Skip to content

GitLab

Commit a0a469e4 authored by aska-0096's avatar aska-0096
Browse files

save progress

parent 3ddd3578
Hide whitespace changes
Inline Side-by-side
Showing with 501 additions and 632 deletions
include/ck/tensor_operation/gpu/block/blockwise_gemm_wmma.hpp
View file @ a0a469e4
......@@ -129,7 +129,11 @@ struct BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle
return make_tuple(c_thread_m, c_thread_n);
}
__host__ __device__ BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle()
using Tuple5 = decltype(CalculateAThreadOriginDataIndex());
__host__ __device__ BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle(
Tuple4 a_origin = CalculateAThreadOriginDataIndex(),
Tuple4 b_origin = CalculateBThreadOriginDataIndex())
: a_thread_copy_(a_origin), b_thread_copy_(b_origin)
{
static_assert(AK0MK1BlockDesc::IsKnownAtCompileTime() &&
BK0NK1BlockDesc::IsKnownAtCompileTime(),
......@@ -299,8 +303,8 @@ struct BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle
B_K1,
B_K1>;
AThreadCopy a_thread_copy_{CalculateAThreadOriginDataIndex()};
BThreadCopy b_thread_copy_{CalculateBThreadOriginDataIndex()};
AThreadCopy a_thread_copy_;
BThreadCopy b_thread_copy_;
};
// block wise level pipe designed for inline asm
......
include/ck/tensor_operation/gpu/device/impl/device_batched_gemm_softmax_gemm_permute_wmma_cshuffle.hpp
View file @ a0a469e4
......@@ -13,7 +13,7 @@
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/matrix_padder.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_wmma_cshuffle_v1.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_wmma_cshuffle.hpp"
#include "ck/tensor_operation/operator_transform/transform_contraction_to_gemm.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
......
include/ck/tensor_operation/gpu/grid/gridwise_batched_gemm_softmax_gemm_wmma_cshuffle.hpp
View file @ a0a469e4
......@@ -18,72 +18,134 @@
namespace ck {
template <typename FloatAB,
typename FloatGemmAcc,
typename FloatCShuffle,
template <typename GridwiseGemm,
typename FloatA,
typename FloatB,
typename FloatC,
typename AGridDesc_K0_M_K1,
typename BGridDesc_K0_N_K1,
typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
typename Block2CTileMap,
bool HasMainKBlockLoop>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
kernel_gemm_wmma(
const FloatA* __restrict__ p_a_grid,
const FloatB* __restrict__ p_b0_grid,
FloatC* __restrict__ p_c_grid,
const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1,
const BGridDesc_K0_N_K1 b0_grid_desc_k0_l_k1,
const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
c_grid_desc_mblock_mperblock_nblock_nperblock,
// const
// CGridDescriptor_MBlockxRepeat_MWave_MSubGroup_MAccVgprs_NBlockxRepeat_NWave_NThreadPerSubGroup
// c_grid_desc_mblockxrepeat_mwave_msubgroup_maccvgprs_nblockxrepeat_nwave_nthreadpersubgroup,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const CElementwiseOperation c_element_op,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx1100__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
p_b0_grid,
p_c_grid,
p_shared,
a_grid_desc_k0_m_k1,
b0_grid_desc_k0_l_k1,
c_grid_desc_mblock_mperblock_nblock_nperblock,
a_element_op,
b_element_op,
c_element_op,
block_2_ctile_map);
#else
ignore = p_a_grid;
ignore = p_b0_grid;
ignore = p_c_grid;
ignore = a_grid_desc_k0_m_k1;
ignore = b0_grid_desc_k0_l_k1;
ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
ignore = a_element_op;
ignore = b_element_op;
ignore = c_element_op;
ignore = block_2_ctile_map;
#endif // end of if (defined(__gfx1100__))
}
// Gemm0: A [M x K] x B0 [K x L] = Acc [M x L]
// Gemm1: Acc [M x L] x B1 [L x N] = C [M x N]
template <index_t BlockSize,
typename FloatA,
typename FloatB0,
typename FloatB1,
typename FloatAcc,
typename FloatCShuffle,
typename FloatC,
typename AElementwiseOperation,
typename B0ElementwiseOperation,
typename AccElementwiseOperation,
typename B1ElementwiseOperation,
typename CElementwiseOperation,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
typename AGridDesc_AK0_M_AK1,
typename BGridDesc_BK0_N_BK1,
typename B1GridDesc_BK0_N_BK1,
typename AGridDesc_K0_M_K1,
typename B0GridDesc_K0_L_K1,
typename B1GridDesc_L0_N_L1,
typename CGridDesc_M_N,
index_t NumGemmKPrefetchStage,
index_t BlockSize,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t Gemm0MPerBlock,
index_t Gemm0LPerBlock,
index_t Gemm0K0PerBlock,
index_t Gemm0K1Value,
index_t Gemm0MPerWmma,
index_t Gemm0LPerWmma,
index_t Gemm0MRepeat,
index_t Gemm0LRepeat,
index_t Gemm1MPerBlock,
index_t Gemm1NPerBlock,
index_t Gemm1KPerBlock,
index_t AK1Value,
index_t BK1Value,
index_t B1K1Value,
index_t MPerXdl,
index_t NPerXdl,
index_t MXdlPerWave,
index_t NXdlPerWave,
index_t Gemm1NXdlPerWave,
typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
index_t Gemm1L0PerBlock,
index_t Gemm1L1Value,
index_t Gemm1MPerWmma,
index_t Gemm1NPerWmma,
index_t Gemm1MRepeat,
index_t Gemm1NRepeat,
typename ABlockTransferThreadClusterLengths_K0_M_K1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
index_t ABlockTransferSrcVectorDim,
index_t ABlockTransferSrcScalarPerVector,
index_t ABlockTransferDstScalarPerVector_AK1,
bool AThreadTransferSrcResetCoordinateAfterRun, // ignored
index_t ABlockLdsExtraM,
typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
index_t BBlockTransferSrcVectorDim,
index_t BBlockTransferSrcScalarPerVector,
index_t BBlockTransferDstScalarPerVector_BK1,
bool BThreadTransferSrcResetCoordinateAfterRun, // ignored
index_t BBlockLdsExtraN,
typename B1BlockTransferThreadClusterLengths_BK0_N_BK1,
index_t ABlockTransferDstScalarPerVector_K1,
bool AThreadTransferSrcResetCoordinateAfterRun,
bool ABlockLdsExtraM,
typename B0BlockTransferThreadClusterLengths_K0_L_K1,
typename B0BlockTransferThreadClusterArrangeOrder,
typename B0BlockTransferSrcAccessOrder,
index_t B0BlockTransferSrcVectorDim,
index_t B0BlockTransferSrcScalarPerVector,
index_t B0BlockTransferDstScalarPerVector_K1,
bool B0ThreadTransferSrcResetCoordinateAfterRun,
bool B0BlockLdsExtraN,
