Skip to content

GitLab

Unverified Commit 13587ab3 authored by arai713's avatar arai713 Committed by GitHub
Browse files

Merge branch 'develop' into gridwise_2d

parents 7e44fd84 685860c2
Hide whitespace changes
Inline Side-by-side
Showing with 1424 additions and 105 deletions
include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp → include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r_xdl_cshuffle.hpp
View file @ 13587ab3
......@@ -13,7 +13,7 @@
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d_multiple_r.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_multiple_r.hpp"
#include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
......@@ -101,9 +101,9 @@ struct ComputePtrOffsetOfStridedBatch
* \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2
* tiles from different matrices. Keep in mind that these 2 matrices can share the same grid
* descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link
* device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for \link
* DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the computing of
* pointer offset into \p ComputePtrOffsetOfStridedBatch.
* impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for
* \link DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the
* computing of pointer offset into \p ComputePtrOffsetOfStridedBatch.
*
* \note \p Block2ETileMap allows customized mapping between a workgroup and the C-tile it computes.
* Together with \p ComputePtrOffsetOfBatch, we can reuse GridwiseGemm (and GridwiseGemm fusion ) to
......
include/ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp → include/ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp
View file @ 13587ab3
......@@ -90,9 +90,9 @@ struct ComputePtrOffsetOfStridedBatch
* \note Using \p ComputePtrOffsetOfBatch gives us the flexibility that 2 workgroups can compute 2
* tiles from different matrices. Keep in mind that these 2 matrices can share the same grid
* descriptor (like in BatchedGEMM), or use their own grid descriptors (in GroupedGemm). \link
* device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for \link
* DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the computing of
* pointer offset into \p ComputePtrOffsetOfStridedBatch.
* impl/device_conv3d_fwd_xdl_ndhwc_kzyxc_ndhwk.hpp kernel_gemm_xdlops_v2r3_for_conv3d \endlink for
* \link DeviceConv3d \endlink uses the same concept, but currently does NOT encapsulate the
* computing of pointer offset into \p ComputePtrOffsetOfStridedBatch.
*
* \note \p Block2ETileMap allows customized mapping between a workgroup and the C-tile it computes.
* Together with \p ComputePtrOffsetOfBatch, we can reuse GridwiseGemm (and GridwiseGemm fusion ) to
......
include/ck/tensor_operation/gpu/device/device_multiple_reduce_multiblock.hpp → include/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_multiblock.hpp
View file @ 13587ab3
......@@ -11,7 +11,7 @@
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/tensor_operation/gpu/device/device_multiple_reduce.hpp"
#include "ck/tensor_operation/gpu/device/device_reduce_common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_multiblock.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_set_multiple_buffer_value.hpp"
......
include/ck/tensor_operation/gpu/device/device_multiple_reduce_threadwise.hpp → include/ck/tensor_operation/gpu/device/impl/device_multiple_reduce_threadwise.hpp
View file @ 13587ab3
......@@ -11,7 +11,7 @@
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/tensor_operation/gpu/device/device_multiple_reduce.hpp"
#include "ck/tensor_operation/gpu/device/device_reduce_common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_2d_multiple_reduction_threadwise.hpp"
#include "ck/host_utility/kernel_launch.hpp"
......
include/ck/tensor_operation/gpu/device/device_layernorm_impl.hpp → include/ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp
View file @ 13587ab3
......@@ -9,7 +9,7 @@
#include "ck/utility/reduction_operator.hpp"
#include "ck/tensor_operation/gpu/device/device_normalization.hpp"
#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
#include "ck/tensor_operation/gpu/device/device_reduce_common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_layernorm_welford_variance.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp"
#include "ck/host_utility/device_prop.hpp"
......@@ -75,14 +75,14 @@ template <typename XDataType,
index_t BetaSrcVectorDim,
index_t BetaSrcVectorSize,
index_t YDstVectorSize>
struct DeviceLayernormImpl : public DeviceLayernorm<XDataType,
GammaDataType,
BetaDataType,
AccDataType,
YDataType,
AccElementwiseOperation,
Rank,
NumReduceDim>
struct DeviceNormalizationImpl : public DeviceNormalization<XDataType,
GammaDataType,
BetaDataType,
AccDataType,
YDataType,
AccElementwiseOperation,
Rank,
NumReduceDim>
{
static_assert(
((GammaSrcVectorDim == 0 && MThreadSliceSize % GammaSrcVectorSize == 0) ||
......@@ -452,7 +452,7 @@ struct DeviceLayernormImpl : public DeviceLayernorm<XDataType,
auto str = std::stringstream();
// clang-format off
str << "DeviceLayernormImpl<" << BlockSize << ",";
str << "DeviceNormalizationImpl<" << BlockSize << ",";
str << "M_C" << MThreadClusterSize << "_S" << MThreadSliceSize << ",";
str << "K_C" << KThreadClusterSize << "_S" << KThreadSliceSize << ",";
str << "XYSrcVectorDim_" << XYSrcVectorDim << ",";
......
include/ck/tensor_operation/gpu/device/device_reduce_multiblock.hpp → include/ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp
View file @ 13587ab3
......@@ -11,7 +11,7 @@
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
#include "ck/tensor_operation/gpu/device/device_reduce_common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_2d_reduction_multiblock.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp"
#include "ck/host_utility/device_prop.hpp"
......
include/ck/tensor_operation/gpu/device/device_reduce_threadwise.hpp → include/ck/tensor_operation/gpu/device/impl/device_reduce_threadwise.hpp
View file @ 13587ab3
......@@ -9,7 +9,7 @@
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
#include "ck/tensor_operation/gpu/device/device_reduce_common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_2d_reduction_multiblock.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_2d_reduction_threadwise.hpp"
......
include/ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp
View file @ 13587ab3
......@@ -10,8 +10,8 @@
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
#include "ck/tensor_operation/gpu/device/device_reduce_multiblock.hpp"
#include "ck/tensor_operation/gpu/device/device_reduce_common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_softmax.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp"
#include "ck/host_utility/device_prop.hpp"
......
include/ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp
View file @ 13587ab3
......@@ -209,6 +209,8 @@ struct BlockToCTileMap_KSplit_M00_N0_M01Adapt
const auto M0 = math::integer_divide_ceil(c_grid_desc_m_n_.GetLength(I0), MPerBlock);
const auto N0 = math::integer_divide_ceil(c_grid_desc_m_n_.GetLength(I1), NPerBlock);
block_1d_id = block_1d_id % (M0 * N0 * KSplit_); // hide groups
const index_t idx_ksplit = block_1d_id / (M0 * N0);
block_1d_id = block_1d_id % (M0 * N0);
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_split_k_multiple_d_xdl_cshuffle.hpp 0 → 100644
View file @ 13587ab3
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/common_header.hpp"
#include "ck/tensor_description/multi_index_transform_helper.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp"
#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v6r1.hpp"
#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp"
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
namespace ck {
// GEMM:
// input : A[M, K]
// input : B[N, K]
// input : D0[M, N], D1[M, N], ...
