate_tuple(
[&](auto i) {
return make_dynamic_buffer<AddressSpaceEnum::Global>(
p_d0s_grid[i],
d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i].GetElementSpaceSize());
},
Number<NumD0Tensor>{});
// z is random number matrix for dropout verify
//
// z vgpr copy to global
......@@ -983,9 +1100,6 @@ struct GridwiseBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
static_cast<ushort*>(p_shared),
z_shuffle_block_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetElementSpaceSize());
const auto wave_id = GetGemm0WaveIdx();
const auto wave_m_n_id = GetGemm0WaveMNIdx(wave_id[I2]); // I2: 0~63
auto z_tmp_thread_copy_vgpr_to_lds = ThreadwiseTensorSliceTransfer_v1r3<
ushort,
ushort,
......@@ -1163,6 +1277,31 @@ struct GridwiseBatchedMultiheadAttentionForward_Xdl_CShuffle_V2
block_sync_lds(); // wait for lds read in gemm0 blockwise gemm
// add bias
static_for<0, NumD0Tensor, 1>{}([&](auto i) {
// get register
using D0DataType = remove_cvref_t<tuple_element_t<i.value, D0sDataType>>;
StaticBuffer<AddressSpaceEnum::Vgpr,
D0DataType,
d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5.GetElementSpaceSize(),
true>
d0_thread_buf;
// load data from global
d0s_threadwise_copy(i).Run(d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
d0s_grid_buf[i],
d0_thread_desc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5,
make_tuple(I0, I0, I0, I0, I0, I0, I0, I0, I0, I0),
d0_thread_buf);
// acc add bias
static_for<0, m0 * n0 * n2 * n4, 1>{}(
[&](auto j) { acc_thread_buf(j) += d0_thread_buf[j]; });
d0s_threadwise_copy(i).MoveSrcSliceWindow(
d0s_griddesc_m0_n0_m1_n1_m2_n2_m3_n3_n4_n5[i],
make_multi_index(0, 1, 0, 0, 0, 0, 0, 0, 0, 0));
});
// softmax
SoftmaxBuf& max = blockwise_softmax.max_value_buf;
SoftmaxBuf& sum = blockwise_softmax.sum_value_buf;
......
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