typename B1BlockTransferThreadClusterLengths_L0_N_L1,
typename B1BlockTransferThreadClusterArrangeOrder,
typename B1BlockTransferSrcAccessOrder,
index_t B1BlockTransferSrcVectorDim,
index_t B1BlockTransferSrcScalarPerVector,
index_t B1BlockTransferDstScalarPerVector_BK1,
index_t B1BlockTransferDstScalarPerVector_L1,
bool B1ThreadTransferSrcResetCoordinateAfterRun,
index_t B1BlockLdsExtraN,
index_t CShuffleMXdlPerWavePerShuffle,
index_t CShuffleNXdlPerWavePerShuffle,
bool B1BlockLdsExtraN,
index_t CShuffleMRepeatPerShuffle,
index_t CShuffleNRepeatPerShuffle,
typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CShuffleBlockTransferScalarPerVector_NPerBlock,
LoopScheduler LoopSched,
bool PadN,
bool MaskOutUpperTriangle,
PipelineVersion PipelineVer = PipelineVersion::v1>
struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
index_t NumGemmKPrefetchStage = 1,
LoopScheduler LoopSched = make_default_loop_scheduler(),
PipelineVersion PipelineVer = PipelineVersion::v1>
struct GridwiseBatchedGemmSoftmaxGemm_Wmma_CShuffle
{
static_assert(LoopSched == LoopScheduler::Default,
"Non-default loop scheduler is currently not supported");
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
......@@ -94,161 +156,127 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
static constexpr auto I7 = Number<7>{};
// K1 should be Number<...>
// Gemm0
static constexpr auto AK0 = Number<KPerBlock / AK1Value>{};
static constexpr auto BK0 = Number<KPerBlock / BK1Value>{};
static constexpr auto AK1 = Number<AK1Value>{};
static constexpr auto BK1 = Number<BK1Value>{};
static constexpr auto Gemm0MWaves = MPerBlock / (MPerXdl * MXdlPerWave);
static constexpr auto Gemm0NWaves = NPerBlock / (NPerXdl * NXdlPerWave);
// Gemm1
static constexpr auto B1K0 = Number<Gemm1KPerBlock / B1K1Value>{};
static constexpr auto B1K1 = Number<B1K1Value>{};
static constexpr auto K1 = Number<Gemm0K1Value>{};
static constexpr auto N1 = Number<Gemm1N1Value>{};
using ThisThreadBlock = ThisThreadBlock<BlockSize>;
using GridwiseGemmPipe = remove_cvref_t<decltype(
GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage>())>;
GridwiseGemmPipeline_Selector<PipelineVer, NumGemmKPrefetchStage, LoopSched>())>;
template <typename ABlockDesc_AK0_M_AK1>
__host__ __device__ static constexpr auto
MakeGemm0AMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
__host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1()
{
constexpr index_t MWaves = MPerBlock / (MXdlPerWave * MPerXdl);
constexpr auto max_lds_align = K1;
return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<MXdlPerWave, MWaves, MPerXdl>(
ABlockDesc_AK0_M_AK1{});
}
template <typename BBlockDesc_BK0_N_BK1>
__host__ __device__ static constexpr auto
MakeGemm0BMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
{
constexpr index_t NWaves = NPerBlock / (NXdlPerWave * NPerXdl);
return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<NXdlPerWave, NWaves, NPerXdl>(
BBlockDesc_BK0_N_BK1{});
}
template <typename ABlockDesc_AK0_M_AK1>
__host__ __device__ static constexpr auto
MakeGemm1AMmaTileDescriptor_M0_M1_M2_K(const ABlockDesc_AK0_M_AK1&)
{
return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<MXdlPerWave, 1, 1>(ABlockDesc_AK0_M_AK1{});
}
// A matrix in LDS memory, dst of blockwise copy
constexpr auto a_block_desc_k0perblock_mperblock_k1 = [&]() {
if constexpr(ABlockLdsExtraM)
{
return make_naive_tensor_descriptor(
make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
}
else
{
return make_naive_tensor_descriptor_aligned(
make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
}
}();
template <typename BBlockDesc_BK0_N_BK1>
__host__ __device__ static constexpr auto
MakeGemm1BMmaTileDescriptor_N0_N1_N2_K(const BBlockDesc_BK0_N_BK1&)
{
constexpr index_t Gemm1NWaves = Gemm1NPerBlock / (Gemm1NXdlPerWave * NPerXdl);
return MakeGemmMmaTileDescriptor_MN0_MN1_MN2_K<Gemm1NXdlPerWave, Gemm1NWaves, NPerXdl>(
BBlockDesc_BK0_N_BK1{});
return a_block_desc_k0perblock_mperblock_k1;
}
__host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
__host__ __device__ static constexpr auto GetBBlockDescriptor_K0PerBlock_NPerBlock_K1()
{
// A matrix in LDS memory, dst of blockwise copy
return make_naive_tensor_descriptor(
make_tuple(AK0, Number<MPerBlock>{}, AK1),
make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
}
constexpr auto max_lds_align = K1;
__host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
{
// B matrix in LDS memory, dst of blockwise copy
return make_naive_tensor_descriptor(
make_tuple(BK0, Number<NPerBlock>{}, BK1),
make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
}
constexpr auto b_block_desc_k0perblock_nperblock_k1 = [&]() {
if constexpr(BBlockLdsExtraN)
{
return make_naive_tensor_descriptor(
make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
}
else
{
return make_naive_tensor_descriptor_aligned(
make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
}
}();
__host__ __device__ static constexpr auto GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1()
{
// B1 matrix in LDS memory, dst of blockwise copy
return make_naive_tensor_descriptor(
make_tuple(B1K0, Number<Gemm1NPerBlock>{}, B1K1),
make_tuple(Number<Gemm1NPerBlock + B1BlockLdsExtraN>{} * B1K1, B1K1, I1));
return b_block_desc_k0perblock_nperblock_k1;
}
__host__ __device__ static constexpr auto
GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
// *Caution Here repeat is shuffle repeat
GetCShuffleBlockDescriptor_MShRepeat_MPerShRepeat_NShRepeat_NPerShRepeat()
{
constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
constexpr index_t NWave = Gemm1NPerBlock / (Gemm1NXdlPerWave * NPerXdl);
constexpr index_t MWave = MPerBlock / (MRepeat * MPerWmma);
constexpr index_t NWave = NPerBlock / (NRepeat * NPerWmma);
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
constexpr auto c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat =
make_naive_tensor_descriptor_packed(
make_tuple(I1,
Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