// output : E[M, N]
// C = a_op(A) * b_op(B)
// E = cde_op(C, D0, D1, ...)
// Assume:
// D0, D1, ... and E have the same layout
template <typename ABDataType, // FIXME: don't assume A/B have same datatype
typename AccDataType,
typename CShuffleDataType,
typename DsDataType,
typename EDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
InMemoryDataOperationEnum EGlobalMemoryDataOperation,
typename AGridDesc_M_K,
typename BGridDesc_N_K,
typename DsGridDesc_M_N,
typename EGridDesc_M_N,
index_t NumGemmKPrefetchStage,
index_t BlockSize,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t AK1Value,
index_t BK1Value,
index_t MPerXdl,
index_t NPerXdl,
index_t MXdlPerWave,
index_t NXdlPerWave,
typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
index_t ABlockTransferSrcVectorDim,
index_t ABlockTransferSrcScalarPerVector,
index_t ABlockTransferDstScalarPerVector_AK1,
bool AThreadTransferSrcResetCoordinateAfterRun,
index_t ABlockLdsExtraM,
typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
index_t BBlockTransferSrcVectorDim,
index_t BBlockTransferSrcScalarPerVector,
index_t BBlockTransferDstScalarPerVector_BK1,
bool BThreadTransferSrcResetCoordinateAfterRun,
index_t BBlockLdsExtraN,
index_t CShuffleMXdlPerWavePerShuffle,
index_t CShuffleNXdlPerWavePerShuffle,
typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CDEShuffleBlockTransferScalarPerVector_NPerBlock,
LoopScheduler LoopSched>
struct GridwiseGemmSplitKMultipleD_xdl_cshuffle
{
static constexpr index_t NumDTensor = DsDataType::Size();
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
static constexpr auto I4 = Number<4>{};
static constexpr auto I5 = Number<5>{};
static constexpr auto I6 = Number<6>{};
static constexpr auto I7 = Number<7>{};
// K1 should be Number<...>
static constexpr auto AK1 = Number<AK1Value>{};
static constexpr auto BK1 = Number<BK1Value>{};
static constexpr auto AK0PerBlock = Number<KPerBlock / AK1Value>{};
static constexpr auto BK0PerBlock = Number<KPerBlock / BK1Value>{};
using ThisThreadBlock = ThisThreadBlock<BlockSize>;
using GridwiseGemmPipe = GridwiseGemmPipeline_v1<NumGemmKPrefetchStage>;
__host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
{
// A matrix in LDS memory, src of blockwise copy
return make_naive_tensor_descriptor(
make_tuple(AK0PerBlock, Number<MPerBlock>{}, AK1),
make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
}
__host__ __device__ static constexpr auto GetABlockDescriptor_AKB_AK0PerBlock_MPerBlock_AK1()
{
// A matrix in LDS memory, dst of blockwise copy
return make_naive_tensor_descriptor(
make_tuple(I1, AK0PerBlock, Number<MPerBlock>{}, AK1),
make_tuple(AK0PerBlock * Number<MPerBlock + ABlockLdsExtraM>{} * AK1,
Number<MPerBlock + ABlockLdsExtraM>{} * AK1,
AK1,
I1));
}
__host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
{
// B matrix in LDS memory, src of blockwise copy
return make_naive_tensor_descriptor(
make_tuple(BK0PerBlock, Number<NPerBlock>{}, BK1),
make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
}
__host__ __device__ static constexpr auto GetBBlockDescriptor_BKB_BK0PerBlock_NPerBlock_BK1()
{
// B matrix in LDS memory, dst of blockwise copy
return make_naive_tensor_descriptor(
make_tuple(I1, BK0PerBlock, Number<NPerBlock>{}, BK1),
make_tuple(BK0PerBlock * Number<NPerBlock + BBlockLdsExtraN>{} * BK1,
Number<NPerBlock + BBlockLdsExtraN>{} * BK1,
BK1,
I1));
}
__host__ __device__ static constexpr auto
GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
{
constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
make_naive_tensor_descriptor_packed(
make_tuple(I1,
Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
I1,
Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
}
// ck::Tuple<const D0DataType*, const D1DataType*, ...>
static constexpr auto MakeDsGridPointer()
{
return generate_tuple(
[&](auto i) {
using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
return static_cast<const DDataType*>(nullptr);
},
Number<NumDTensor>{});
}
__host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
{
// LDS allocation for A and B: be careful of alignment
constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
// lds max alignment
constexpr auto max_lds_align = math::lcm(AK1, BK1);
constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
// LDS allocation for C shuffle in LDS
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
constexpr auto c_block_size =
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
sizeof(ABDataType),
c_block_size * sizeof(CShuffleDataType));
}
// A desc for source in blockwise copy
__host__ __device__ static constexpr auto
MakeDefaultAGridDescriptor_AKB_AK0_M_AK1(const AGridDesc_M_K& a_grid_desc_m_k,
const int split_k)
{
const auto MRaw = a_grid_desc_m_k.GetLength(I0);
const auto KRaw = a_grid_desc_m_k.GetLength(I1);
const index_t AK0 =
(math::integer_divide_ceil(KRaw, KPerBlock * split_k) * KPerBlock) / AK1;
const index_t K = split_k * AK0 * AK1;
const auto KPad = K - KRaw;
const auto a_grid_desc_m_kpad = transform_tensor_descriptor(
a_grid_desc_m_k,
make_tuple(make_pass_through_transform(MRaw), make_right_pad_transform(KRaw, KPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return transform_tensor_descriptor(
a_grid_desc_m_kpad,
make_tuple(make_unmerge_transform(make_tuple(split_k, AK0, AK1)),
make_pass_through_transform(MRaw)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
}
// B desc for source in blockwise copy
__host__ __device__ static constexpr auto
MakeDefaultBGridDescriptor_BKB_BK0_N_BK1(const BGridDesc_N_K& b_grid_desc_n_k,
const int split_k)
{
const auto NRaw = b_grid_desc_n_k.GetLength(I0);
const auto KRaw = b_grid_desc_n_k.GetLength(I1);
const index_t BK0 =
(math::integer_divide_ceil(KRaw, KPerBlock * split_k) * KPerBlock) / BK1;
const index_t K = split_k * BK0 * BK1;
const auto KPad = K - KRaw;
const auto b_grid_desc_n_kpad = transform_tensor_descriptor(
b_grid_desc_n_k,