Number<CShuffleMRepeatPerShuffle * MWave * MPerWmma>{},
I1,
Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
Number<CShuffleNRepeatPerShuffle * NWave * NPerWmma>{}));
return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
return c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat;
}
__host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
{
const index_t gemm0_bytes_end = (SharedMemTrait::a_block_space_size_aligned +
SharedMemTrait::b_block_space_size_aligned) *
sizeof(FloatAB);
const index_t gemm1_bytes_end =
(SharedMemTrait::b1_block_space_offset + SharedMemTrait::b1_block_space_size_aligned) *
sizeof(FloatAB);
const index_t softmax_bytes_end = (SharedMemTrait::reduction_space_offset +
SharedMemTrait::reduction_space_size_aligned) *
sizeof(FloatGemmAcc);
const index_t c_block_bytes_end =
SharedMemTrait::c_block_space_size * sizeof(FloatCShuffle);
return math::max(gemm0_bytes_end, gemm1_bytes_end, softmax_bytes_end, c_block_bytes_end);
// LDS allocation for A and B: be careful of alignment
constexpr auto a_block_desc_k0perblock_mperblock_k1 =
GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
constexpr auto b_block_desc_k0perblock_nperblock_k1 =
GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
constexpr auto max_lds_align = K1;
constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
a_block_desc_k0perblock_mperblock_k1.GetElementSpaceSize(), max_lds_align);
constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
b_block_desc_k0perblock_nperblock_k1.GetElementSpaceSize(), max_lds_align);
return (a_block_space_size_aligned * sizeof(FloatA) +
b_block_space_size_aligned * sizeof(FloatB));
}
// block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
template <typename Block2CTileMap>
__host__ __device__ static constexpr bool
CheckValidity(const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
const B1GridDesc_BK0_N_BK1& b1_grid_desc_bk0_n_bk1,
CheckValidity(const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
const BGridDesc_K0_N_K1& b0_grid_desc_k0_l_k1,
const CGridDesc_M_N& c_grid_desc_m_n,
const Block2CTileMap& block_2_ctile_map)
{
static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
(NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value,
"wrong! K1 need to be known at compile-time");
static_assert((MPerBlock % (MPerWmma * MRepeat) == 0) &&
(NPerBlock % (NRepeat * NPerWmma)) == 0,
"Invalid tuning param!");
const auto M = a_grid_desc_ak0_m_ak1.GetLength(I1);
const auto N = b_grid_desc_bk0_n_bk1.GetLength(I1);
const auto K = a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2);
const auto Gemm1N = b1_grid_desc_bk0_n_bk1.GetLength(I1);
const auto M = a_grid_desc_k0_m_k1.GetLength(I1);
const auto N = b0_grid_desc_k0_l_k1.GetLength(I1);
const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
if(!(M == c_grid_desc_m_n.GetLength(I0) && Gemm1N == c_grid_desc_m_n.GetLength(I1)))
{
if(!(M == c_grid_desc_m_n.GetLength(I0) && N == c_grid_desc_m_n.GetLength(I1) &&
K0 == b0_grid_desc_k0_l_k1.GetLength(I0) && K1 == a_grid_desc_k0_m_k1.GetLength(I2) &&
K1 == b0_grid_desc_k0_l_k1.GetLength(I2)))
return false;
}
if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0 &&
Gemm1N % Gemm1NPerBlock == 0))
{
if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
return false;
}
// check gemm0 gridwise gemm pipeline
const auto num_gemm0_k_loop = K / KPerBlock;
if(!GridwiseGemmPipe::IsSupported(num_gemm0_k_loop))
{
return false;
}
// check gridwise gemm pipeline
const auto num_k_loop = K0 / K0PerBlock;
// check gemm1 gridwise gemm pipeline
if(!(NPerBlock % Gemm1KPerBlock == 0))
{
return false;
}
const auto num_gemm1_k_inner_loop = NPerBlock / Gemm1KPerBlock;
if(!GridwiseGemmPipe::IsSupported(num_gemm1_k_inner_loop))
if(!GridwiseGemmPipe::IsSupported(num_k_loop))
{
return false;
}
......@@ -264,7 +292,7 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
__host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
{
const index_t num_loop = K / KPerBlock;
const index_t num_loop = K / (K0PerBlock * K1);
return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
}
......@@ -276,12 +304,12 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
const auto N = c_grid_desc_m_n.GetLength(I1);
const auto MBlock = M / MPerBlock;
const auto NBlock = N / Gemm1NPerBlock;
const auto NBlock = N / NPerBlock;
const auto c_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
c_grid_desc_m_n,
make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
make_unmerge_transform(make_tuple(NBlock, Number<Gemm1NPerBlock>{}))),
make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
......@@ -289,284 +317,225 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
}
// return block_id to C matrix tile idx (m0, n0) mapping
__host__ __device__ static constexpr auto
MakeDefaultBlock2CTileMap(const CGridDesc_M_N& c_grid_desc_m_n)
__host__ __device__ static constexpr auto MakeDefaultBlock2CTileMap(
const CGridDesc_M_N& c_grid_desc_m_n, index_t /* M01 */, index_t /* N01 */)
{
return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, Gemm1NPerBlock, CGridDesc_M_N>(
return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc_M_N>(
c_grid_desc_m_n);
}
using CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
MakeCGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(CGridDesc_M_N{}))>;
using DefaultBlock2CTileMap =
remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}))>;
remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}, 1, 1))>;
struct SharedMemTrait
{
// LDS allocation for A and B: be careful of alignment
static constexpr auto a_block_desc_ak0_m_ak1 =
GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
static constexpr auto b_block_desc_bk0_n_bk1 =
GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
static constexpr auto b1_block_desc_bk0_n_bk1 =
GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1();
static constexpr auto max_lds_align = math::lcm(math::lcm(AK1, BK1), B1K1);
static constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
static constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
static constexpr auto b1_block_space_size_aligned = math::integer_least_multiple(
b1_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
static constexpr auto a_block_space_offset = 0;
static constexpr auto b_block_space_offset = a_block_space_size_aligned.value;
static constexpr auto b1_block_space_offset = 0;
// LDS allocation for reduction
static constexpr index_t reduction_space_size_aligned =
math::integer_least_multiple(BlockSize, max_lds_align);
static constexpr auto reduction_space_offset = 0;
// LDS allocation for C shuffle in LDS
static constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
static constexpr auto c_block_space_size =
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
};
template <bool HasMainKBlockLoop, typename Block2CTileMap, typename C0MatrixMask>
__device__ static void Run(const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
const FloatAB* __restrict__ p_b1_grid,
template <bool HasMainKBlockLoop, typename C0MatrixMask, typename Block2CTileMap = DefaultBlock2CTileMap>
__device__ static void Run(const FloatA* __restrict__ p_a_grid,
const FloatB0* __restrict__ p_b0_grid,
const FloatB1* __restrict__ p_b1_grid,
FloatC* __restrict__ p_c_grid,
void* __restrict__ p_shared,
const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
const B0GridDesc_K0_L_K1& b0_grid_desc_k0_l_k1,
const B1GridDesc_L0_N_L1& b1_grid_desc_l0_n_l1,
const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
c_grid_desc_mblock_mperblock_nblock_nperblock,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const B0ElementwiseOperation& b0_element_op,
const AccElementwiseOperation& acc_element_op,
const B1ElementwiseOperation& b1_element_op,
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 B1GridDesc_BK0_N_BK1& b1_grid_desc_bk0_n_bk1,
const CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
c_grid_desc_mblock_mperblock_nblock_nperblock,
const Block2CTileMap& block_2_ctile_map,
const C0MatrixMask& c0_matrix_mask)
const C0MatrixMask& c0_matrix_mask,
const Block2CTileMap& block_2_ctile_map)
{
// clang-format off
/*******************************************************************************/
// Memory buffer zone.
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
p_a_grid, a_grid_desc_k0_m_k1.GetElementSpaceSize());
const auto b0_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_b0_grid, b0_grid_desc_k0_l_k1.GetElementSpaceSize());
const auto b1_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_b1_grid, b1_grid_desc_bk0_n_bk1.GetElementSpaceSize());
p_b1_grid, b1_grid_desc_l0_n_l1.GetElementSpaceSize());
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
// divide block work by [M, N]
const auto block_work_idx =
block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
/*******************************************************************************/
// BlockIdx.x -> [BlockId.m, BlockId.n]
const auto block_work_idx = block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
if(!block_2_ctile_map.ValidCTileIndex(
block_work_idx,
make_tuple(c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
c_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
{
return;
}
{ return; }
// HACK: this force m/gemm1_n_block_data_idx_on_grid into SGPR
const index_t m_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
// Store BlockId into SGPR
const index_t m_block_data_idx_on_grid = __builtin_amdgcn_readfirstlane(block_work_idx[I0] * MPerBlock);
const index_t n_block_data_idx_on_grid = __builtin_amdgcn_readfirstlane(block_work_idx[I1] * NPerBlock);
const index_t gemm1_n_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_work_idx[I1] * Gemm1NPerBlock);
/*******************************************************************************/
// set up Gemm0
/*******************************************************************************/
// A matrix in LDS memory, dst of blockwise copy
constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
// B matrix in LDS memory, dst of blockwise copy
constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
//
// set up Gemm0
//
/*******************************************************************************/
// BlockLevel, A/B Matrix ThreadMapping in LDS, As Destinaion of BlockWise_Copy
const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
constexpr auto max_lds_align = K1;
constexpr auto a_block_desc_k0perblock_mperblock_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
constexpr auto b_block_desc_k0perblock_nperblock_k1 = GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
// A matrix blockwise copy
auto a_blockwise_copy =
ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
AElementwiseOperation,
tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum::Set,
Sequence<AK0, MPerBlock, AK1>,
ABlockTransferThreadClusterLengths_AK0_M_AK1,
ABlockTransferThreadClusterArrangeOrder,
FloatAB,
FloatAB,
decltype(a_grid_desc_ak0_m_ak1),
decltype(a_block_desc_ak0_m_ak1),
ABlockTransferSrcAccessOrder,
Sequence<1, 0, 2>,
ABlockTransferSrcVectorDim,
2,
ABlockTransferSrcScalarPerVector,
ABlockTransferDstScalarPerVector_AK1,
1,
1,
true, // SrcResetCoord
true, // DstResetCoord
NumGemmKPrefetchStage>(
a_grid_desc_ak0_m_ak1,
ThreadGroupTensorSliceTransfer_v4r1< ThisThreadBlock,
/* typename SrcElementwiseOperation, */ AElementwiseOperation,
/* typename DstElementwiseOperation, */ ck::tensor_operation::element_wise::PassThrough,
/* InMemoryDataOperationEnum DstInMemOp, */ InMemoryDataOperationEnum::Set,
/* typename BlockSliceLengths, */ Sequence<K0PerBlock, MPerBlock, K1>,
/* typename ThreadClusterLengths, */ ABlockTransferThreadClusterLengths_K0_M_K1,
/* typename ThreadClusterArrangeOrder, */ ABlockTransferThreadClusterArrangeOrder,
/* typename SrcData, */ FloatA,
/* typename DstData, */ FloatA,
/* typename SrcDesc, */ decltype(a_grid_desc_k0_m_k1),
/* typename DstDesc, */ decltype(a_block_desc_k0perblock_mperblock_k1),