make_tuple(make_pass_through_transform(NRaw), make_right_pad_transform(KRaw, KPad)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return transform_tensor_descriptor(
b_grid_desc_n_kpad,
make_tuple(make_unmerge_transform(make_tuple(split_k, BK0, BK1)),
make_pass_through_transform(NRaw)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
}
// E desc for destination in blockwise copy
template <typename EGridDescriptor_M_N>
__host__ __device__ static constexpr auto MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
const EGridDescriptor_M_N& e_grid_desc_m_n)
{
const auto M = e_grid_desc_m_n.GetLength(I0);
const auto N = e_grid_desc_m_n.GetLength(I1);
const auto MBlock = M / MPerBlock;
const auto NBlock = N / NPerBlock;
const auto e_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
e_grid_desc_m_n,
make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
return e_grid_desc_mblock_mperblock_nblock_nperblock;
}
// Ds desc for source in blockwise copy
template <typename DsGridDescriptor_M_N>
__host__ __device__ static constexpr auto
MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
const DsGridDescriptor_M_N& ds_grid_desc_m_n)
{
return generate_tuple(
[&](auto i) {
return MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[i]);
},
Number<NumDTensor>{});
}
// return block_id to E matrix tile idx (m0, n0) mapping
__host__ __device__ static constexpr auto
MakeDefaultBlock2ETileMap(const EGridDesc_M_N& e_grid_desc_m_n, const int split_k)
{
return BlockToCTileMap_KSplit_M00_N0_M01Adapt<MPerBlock, NPerBlock, EGridDesc_M_N>(
e_grid_desc_m_n, 8, split_k);
}
// block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
template <typename AGridDesc_AKB_AK0_M_AK1,
typename BGridDesc_BKB_BK0_N_BK1,
typename Block2ETileMap>
__host__ __device__ static constexpr bool
CheckValidity(const AGridDesc_AKB_AK0_M_AK1& a_grid_desc_akb_ak0_m_ak1,
const BGridDesc_BKB_BK0_N_BK1& b_grid_desc_bkb_bk0_n_bk1,
const DsGridDesc_M_N& ds_grid_desc_m_n,
const EGridDesc_M_N& e_grid_desc_m_n,
const Block2ETileMap& block_2_etile_map)
{
static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
(NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
"Invalid tuning param!");
const auto M = a_grid_desc_akb_ak0_m_ak1.GetLength(I2);
const auto N = b_grid_desc_bkb_bk0_n_bk1.GetLength(I2);
const auto K =
a_grid_desc_akb_ak0_m_ak1.GetLength(I1) * a_grid_desc_akb_ak0_m_ak1.GetLength(I3);
if(K != b_grid_desc_bkb_bk0_n_bk1.GetLength(I1) * b_grid_desc_bkb_bk0_n_bk1.GetLength(I3))
{
return false;
}
if(a_grid_desc_akb_ak0_m_ak1.GetLength(I0) != b_grid_desc_bkb_bk0_n_bk1.GetLength(I0))
{
return false;
}
// check consistency of desc
if(!(M == e_grid_desc_m_n.GetLength(I0) && N == e_grid_desc_m_n.GetLength(I1)))
{
return false;
}
bool valid = true;
static_for<0, NumDTensor, 1>{}([&](auto i) {
valid = valid && (M == ds_grid_desc_m_n[i].GetLength(I0) &&
N == ds_grid_desc_m_n[i].GetLength(I1));
});
if(!valid)
{
return false;
}
// check tile size
if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
{
return false;
}
// check gridwise gemm pipeline
const auto num_k_loop = K / KPerBlock;
if(!GridwiseGemmPipe::IsSupported(num_k_loop))
{
return false;
}
// check block-to-E-tile
if(!block_2_etile_map.CheckValidity(e_grid_desc_m_n))
{
return false;
}
// TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
// check tensor size: cannot be larger than 2GB each
constexpr long_index_t TwoGB = (long_index_t{1} << 31);
if(!(a_grid_desc_akb_ak0_m_ak1.GetElementSpaceSize() * sizeof(ABDataType) <= TwoGB &&
b_grid_desc_bkb_bk0_n_bk1.GetElementSpaceSize() * sizeof(ABDataType) <= TwoGB &&
e_grid_desc_m_n.GetElementSpaceSize() * sizeof(EDataType) <= TwoGB))
{
return false;
}
return true;
}
__host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
{
const index_t num_loop = K / KPerBlock;
return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
}
using DefaultAGridDesc_AK0_M_AK1 =
remove_cvref_t<decltype(MakeDefaultAGridDescriptor_AKB_AK0_M_AK1(AGridDesc_M_K{}, 1))>;
using DefaultBGridDesc_BK0_N_BK1 =
remove_cvref_t<decltype(MakeDefaultBGridDescriptor_BKB_BK0_N_BK1(BGridDesc_N_K{}, 1))>;
using EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}))>;
using DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(DsGridDesc_M_N{}))>;
using DefaultBlock2ETileMap =
remove_cvref_t<decltype(MakeDefaultBlock2ETileMap(EGridDesc_M_N{}, 1))>;
using DsGridPointer = decltype(MakeDsGridPointer());
template <bool HasMainKBlockLoop,
typename AGridDesc_AKB_AK0_M_AK1,
typename BGridDesc_BKB_BK0_N_BK1,
typename Block2ETileMap>
__device__ static void Run(const ABDataType* __restrict__ p_a_grid,
const ABDataType* __restrict__ p_b_grid,
DsGridPointer p_ds_grid,
EDataType* __restrict__ p_e_grid,
void* __restrict__ p_shared,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const CDEElementwiseOperation& cde_element_op,
const AGridDesc_AKB_AK0_M_AK1& a_grid_desc_akb_ak0_m_ak1,
const BGridDesc_BKB_BK0_N_BK1& b_grid_desc_bkb_bk0_n_bk1,
const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
ds_grid_desc_mblock_mperblock_nblock_nperblock,
const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
e_grid_desc_mblock_mperblock_nblock_nperblock,
const Block2ETileMap& block_2_etile_map)
{
const auto block_work_idx =
block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
if(block_work_idx[Number<0>{}] == 0)
{
Run0<HasMainKBlockLoop>(p_a_grid,
p_b_grid,
p_ds_grid,
p_e_grid,
p_shared,
a_element_op,
b_element_op,
cde_element_op,
a_grid_desc_akb_ak0_m_ak1,
b_grid_desc_bkb_bk0_n_bk1,
ds_grid_desc_mblock_mperblock_nblock_nperblock,
e_grid_desc_mblock_mperblock_nblock_nperblock,
block_2_etile_map);
}
else
{
Run1<HasMainKBlockLoop>(p_a_grid,
p_b_grid,
p_e_grid,
p_shared,
a_element_op,
b_element_op,
a_grid_desc_akb_ak0_m_ak1,
b_grid_desc_bkb_bk0_n_bk1,
ds_grid_desc_mblock_mperblock_nblock_nperblock,
e_grid_desc_mblock_mperblock_nblock_nperblock,
block_2_etile_map);
}
}
template <bool HasMainKBlockLoop,
typename AGridDesc_AKB_AK0_M_AK1,