/* typename SrcDimAccessOrder, */ ABlockTransferSrcAccessOrder,
/* typename DstDimAccessOrder, */ Sequence<0, 1, 2>,
/* index_t SrcVectorDim, */ ABlockTransferSrcVectorDim,
/* index_t DstVectorDim, */ 2,
/* index_t SrcScalarPerVector, */ ABlockTransferSrcScalarPerVector,
/* index_t DstScalarPerVector, */ ABlockTransferDstScalarPerVector_K1,
/* index_t SrcScalarStrideInVector, */ 1,
/* index_t DstScalarStrideInVector, */ 1,
/* bool ThreadTransferSrcResetCoordinateAfterRun, */ AThreadTransferSrcResetCoordinateAfterRun,
/* bool ThreadTransferDstResetCoordinateAfterRun, */ true>(
a_grid_desc_k0_m_k1,
make_multi_index(0, m_block_data_idx_on_grid, 0),
a_element_op,
a_block_desc_ak0_m_ak1,
a_block_desc_k0perblock_mperblock_k1,
make_multi_index(0, 0, 0),
tensor_operation::element_wise::PassThrough{});
ck::tensor_operation::element_wise::PassThrough{});
// B matrix blockwise copy
auto b_blockwise_copy =
ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
BElementwiseOperation,
tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum::Set,
Sequence<BK0, NPerBlock, BK1>,
BBlockTransferThreadClusterLengths_BK0_N_BK1,
Sequence<K0PerBlock, NPerBlock, K1>,
BBlockTransferThreadClusterLengths_K0_N_K1,
BBlockTransferThreadClusterArrangeOrder,
FloatAB,
FloatAB,
decltype(b_grid_desc_bk0_n_bk1),
decltype(b_block_desc_bk0_n_bk1),
FloatB,
FloatB,
decltype(b0_grid_desc_k0_l_k1),
decltype(b_block_desc_k0perblock_nperblock_k1),
BBlockTransferSrcAccessOrder,
Sequence<1, 0, 2>,
Sequence<0, 1, 2>,
BBlockTransferSrcVectorDim,
2,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_BK1,
BBlockTransferDstScalarPerVector_K1,
1,
1,
true, // SrcResetCoord
true, // DstResetCoord
NumGemmKPrefetchStage>(
b_grid_desc_bk0_n_bk1,
make_multi_index(0, 0, 0), // will loop over GemmN dimension
BThreadTransferSrcResetCoordinateAfterRun,
true>(
b0_grid_desc_k0_l_k1,
make_multi_index(0, n_block_data_idx_on_grid, 0),
b_element_op,
b_block_desc_bk0_n_bk1,
b_block_desc_k0perblock_nperblock_k1,
make_multi_index(0, 0, 0),
tensor_operation::element_wise::PassThrough{});
// Fused Gemm+Gemm pipeline
// for n in N0:
// for k in K0:
// acc[m][n] += A[m][k] * B0[k][n]
// acc1[m][o] += acc[m][n] * B1[n][o]
// sanity check
constexpr index_t KPack = math::max(
math::lcm(AK1, BK1), MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
auto blockwise_gemm = BlockwiseGemmXdlops_v2<
BlockSize,
FloatAB,
FloatGemmAcc,
decltype(a_block_desc_ak0_m_ak1),
decltype(b_block_desc_bk0_n_bk1),
decltype(MakeGemm0AMmaTileDescriptor_M0_M1_M2_K(a_block_desc_ak0_m_ak1)),
decltype(MakeGemm0BMmaTileDescriptor_N0_N1_N2_K(b_block_desc_bk0_n_bk1)),
MPerBlock,
NPerBlock,
KPerBlock,
MPerXdl,
NPerXdl,
MXdlPerWave,
NXdlPerWave,
KPack,
true>{}; // TransposeC
ck::tensor_operation::element_wise::PassThrough{});
/*******************************************************************************/
// Gemm0
constexpr auto WmmaK = 16;
constexpr auto KPack = math::integer_least_multiple(K1, WmmaK);
auto blockwise_gemm0 =
BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle_FIFO<BlockSize,
FloatA,
FloatB,
FloatAcc,
decltype(a_block_desc_k0perblock_mperblock_k1),
decltype(b_block_desc_k0perblock_nperblock_k1),
MPerWmma,
NPerWmma,
MRepeat,
NRepeat,
KPack>{};
// Prepare Register for A*B0 matrix
auto acc_thread_buf = blockwise_gemm.GetCThreadBuffer();
/*******************************************************************************/
constexpr auto a_block_space_size_aligned = math::integer_least_multiple(a_block_desc_k0perblock_mperblock_k1.GetElementSpaceSize(), max_lds_align);
// LDS allocation for A and B: be careful of alignment
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<FloatAB*>(p_shared) + SharedMemTrait::a_block_space_offset,
a_block_desc_ak0_m_ak1.GetElementSpaceSize());
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<FloatAB*>(p_shared) + SharedMemTrait::b_block_space_offset,
b_block_desc_bk0_n_bk1.GetElementSpaceSize());
constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
const auto a_block_reset_copy_step =
make_multi_index(-a_grid_desc_ak0_m_ak1.GetLength(I0), 0, 0);
const auto b_block_reset_copy_step =
make_multi_index(-b_grid_desc_bk0_n_bk1.GetLength(I0), NPerBlock, 0);
// gridwise GEMM pipeline
// Only supports LoopScheduler::Default
const auto gridwise_gemm_pipeline = GridwiseGemmPipeline_Selector<PipelineVer,
NumGemmKPrefetchStage,
LoopScheduler::Default>();
const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
(a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
KPerBlock);
//
// set up Gemm1
//
// Acc matrix threadwise copy: AccVGPR to VGPR and downcast to XDL input data type
constexpr auto acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
constexpr auto m0 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I0);
constexpr auto n0 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I1);
constexpr auto m1 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I2);
constexpr auto n1 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I3);
constexpr auto m2 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I4);
constexpr auto n2 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I5);
constexpr auto n3 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I6);
constexpr auto n4 = acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I7);
constexpr auto b1_block_slice_copy_step = make_multi_index(Gemm1KPerBlock / B1K1, 0, 0);
// acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 to acc_thread_desc_k0_m_k1
// n0_n1_n2_n3 -> k0
// m0_m1_m2 -> m
// n4 -> k1
// NOTE: had to use merge_v3 or will spit out compilation errors
constexpr auto acc_thread_desc_k0_m_k1 = transform_tensor_descriptor(
acc_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4,
make_tuple(make_merge_transform_v3_division_mod(make_tuple(n0, n1, n2, n3)),
make_merge_transform_v3_division_mod(make_tuple(m0, m1, m2)),
make_pass_through_transform(n4)),
make_tuple(Sequence<1, 3, 5, 6>{}, Sequence<0, 2, 4>{}, Sequence<7>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}));