typename BGridDesc_BKB_BK0_N_BK1,
typename Block2ETileMap>
__device__ static void Run0(const ABDataType* __restrict__ p_a_grid,
const ABDataType* __restrict__ p_b_grid,
DsGridPointer p_ds_grid,
EDataType* __restrict__ p_e_grid,
void* __restrict__ p_shared,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const CDEElementwiseOperation& cde_element_op,
const AGridDesc_AKB_AK0_M_AK1& a_grid_desc_akb_ak0_m_ak1,
const BGridDesc_BKB_BK0_N_BK1& b_grid_desc_bkb_bk0_n_bk1,
const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
ds_grid_desc_mblock_mperblock_nblock_nperblock,
const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
e_grid_desc_mblock_mperblock_nblock_nperblock,
const Block2ETileMap& block_2_etile_map)
{
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_grid, a_grid_desc_akb_ak0_m_ak1.GetElementSpaceSize());
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_b_grid, b_grid_desc_bkb_bk0_n_bk1.GetElementSpaceSize());
const auto ds_grid_buf = generate_tuple(
[&](auto i) {
return make_dynamic_buffer<AddressSpaceEnum::Global>(
p_ds_grid[i],
ds_grid_desc_mblock_mperblock_nblock_nperblock[i].GetElementSpaceSize());
},
Number<NumDTensor>{});
auto e_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_e_grid, e_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
// divide block work by [M, N]
const auto block_work_idx =
block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
if(!block_2_etile_map.ValidCTileIndex(
make_tuple(block_work_idx[I1], block_work_idx[I2]),
make_tuple(e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
{
return;
}
// HACK: this force m/n_block_data_idx_on_grid into SGPR
const index_t k_batch_id = block_work_idx[I0];
const index_t m_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_work_idx[I1] * MPerBlock);
const index_t n_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_work_idx[I2] * NPerBlock);
// lds max alignment
constexpr auto max_lds_align = math::lcm(AK1, BK1);
// A matrix in LDS memory, dst of blockwise copy
constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
constexpr auto a_block_desc_akb_ak0_m_ak1 =
GetABlockDescriptor_AKB_AK0PerBlock_MPerBlock_AK1();
// B matrix in LDS memory, dst of blockwise copy
constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
constexpr auto b_block_desc_bkb_bk0_n_bk1 =
GetBBlockDescriptor_BKB_BK0PerBlock_NPerBlock_BK1();
// A matrix blockwise copy
auto a_blockwise_copy =
ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
AElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum::Set,
Sequence<I1, AK0PerBlock, MPerBlock, AK1>,
ABlockTransferThreadClusterLengths_AK0_M_AK1,
ABlockTransferThreadClusterArrangeOrder,
ABDataType,
ABDataType,
decltype(a_grid_desc_akb_ak0_m_ak1),
decltype(a_block_desc_akb_ak0_m_ak1),
ABlockTransferSrcAccessOrder,
Sequence<0, 2, 1, 3>,
ABlockTransferSrcVectorDim,
3,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_AK1,
1,
1,
AThreadTransferSrcResetCoordinateAfterRun,
true,
NumGemmKPrefetchStage>(
a_grid_desc_akb_ak0_m_ak1,
make_multi_index(k_batch_id, 0, m_block_data_idx_on_grid, 0),
a_element_op,
a_block_desc_akb_ak0_m_ak1,
make_multi_index(0, 0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
// B matrix blockwise copy
auto b_blockwise_copy =
ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
BElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum::Set,
Sequence<I1, BK0PerBlock, NPerBlock, BK1>,
BBlockTransferThreadClusterLengths_BK0_N_BK1,
BBlockTransferThreadClusterArrangeOrder,
ABDataType,
ABDataType,
decltype(b_grid_desc_bkb_bk0_n_bk1),
decltype(b_block_desc_bkb_bk0_n_bk1),
BBlockTransferSrcAccessOrder,
Sequence<0, 2, 1, 3>,
BBlockTransferSrcVectorDim,
3,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_BK1,
1,
1,
BThreadTransferSrcResetCoordinateAfterRun,
true,
NumGemmKPrefetchStage>(
b_grid_desc_bkb_bk0_n_bk1,
make_multi_index(k_batch_id, 0, n_block_data_idx_on_grid, 0),
b_element_op,
b_block_desc_bkb_bk0_n_bk1,
make_multi_index(0, 0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
// GEMM definition
// c_mtx += transpose(a_mtx) * b_mtx
// a_mtx[K0PerBlock, MPerBlock] is in LDS
// b_mtx[K0PerBlock, NPerBlock] is in LDS
// c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
// register
// sanity check
constexpr index_t KPack =
math::max(math::lcm(AK1, BK1),
MfmaSelector<ABDataType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
BlockSize,
ABDataType,
AccDataType,
decltype(a_block_desc_ak0_m_ak1),
decltype(b_block_desc_bk0_n_bk1),
MPerXdl,
NPerXdl,
MXdlPerWave,
NXdlPerWave,
KPack,
LoopSched>();
auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
// LDS allocation for A and B: be careful of alignment
constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<ABDataType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<ABDataType*>(p_shared) + a_block_space_size_aligned,
b_block_desc_bk0_n_bk1.GetElementSpaceSize());
constexpr auto a_block_slice_copy_step = make_multi_index(0, KPerBlock / AK1, 0, 0);
constexpr auto b_block_slice_copy_step = make_multi_index(0, KPerBlock / BK1, 0, 0);
// gridwise GEMM pipeline
const auto gridwise_gemm_pipeline =
GridwiseGemmPipeline_v1_Selector<NumGemmKPrefetchStage, LoopSched>();
const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
(a_grid_desc_akb_ak0_m_ak1.GetLength(I1) * a_grid_desc_akb_ak0_m_ak1.GetLength(I3)) /
KPerBlock);
gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_akb_ak0_m_ak1,
a_block_desc_akb_ak0_m_ak1,
a_blockwise_copy,
a_grid_buf,
a_block_buf,
a_block_slice_copy_step,
b_grid_desc_bkb_bk0_n_bk1,
b_block_desc_bkb_bk0_n_bk1,
b_blockwise_copy,
b_grid_buf,
b_block_buf,
b_block_slice_copy_step,
blockwise_gemm,
c_thread_buf,
num_k_block_main_loop);
// shuffle C and write out
{
static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
"wrong!");
constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
// TODO: hacky, fix it!
constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
// TODO: hacky, fix it!
// c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<CShuffleDataType*>(p_shared),
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
make_tuple(
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
M1, // M1 = MWave
M2, // M2 * M3 * M4 = MPerXdl
M3,
M4)),
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
N1, // N1 = NWave
N2))), // N2 = NPerXdl
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(
Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
// calculate origin of thread output tensor on global memory
// blockwise GEMM c matrix starting index
const auto c_thread_mtx_on_block =
blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
make_tuple(Sequence<0, 1, 2, 3, 4>{}),
make_tuple(Sequence<0>{}));
const auto m_thread_data_on_block_idx =
m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
make_multi_index(m_thread_data_on_block));
const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
make_tuple(Sequence<0, 1, 2>{}),
make_tuple(Sequence<0>{}));
const auto n_thread_data_on_block_idx =
n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
make_multi_index(n_thread_data_on_block));
// shuffle: threadwise copy C from VGPR to LDS
auto c_thread_copy_vgpr_to_lds =
ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
CShuffleDataType,
decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
ck::tensor_operation::element_wise::PassThrough,
Sequence<CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
I1,
I1,
M2,
I1,
M4,
I1>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
7,
1,
InMemoryDataOperationEnum::Set,
1,
true>{
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
make_multi_index(0,
0,
m_thread_data_on_block_idx[I1],
n_thread_data_on_block_idx[I1],
m_thread_data_on_block_idx[I2],
m_thread_data_on_block_idx[I3],
m_thread_data_on_block_idx[I4],
n_thread_data_on_block_idx[I2]),
ck::tensor_operation::element_wise::PassThrough{}};
{
// tuple of reference to C/Ds tensor descriptors
const auto c_ds_desc_refs = concat_tuple_of_reference(
tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
generate_tie(
[&](auto i) -> const auto& // return type should be reference
{ return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
Number<NumDTensor>{}));
// tuple of reference to C/Ds tensor descriptors
const auto c_ds_buf_refs = concat_tuple_of_reference(
tie(c_shuffle_block_buf),
generate_tie(
[&](auto i) -> const auto& // return type should be reference
{ return ds_grid_buf[i]; },
Number<NumDTensor>{}));
// tuple of starting index of C/Ds blockwise copy
const auto idx_c_ds_block_begin = container_concat(
make_tuple(make_multi_index(0, 0, 0, 0)),
generate_tuple(
[&](auto) {
return make_multi_index(block_work_idx[I1], 0, block_work_idx[I2], 0);
},
Number<NumDTensor>{}));
// blockwise copy C/D/E between LDS and global
auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7<
ThisThreadBlock,
decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})),
Tuple<EDataType>,
decltype(c_ds_desc_refs),
decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
CDEElementwiseOperation,
Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // FIXME: make
// Sequence support
// arbitray type
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
3, // index_t VectorDim,
CDEShuffleBlockTransferScalarPerVector_NPerBlock,
sequence_merge_t<Sequence<true>,
uniform_sequence_gen_t<
NumDTensor,
false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
Sequence<false>> // ThreadTransferDstResetCoordinateAfterRunFlags
{c_ds_desc_refs,
idx_c_ds_block_begin,
tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
make_tuple(make_multi_index(block_work_idx[I1], 0, block_work_idx[I2], 0)),
cde_element_op};
// space filling curve for threadwise C in VGPR before shuffle
constexpr auto sfc_c_vgpr =
SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
Sequence<CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
1,
1,
M2,
1,
M4,
1>>{};
// space filling curve for shuffled blockwise C/D/E
constexpr auto sfc_cde_block =
SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
Sequence<0, 2, 1, 3>,
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
static_assert(num_access == sfc_cde_block.GetNumOfAccess(), "wrong!");
static_for<0, num_access, 1>{}([&](auto access_id) {
// make sure it's safe to write to LDS
block_sync_lds();
// each thread write its data from VGPR to LDS
c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
c_thread_buf,
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
c_shuffle_block_buf);
// make sure it's safe to read from LDS
block_sync_lds();
// each block copy its data from LDS to global
cde_block_copy_lds_and_global.Run(
c_ds_desc_refs,
c_ds_buf_refs,
tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
tie(e_grid_buf));
if constexpr(access_id < num_access - 1)
{
constexpr auto cde_lds_and_global_step =
sfc_cde_block.GetForwardStep(access_id);
// move on Ds
static_for<0, NumDTensor, 1>{}([&](auto i) {
cde_block_copy_lds_and_global.MoveSrcSliceWindow(
c_ds_desc_refs, i + I1, cde_lds_and_global_step);
});
// move on E
cde_block_copy_lds_and_global.MoveDstSliceWindow(
tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
I0,
cde_lds_and_global_step);
}
});
}
}
}
template <bool HasMainKBlockLoop,
typename AGridDesc_AKB_AK0_M_AK1,
typename BGridDesc_BKB_BK0_N_BK1,
typename Block2ETileMap>
__device__ static void Run1(const ABDataType* __restrict__ p_a_grid,
const ABDataType* __restrict__ p_b_grid,
EDataType* __restrict__ p_e_grid,
void* __restrict__ p_shared,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const AGridDesc_AKB_AK0_M_AK1& a_grid_desc_akb_ak0_m_ak1,
const BGridDesc_BKB_BK0_N_BK1& b_grid_desc_bkb_bk0_n_bk1,
const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&,
const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
e_grid_desc_mblock_mperblock_nblock_nperblock,
const Block2ETileMap& block_2_etile_map)
{
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_grid, a_grid_desc_akb_ak0_m_ak1.GetElementSpaceSize());
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_b_grid, b_grid_desc_bkb_bk0_n_bk1.GetElementSpaceSize());
auto e_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_e_grid, e_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
// divide block work by [M, N]
const auto block_work_idx =
block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
if(!block_2_etile_map.ValidCTileIndex(
make_tuple(block_work_idx[I1], block_work_idx[I2]),
make_tuple(e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
{
return;
}