// A1 matrix in AccVGPR
// N2 num_groups_per_blk, N3 num_input_blks, N4 group_size
constexpr auto AccN3 =
blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLength(I6);
constexpr auto A1ThreadSlice_K0_M_K1 =
make_tuple(Number<Gemm1KPerBlock / n4 / AccN3>{}, Number<m0 * m1 * m2>{}, Number<n4>{});
constexpr auto A1ThreadSliceK0 = A1ThreadSlice_K0_M_K1[I0];
constexpr auto A1ThreadSliceM = A1ThreadSlice_K0_M_K1[I1];
constexpr auto A1ThreadSliceK1 = A1ThreadSlice_K0_M_K1[I2];
constexpr auto a1_thread_desc_k0_m_k1 = make_naive_tensor_descriptor(
A1ThreadSlice_K0_M_K1,
make_tuple(A1ThreadSliceM * A1ThreadSliceK1, A1ThreadSliceK1, I1));
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(static_cast<FloatA*>(p_shared), a_block_desc_k0perblock_mperblock_k1.GetElementSpaceSize());
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(static_cast<FloatB*>(p_shared) + a_block_space_size_aligned, b_block_desc_k0perblock_nperblock_k1.GetElementSpaceSize());
// Shift Per SUB_K
constexpr auto a_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
constexpr auto b_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
const auto a_block_reset_copy_step = make_multi_index(-a_grid_desc_k0_m_k1.GetLength(I0), 0, 0);
const auto b_block_reset_copy_step = make_multi_index(-b0_grid_desc_k0_l_k1.GetLength(I0), LPerBlock, 0);
/*******************************************************************************/
// softmax
/*******************************************************************************/
auto workspace_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(static_cast<FloatAcc*>(p_shared), math::integer_least_multiple(BlockSize, max_lds_align));
// get acc0 8D thread cluster
constexpr auto thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4 =
blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths() /
blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
constexpr auto tm0 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I0);
constexpr auto tn0 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I1);
constexpr auto tm1 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I2);
constexpr auto tn1 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I3);
constexpr auto tm2 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I4);
constexpr auto tn2 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I5);
constexpr auto tn3 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I6);
constexpr auto tn4 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I7);
// get acc0 thread map
constexpr auto m0_n_m1_to_m_n_adaptor = make_single_stage_tensor_adaptor(
make_tuple(make_unmerge_transform(make_tuple(tm0 * tm1, tm2)),
make_pass_through_transform(I1)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
constexpr auto threadid_to_m0_n_m1_adaptor = make_single_stage_tensor_adaptor(
make_tuple(
make_merge_transform(make_tuple(tm0 * tm1, tn0 * tn1 * tn2 * tn3 * tn4, tm2))),
make_tuple(Sequence<0, 1, 2>{}),
make_tuple(Sequence<0>{}));
const auto threadid_to_m_n_thread_cluster_adaptor =
chain_tensor_adaptors(m0_n_m1_to_m_n_adaptor, threadid_to_m0_n_m1_adaptor);
// get acc0 2D thread cluster & 2D thread slice
constexpr auto thread_cluster_desc_m_n = make_naive_tensor_descriptor_packed(
make_tuple(tm0 * tm1 * tm2, tn0 * tn1 * tn2 * tn3 * tn4));
constexpr auto thread_slice_desc_m_n =
make_naive_tensor_descriptor_packed(make_tuple(m0 * m1 * m2, n0 * n1 * n2 * n3 * n4));
auto blockwise_softmax = BlockwiseSoftmax<BlockSize,
FloatGemmAcc,
decltype(threadid_to_m_n_thread_cluster_adaptor),
decltype(thread_cluster_desc_m_n),
decltype(thread_slice_desc_m_n)>{};
// Initialize running sum and max of exponentiating row vectors
using SoftmaxBuf = typename decltype(blockwise_softmax)::BufferType;
SoftmaxBuf running_sum, running_sum_new, running_max, running_max_new;
running_sum = 0;
running_sum_new = 0;
running_max = NumericLimits<FloatGemmAcc>::Lowest();
running_max_new = NumericLimits<FloatGemmAcc>::Lowest();
/*******************************************************************************/
// set up Gemm1
/*******************************************************************************/
// B1 matrix in LDS memory, dst of blockwise copy
constexpr auto b1_block_desc_bk0_n_bk1 = GetB1BlockDescriptor_BK0PerBlock_NPerBlock_BK1();
constexpr auto b1_block_desc_l0perblock_nperblock_l1 = GetB1BlockDescriptor_L0PerBlock_NPerBlock_L1();
// A1 matrix blockwise copy
auto a1_blockwise_copy = ThreadwiseTensorSliceTransfer_StaticToStatic<
FloatGemmAcc,
FloatAB,
FloatAcc,
FloatA,
decltype(acc_thread_desc_k0_m_k1),
decltype(a1_thread_desc_k0_m_k1),
tensor_operation::element_wise::PassThrough,
......@@ -574,7 +543,7 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
Sequence<1, 0, 2>,
2,
n4>{tensor_operation::element_wise::PassThrough{}};
// B1 matrix blockwise copy
auto b1_blockwise_copy =
ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
......@@ -605,142 +574,56 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
b1_block_desc_bk0_n_bk1,
make_multi_index(0, 0, 0),
tensor_operation::element_wise::PassThrough{});
auto a1_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatAB>(
a1_thread_desc_k0_m_k1.GetElementSpaceSize());
// reuse LDS space for gemm0's b_block_buf
auto b1_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<FloatAB*>(p_shared) + SharedMemTrait::b1_block_space_offset,
b1_block_desc_bk0_n_bk1.GetElementSpaceSize());
// selected_mfma.group_size or B1K1 <= Gemm1KPack <= selected_mfma.group_size
// selected_mfma.k_per_blk <= Gemm1KPack
//
// Following similar rationale behind Gemm0KPack, let Gemm1KPack be the lowest common
// multiples of A1K1 (predetermined by selected_mfma.group_size) and B1K1. But in this case
// Gemm1KPack can't be higher than A1K1 itself because A1 matrix is distributed in VGPRs
// with 'group_size' amount of contiguous elements. Having Gemm1KPack greater than A1K1 will
// cause mismatch in summation index for example c[0:7] = a1[[0:3, 8:11]] * b1[0:7].