// HACK: this force m/n_block_data_idx_on_grid into SGPR
const index_t k_batch_id = block_work_idx[I0];
const index_t m_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_work_idx[I1] * MPerBlock);
const index_t n_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_work_idx[I2] * NPerBlock);
// lds max alignment
constexpr auto max_lds_align = math::lcm(AK1, BK1);
// A matrix in LDS memory, dst of blockwise copy
constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
constexpr auto a_block_desc_akb_ak0_m_ak1 =
GetABlockDescriptor_AKB_AK0PerBlock_MPerBlock_AK1();
// B matrix in LDS memory, dst of blockwise copy
constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
constexpr auto b_block_desc_bkb_bk0_n_bk1 =
GetBBlockDescriptor_BKB_BK0PerBlock_NPerBlock_BK1();
// A matrix blockwise copy
auto a_blockwise_copy =
ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
AElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum::Set,
Sequence<I1, AK0PerBlock, MPerBlock, AK1>,
ABlockTransferThreadClusterLengths_AK0_M_AK1,
ABlockTransferThreadClusterArrangeOrder,
ABDataType,
ABDataType,
decltype(a_grid_desc_akb_ak0_m_ak1),
decltype(a_block_desc_akb_ak0_m_ak1),
ABlockTransferSrcAccessOrder,
Sequence<0, 2, 1, 3>,
ABlockTransferSrcVectorDim,
3,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_AK1,
1,
1,
AThreadTransferSrcResetCoordinateAfterRun,
true,
NumGemmKPrefetchStage>(
a_grid_desc_akb_ak0_m_ak1,
make_multi_index(k_batch_id, 0, m_block_data_idx_on_grid, 0),
a_element_op,
a_block_desc_akb_ak0_m_ak1,
make_multi_index(0, 0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
// B matrix blockwise copy
auto b_blockwise_copy =
ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
BElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum::Set,
Sequence<I1, BK0PerBlock, NPerBlock, BK1>,
BBlockTransferThreadClusterLengths_BK0_N_BK1,
BBlockTransferThreadClusterArrangeOrder,
ABDataType,
ABDataType,
decltype(b_grid_desc_bkb_bk0_n_bk1),
decltype(b_block_desc_bkb_bk0_n_bk1),
BBlockTransferSrcAccessOrder,
Sequence<0, 2, 1, 3>,
BBlockTransferSrcVectorDim,
3,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_BK1,
1,
1,
BThreadTransferSrcResetCoordinateAfterRun,
true,
NumGemmKPrefetchStage>(
b_grid_desc_bkb_bk0_n_bk1,
make_multi_index(k_batch_id, 0, n_block_data_idx_on_grid, 0),
b_element_op,
b_block_desc_bkb_bk0_n_bk1,
make_multi_index(0, 0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
// GEMM definition
// c_mtx += transpose(a_mtx) * b_mtx
// a_mtx[K0PerBlock, MPerBlock] is in LDS
// b_mtx[K0PerBlock, NPerBlock] is in LDS
// c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
// register
// sanity check
constexpr index_t KPack =
math::max(math::lcm(AK1, BK1),
MfmaSelector<ABDataType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
BlockSize,
ABDataType,
AccDataType,
decltype(a_block_desc_ak0_m_ak1),
decltype(b_block_desc_bk0_n_bk1),
MPerXdl,
NPerXdl,
MXdlPerWave,
NXdlPerWave,
KPack,
LoopSched>();
auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
// LDS allocation for A and B: be careful of alignment
constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<ABDataType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<ABDataType*>(p_shared) + a_block_space_size_aligned,
b_block_desc_bk0_n_bk1.GetElementSpaceSize());
constexpr auto a_block_slice_copy_step = make_multi_index(0, KPerBlock / AK1, 0, 0);
constexpr auto b_block_slice_copy_step = make_multi_index(0, KPerBlock / BK1, 0, 0);
// gridwise GEMM pipeline
const auto gridwise_gemm_pipeline =
GridwiseGemmPipeline_v1_Selector<NumGemmKPrefetchStage, LoopSched>();
const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
(a_grid_desc_akb_ak0_m_ak1.GetLength(I1) * a_grid_desc_akb_ak0_m_ak1.GetLength(I3)) /
KPerBlock);
gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_akb_ak0_m_ak1,
a_block_desc_akb_ak0_m_ak1,
a_blockwise_copy,
a_grid_buf,
a_block_buf,
a_block_slice_copy_step,
b_grid_desc_bkb_bk0_n_bk1,
b_block_desc_bkb_bk0_n_bk1,
b_blockwise_copy,
b_grid_buf,
b_block_buf,
b_block_slice_copy_step,
blockwise_gemm,
c_thread_buf,
num_k_block_main_loop);
// shuffle C and write out
{
static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
"wrong!");
constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
// TODO: hacky, fix it!
constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
// TODO: hacky, fix it!
// c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<CShuffleDataType*>(p_shared),
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
make_tuple(
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
M1, // M1 = MWave
M2, // M2 * M3 * M4 = MPerXdl
M3,
M4)),
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
N1, // N1 = NWave
N2))), // N2 = NPerXdl
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(
Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
// calculate origin of thread output tensor on global memory
// blockwise GEMM c matrix starting index
const auto c_thread_mtx_on_block =
blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
make_tuple(Sequence<0, 1, 2, 3, 4>{}),
make_tuple(Sequence<0>{}));
const auto m_thread_data_on_block_idx =
m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
make_multi_index(m_thread_data_on_block));
const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
make_tuple(Sequence<0, 1, 2>{}),
make_tuple(Sequence<0>{}));
const auto n_thread_data_on_block_idx =
n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
make_multi_index(n_thread_data_on_block));
// shuffle: threadwise copy C from VGPR to LDS
auto c_thread_copy_vgpr_to_lds =
ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
CShuffleDataType,
decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
ck::tensor_operation::element_wise::PassThrough,
Sequence<CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
I1,
I1,
M2,
I1,
M4,
I1>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
7,
1,
InMemoryDataOperationEnum::Set,
1,
true>{
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
make_multi_index(0,
0,
m_thread_data_on_block_idx[I1],
n_thread_data_on_block_idx[I1],
m_thread_data_on_block_idx[I2],
m_thread_data_on_block_idx[I3],
m_thread_data_on_block_idx[I4],
n_thread_data_on_block_idx[I2]),
ck::tensor_operation::element_wise::PassThrough{}};
{
// shuffle: blockwise copy C from LDS to global
auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
ThisThreadBlock, // ThreadGroup
ck::tensor_operation::element_wise::PassThrough, // ElementwiseOperation,
EGlobalMemoryDataOperation, // DstInMemOp,