// therefore we may just as well assign Gemm1KPack = group_size
constexpr index_t Gemm1KPack =
MfmaSelector<FloatAB, MPerXdl, NPerXdl>::selected_mfma.group_size;
auto gemm1_blockwise_gemm = BlockwiseGemmXdlops_v2<
BlockSize,
FloatAB,
FloatGemmAcc,
decltype(a1_thread_desc_k0_m_k1),
decltype(b1_block_desc_bk0_n_bk1),
decltype(MakeGemm1AMmaTileDescriptor_M0_M1_M2_K(a1_thread_desc_k0_m_k1)),
decltype(MakeGemm1BMmaTileDescriptor_N0_N1_N2_K(b1_block_desc_bk0_n_bk1)),
MPerBlock,
Gemm1NPerBlock,
Gemm1KPerBlock,
MPerXdl,
NPerXdl,
MXdlPerWave,
Gemm1NXdlPerWave,
Gemm1KPack,
true, // TransposeC
Gemm1KPack, // AMmaKStride
Gemm1KPack * XdlopsGemm<FloatAB, MPerXdl, NPerXdl, Gemm1KPack, false>{}.K0PerXdlops>{
// BMmaKStride
make_tuple(0, 0, 0, 0)}; // A_origin
auto acc1_thread_buf = gemm1_blockwise_gemm.GetCThreadBuffer();
//
// Blockwise softmax
//
auto workspace_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<FloatGemmAcc*>(p_shared) + SharedMemTrait::reduction_space_offset,
SharedMemTrait::reduction_space_size_aligned);
// get acc0 8D thread cluster
constexpr auto thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4 =
blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths() /
blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4().GetLengths();
constexpr auto tm0 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I0);
constexpr auto tn0 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I1);
constexpr auto tm1 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I2);
constexpr auto tn1 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I3);
constexpr auto tm2 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I4);
constexpr auto tn2 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I5);
constexpr auto tn3 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I6);
constexpr auto tn4 = thread_cluster_m0_n0_m1_n1_m2_n2_n3_n4.At(I7);
// get acc0 thread map
constexpr auto m0_n_m1_to_m_n_adaptor = make_single_stage_tensor_adaptor(
make_tuple(make_unmerge_transform(make_tuple(tm0 * tm1, tm2)),
make_pass_through_transform(I1)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
constexpr auto threadid_to_m0_n_m1_adaptor = make_single_stage_tensor_adaptor(
make_tuple(
make_merge_transform(make_tuple(tm0 * tm1, tn0 * tn1 * tn2 * tn3 * tn4, tm2))),
make_tuple(Sequence<0, 1, 2>{}),
make_tuple(Sequence<0>{}));
const auto threadid_to_m_n_thread_cluster_adaptor =
chain_tensor_adaptors(m0_n_m1_to_m_n_adaptor, threadid_to_m0_n_m1_adaptor);
// get acc0 2D thread cluster & 2D thread slice
constexpr auto thread_cluster_desc_m_n = make_naive_tensor_descriptor_packed(
make_tuple(tm0 * tm1 * tm2, tn0 * tn1 * tn2 * tn3 * tn4));
constexpr auto thread_slice_desc_m_n =
make_naive_tensor_descriptor_packed(make_tuple(m0 * m1 * m2, n0 * n1 * n2 * n3 * n4));
auto blockwise_softmax = BlockwiseSoftmax<BlockSize,
FloatGemmAcc,
decltype(threadid_to_m_n_thread_cluster_adaptor),
decltype(thread_cluster_desc_m_n),
decltype(thread_slice_desc_m_n)>{};
const index_t num_gemm1_k_block_outer_loop =
b_grid_desc_bk0_n_bk1.GetLength(I1) / NPerBlock;
auto a1_thread_buf = make_static_buffer<AddressSpaceEnum::Vgpr, FloatA>(a1_thread_desc_k0_m_k1.GetElementSpaceSize());
auto b1_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(static_cast<FloatB*>(p_shared), b_block_desc_k0perblock_nperblock_k1.GetElementSpaceSize());
auto blockwise_gemm1 =
BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle<BlockSize,
FloatA,
FloatB,
FloatAcc,
decltype(a1_thread_desc_k0perblock_mperblock_k1),
decltype(b1_block_desc_k0perblock_nperblock_k1),
MPerWmma,
NPerWmma,
MRepeat,
NRepeat,
KPack>{make_tuple(0, 0, 0, 0, 0)};
auto acc1_thread_buf = blockwise_gemm1.GetCThreadBuffer();
const index_t num_gemm1_k_block_outer_loop = b_grid_desc_bk0_n_bk1.GetLength(I1) / NPerBlock;
constexpr index_t num_gemm1_k_block_inner_loop = NPerBlock / Gemm1KPerBlock;
// Initialize C
StaticBuffer<AddressSpaceEnum::Vgpr, FloatGemmAcc, acc1_thread_buf.Size(), true>
c_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, FloatAcc, acc1_thread_buf.Size(), true> c_thread_buf;
c_thread_buf.Clear();
// Initialize running sum and max of exponentiating row vectors
using SoftmaxBuf = typename decltype(blockwise_softmax)::BufferType;
SoftmaxBuf running_sum, running_sum_new, running_max, running_max_new;
running_sum = 0;
running_sum_new = 0;
running_max = NumericLimits<FloatGemmAcc>::Lowest();
running_max_new = NumericLimits<FloatGemmAcc>::Lowest();
// gemm1 K loop
/*******************************************************************************/
// Flash Attention
// Dao, Tri, et al. "Flashattention: Fast and memory-efficient exact attention with io-awareness." arXiv preprint arXiv:2205.14135 (2022).