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
CShuffleDataType, // typename SrcData,
EDataType, // typename DstData,
decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
decltype(e_grid_desc_mblock_mperblock_nblock_nperblock),
Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
3, // index_t VectorDim,
CDEShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
true, // bool ThreadTransferSrcResetCoordinateAfterRun,
false> // bool ThreadTransferDstResetCoordinateAfterRun>
{c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(0, 0, 0, 0),
e_grid_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(block_work_idx[I1], 0, block_work_idx[I2], 0),
ck::tensor_operation::element_wise::PassThrough{}};
// space filling curve for threadwise C in VGPR
constexpr auto sfc_c_vgpr =
SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
Sequence<CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
1,
1,
M2,
1,
M4,
1>>{};
// space filling curve for shuffled blockwise C in global mem
constexpr auto sfc_c_global =
SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
Sequence<0, 2, 1, 3>,
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
static_assert(num_access == sfc_c_global.GetNumOfAccess(), "wrong!");
static_for<0, num_access, 1>{}([&](auto access_id) {
// make sure it's safe to write to LDS
block_sync_lds();
// each thread write its data from VGPR to LDS
c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
c_thread_buf,
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
c_shuffle_block_buf);
// make sure it's safe to read from LDS
block_sync_lds();
// each block copy its data from LDS to global
c_shuffle_block_copy_lds_to_global.Run(
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
c_shuffle_block_buf,
e_grid_desc_mblock_mperblock_nblock_nperblock,
e_grid_buf);
if constexpr(access_id < num_access - 1)
{
constexpr auto c_global_step = sfc_c_global.GetForwardStep(access_id);
// move on C
c_shuffle_block_copy_lds_to_global.MoveDstSliceWindow(
e_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
}
});
}
}
}
};
} // namespace ck
include/ck/tensor_operation/gpu/grid/gridwise_layernorm_welford_variance.hpp
View file @ 13587ab3
......@@ -57,7 +57,7 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk
make_cluster_descriptor(ThreadClusterLengths_M_K{}, ThreadClusterArrangeOrder{});
using ThreadReduceSrcDesc_M_K = decltype(make_naive_tensor_descriptor_packed(
make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{})));
make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{})));
using ThreadReduceDstDesc_M =
decltype(make_naive_tensor_descriptor_packed(make_tuple(Number<MThreadSliceSize>{})));
......@@ -73,8 +73,14 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
static constexpr index_t K_BlockTileStepSize = KThreadClusterSize * XSrcVectorSize;
static constexpr auto XThreadBufferNumber = Number<KThreadSliceSize / XSrcVectorSize>{};
static constexpr auto GammaThreadBufferNumber = Number<KThreadSliceSize / XSrcVectorSize>{};
static constexpr auto BetaThreadBufferNumber = Number<KThreadSliceSize / XSrcVectorSize>{};
static constexpr auto YThreadBufferNumber = Number<KThreadSliceSize / XSrcVectorSize>{};
__device__ static int GetKPerThread(const GridDesc_M_K& x_grid_desc_m_k,
int thread_k_cluster_id)
......@@ -87,10 +93,13 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk
if(kPerBlockTail > 0)
{
int thread_max_len = (thread_k_cluster_id + 1) * KThreadSliceSize;
int delta = thread_max_len - kPerBlockTail;
delta = math::clamp(thread_max_len - kPerBlockTail, 0, KThreadSliceSize);
kPerThread += KThreadSliceSize - delta;
static_for<0, XThreadBufferNumber, 1>{}([&](auto i) {
int thread_max_len =
(thread_k_cluster_id + 1) * XSrcVectorSize + K_BlockTileStepSize * i;
int delta = thread_max_len - kPerBlockTail;
delta = math::clamp(thread_max_len - kPerBlockTail, 0, XSrcVectorSize);
kPerThread += XSrcVectorSize - delta;
});
}
return kPerThread;
......@@ -116,19 +125,41 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk
auto y_global_val_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_y_global, y_grid_desc_m_k.GetElementSpaceSize());
StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
x_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
gamma_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr,
AccDataType,
MThreadSliceSize * KThreadSliceSize,
true>& beta_thread_buf = gamma_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize * KThreadSliceSize, true>
y_thread_buf;
auto x_thread_buf = generate_tuple(
[&](auto) {
return StaticBuffer<AddressSpaceEnum::Vgpr,
AccDataType,
MThreadSliceSize * XSrcVectorSize,
true>{};
},
Number<XThreadBufferNumber>{});
auto gamma_thread_buf = generate_tuple(
[&](auto) {
return StaticBuffer<AddressSpaceEnum::Vgpr,
AccDataType,
MThreadSliceSize * GammaSrcVectorSize,
true>{};
},
Number<GammaThreadBufferNumber>{});
auto beta_thread_buf = generate_tuple(
[&](auto) {
return StaticBuffer<AddressSpaceEnum::Vgpr,
AccDataType,
MThreadSliceSize * BetaSrcVectorSize,
true>{};
},
Number<BetaThreadBufferNumber>{});
auto y_thread_buf = generate_tuple(
[&](auto) {
return StaticBuffer<AddressSpaceEnum::Vgpr,
AccDataType,
MThreadSliceSize * YDstVectorSize,
true>{};
},
Number<YThreadBufferNumber>{});
StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> mean_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, AccDataType, MThreadSliceSize, true> var_thread_buf;
......@@ -142,9 +173,9 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk
const auto thread_m_cluster_id = thread_cluster_idx[I0];
const auto thread_k_cluster_id = thread_cluster_idx[I1];
using ThreadBufferLengths_M_K = Sequence<MThreadSliceSize, KThreadSliceSize>;
using ThreadBufferLengths_M_K = Sequence<MThreadSliceSize, XSrcVectorSize>;
constexpr auto thread_buffer_desc_m_k = make_naive_tensor_descriptor_packed(
make_tuple(Number<MThreadSliceSize>{}, Number<KThreadSliceSize>{}));
make_tuple(Number<MThreadSliceSize>{}, Number<XSrcVectorSize>{}));
auto threadwise_x_load = ThreadwiseTensorSliceTransfer_v2<XDataType,
AccDataType,
......@@ -159,7 +190,7 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk
x_grid_desc_m_k,
make_multi_index(block_global_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize,
thread_k_cluster_id * KThreadSliceSize));
thread_k_cluster_id * XSrcVectorSize));
auto threadwise_gamma_load =
ThreadwiseTensorSliceTransfer_v2<GammaDataType,
......@@ -175,7 +206,7 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk
gamma_grid_desc_m_k,
make_multi_index(block_global_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize,
thread_k_cluster_id * KThreadSliceSize));
thread_k_cluster_id * GammaSrcVectorSize));
auto threadwise_beta_load =
ThreadwiseTensorSliceTransfer_v2<BetaDataType,
......@@ -191,7 +222,7 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk
beta_grid_desc_m_k,
make_multi_index(block_global_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize,
thread_k_cluster_id * KThreadSliceSize));
thread_k_cluster_id * BetaSrcVectorSize));
auto threadwise_y_store =
ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
......@@ -209,13 +240,10 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk
y_grid_desc_m_k,
make_multi_index(block_global_id * M_BlockTileSize +
thread_m_cluster_id * MThreadSliceSize,
thread_k_cluster_id * KThreadSliceSize),
thread_k_cluster_id * YDstVectorSize),
acc_elementwise_op);
// Copy x from Cache
// one pass: fwd, second pass: bwd
constexpr auto thread_copy_fwd_step_m_k =
make_multi_index(0, SweepOnce ? 0 : K_BlockTileSize);
constexpr auto thread_copy_fwd_step_m_k = make_multi_index(0, K_BlockTileStepSize);
constexpr auto thread_copy_bwd_step_m_k =
make_multi_index(0, SweepOnce ? 0 : -K_BlockTileSize);
......@@ -238,14 +266,15 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk
for(index_t reducedTiles = 0; reducedTiles < num_k_block_tile_iteration; ++reducedTiles)
{
threadwise_x_load.Run(x_grid_desc_m_k,
x_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
x_thread_buf);
threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
threadwise_welford.Run(x_thread_buf, mean_thread_buf, var_thread_buf);
static_for<0, XThreadBufferNumber, 1>{}([&](auto i) {
threadwise_x_load.Run(x_grid_desc_m_k,
x_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
x_thread_buf(i));
threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
threadwise_welford.Run(x_thread_buf[i], mean_thread_buf, var_thread_buf);
});
}
static_for<0, MThreadSliceSize, 1>{}([&](auto I) {
......@@ -256,7 +285,8 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk
BlockwiseWelford::Run(mean_thread_buf(I), var_thread_buf(I), count);
});
auto thread_copy_tail_m_k = (num_k_block_tile_iteration - 1) * thread_copy_fwd_step_m_k;
auto thread_copy_tail_m_k =
(num_k_block_tile_iteration - 1) * XThreadBufferNumber * thread_copy_fwd_step_m_k;
threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k);
threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k, thread_copy_tail_m_k);
......@@ -267,62 +297,86 @@ struct GridwiseLayernormWelfordVariance_mk_to_mk
{
if constexpr(!SweepOnce)
{
threadwise_x_load.Run(x_grid_desc_m_k,
x_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
x_thread_buf);
static_for<0, XThreadBufferNumber, 1>{}([&](auto i) {
threadwise_x_load.Run(x_grid_desc_m_k,
x_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
x_thread_buf(i));
threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_fwd_step_m_k);
});
}
threadwise_gamma_load.Run(gamma_grid_desc_m_k,
gamma_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
gamma_thread_buf);
static_for<0, GammaThreadBufferNumber, 1>{}([&](auto i) {
threadwise_gamma_load.Run(gamma_grid_desc_m_k,
gamma_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
gamma_thread_buf(i));
threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
thread_copy_fwd_step_m_k);
});
static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
constexpr auto offset_m_k =
thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
// normalize
y_thread_buf(Number<offset_m_k>{}) =
(x_thread_buf(Number<offset_m_k>{}) - mean_thread_buf(iM)) /
sqrt(var_thread_buf(iM) + epsilon);
// gamma
y_thread_buf(Number<offset_m_k>{}) =
y_thread_buf(Number<offset_m_k>{}) * gamma_thread_buf(Number<offset_m_k>{});
auto divisor = 1 / __builtin_amdgcn_sqrtf(var_thread_buf(iM) + epsilon);
static_for<0, XThreadBufferNumber, 1>{}([&](auto iK0) {
static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
constexpr auto offset_m_k =
thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
// normalize
y_thread_buf(iK0)(Number<offset_m_k>{}) =
(x_thread_buf(iK0)(Number<offset_m_k>{}) - mean_thread_buf(iM)) *
divisor;
// gamma
y_thread_buf(iK0)(Number<offset_m_k>{}) =
y_thread_buf(iK0)(Number<offset_m_k>{}) *
gamma_thread_buf(iK0)(Number<offset_m_k>{});
});
});
});
threadwise_beta_load.Run(beta_grid_desc_m_k,
beta_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
beta_thread_buf);
static_for<0, BetaThreadBufferNumber, 1>{}([&](auto i) {
threadwise_beta_load.Run(beta_grid_desc_m_k,
beta_global_val_buf,
thread_buffer_desc_m_k,
make_tuple(I0, I0),
beta_thread_buf(i));
threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
thread_copy_fwd_step_m_k);
});
static_for<0, MThreadSliceSize, 1>{}([&](auto iM) {
static_for<0, KThreadSliceSize, 1>{}([&](auto iK) {
constexpr auto offset_m_k =
thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK));
// beta
y_thread_buf(Number<offset_m_k>{}) =
y_thread_buf(Number<offset_m_k>{}) + beta_thread_buf(Number<offset_m_k>{});
static_for<0, XThreadBufferNumber, 1>{}([&](auto iK0) {
static_for<0, XSrcVectorSize, 1>{}([&](auto iK1) {
constexpr auto offset_m_k =
thread_buffer_desc_m_k.CalculateOffset(make_tuple(iM, iK1));
// beta
y_thread_buf(iK0)(Number<offset_m_k>{}) =
y_thread_buf(iK0)(Number<offset_m_k>{}) +
beta_thread_buf(iK0)(Number<offset_m_k>{});
});
});
});
threadwise_y_store.Run(thread_buffer_desc_m_k,
make_tuple(I0, I0),
y_thread_buf,
y_grid_desc_m_k,
y_global_val_buf);
static_for<0, YThreadBufferNumber, 1>{}([&](auto i) {
threadwise_y_store.Run(thread_buffer_desc_m_k,
make_tuple(I0, I0),
y_thread_buf(i),
y_grid_desc_m_k,
y_global_val_buf);
threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_fwd_step_m_k);
});
threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, thread_copy_bwd_step_m_k);
threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k, thread_copy_bwd_step_m_k);
threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k, thread_copy_bwd_step_m_k);
threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, thread_copy_bwd_step_m_k);
threadwise_x_load.MoveSrcSliceWindow(x_grid_desc_m_k, 2 * thread_copy_bwd_step_m_k);
threadwise_gamma_load.MoveSrcSliceWindow(gamma_grid_desc_m_k,
2 * thread_copy_bwd_step_m_k);
threadwise_beta_load.MoveSrcSliceWindow(beta_grid_desc_m_k,
2 * thread_copy_bwd_step_m_k);
threadwise_y_store.MoveDstSliceWindow(y_grid_desc_m_k, 2 * thread_copy_bwd_step_m_k);
}
}
};
......
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