index_t gemm1_k_block_outer_index = 0;
do
{
auto n_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(gemm1_k_block_outer_index * NPerBlock);
if(c0_matrix_mask.IsTileSkippable(
m_block_data_idx_on_grid, n_block_data_idx_on_grid, MPerBlock, NPerBlock))
{
continue;
}
// gemm0
gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_grid_desc_ak0_m_ak1,
a_block_desc_ak0_m_ak1,
a_blockwise_copy,
a_grid_buf,
a_block_buf,
a_block_slice_copy_step,
b_grid_desc_bk0_n_bk1,
b_block_desc_bk0_n_bk1,
b_blockwise_copy,
b_grid_buf,
b_block_buf,
b_block_slice_copy_step,
blockwise_gemm,
acc_thread_buf,
num_k_block_main_loop);
// Outer loop, along GEMM_L
// Inner loop, along GEMM_K
do{
// gemm0 start, A-B swaped
const index_t K0BlockMainLoop = __builtin_amdgcn_readfirstlane(K0 / K0PerBlock);
GridwiseGemmPipe::template Run<HasMainKBlockLoop>(a_grid_desc_k0_m_k1,
a_block_desc_k0perblock_mperblock_k1,
a_blockwise_copy,
a_grid_buf,
a_block_buf,
a_block_slice_copy_step,
b0_grid_desc_k0_l_k1,
b_block_desc_k0perblock_nperblock_k1,
b_blockwise_copy,
b0_grid_buf,
b_block_buf,
b_block_slice_copy_step,
blockwise_gemm,
acc_thread_buf,
K0BlockMainLoop);
// do MNK padding or upper triangular masking
if constexpr(MaskOutUpperTriangle || PadN)
{
......@@ -797,13 +680,15 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
});
}
else
{
static_for<0, acc_thread_buf.Size(), 1>{}(
{ static_for<0, acc_thread_buf.Size(), 1>{}(
[&](auto i) { acc_element_op(acc_thread_buf(i), acc_thread_buf[i]); });
}
block_sync_lds(); // wait for lds read in gemm0 blockwise gemm
block_sync_lds();
// gemm0 end
// Tiled softmax start
// softmax
SoftmaxBuf& max = blockwise_softmax.max_value_buf;
SoftmaxBuf& sum = blockwise_softmax.sum_value_buf;
......@@ -814,7 +699,13 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
running_max_new = mathext::max(max, running_max);
running_sum_new = mathext::exp(running_max - running_max_new) * running_sum +
mathext::exp(max - running_max_new) * sum;
// Intra-Row data permutation, make swizzled A input for WMMA
__builtin_amdgcn_permlane16(0xeca86420, 0xfdb97531);
// Low/high row move data to low/high half of thread buffer
/* thread copy*/
// Inter-Row data permutation, fullfill data duplication requirement
__builtin_amdgcn_permlanex16(0x76543210, 0xfedcba98);
// gemm1
{
// TODO: explore using dynamic buffer for a1 thread buffer
......@@ -841,6 +732,7 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
if constexpr(num_gemm1_k_block_inner_loop > 1)
{
static_for<0, num_gemm1_k_block_inner_loop - 1, 1>{}([&](auto i) {
// Data cast from FloatAcc to FloatA happen here
a1_blockwise_copy.Run(acc_thread_desc_k0_m_k1,
make_tuple(Number<i * A1ThreadSliceK0>{}, I0, I0),
acc_thread_buf,
......@@ -879,14 +771,6 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
}
} // end gemm1
// workaround compiler issue; see ck/ck.hpp
if constexpr(CK_WORKAROUND_SWDEV_XXXXXX_BF16_ATTEN_FWD_GFX908_ISSUE == 1 &&
is_same_v<FloatAB, bhalf_t> && MPerBlock == 256 && NPerBlock == 128 &&
Gemm1NPerBlock == 128)
{
__builtin_amdgcn_sched_barrier(0);
}
constexpr auto c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
gemm1_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
constexpr auto cm0 = c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4.GetLength(I0);
......@@ -910,8 +794,7 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
FloatGemmAcc c_new =
(running_sum[iM] * math::exp(running_max[iM] - running_max_new[iM]) * c +
math::exp(max[iM] - running_max_new[iM]) * acc1) /
running_sum_new[iM]; // Formula by Dao et al.,
// https://arxiv.org/pdf/2205.14135v2.pdf section 3.1
running_sum_new[iM];
c_thread_buf(I) = c_new; // O_new
});
......@@ -927,120 +810,102 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
running_sum = running_sum_new;
block_sync_lds(); // wait for gemm1 LDS read
} while(++gemm1_k_block_outer_index < num_gemm1_k_block_outer_loop); // end j loop
// shuffle C and write out
}while(++gemm1_k_block_outer_index < num_gemm1_k_block_outer_loop);
/*******************************************************************************/
// write out to C, implement shuffle
{
static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
Gemm1NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
"wrong!");
constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
constexpr index_t NWave = Gemm1NPerBlock / (Gemm1NXdlPerWave * NPerXdl);
constexpr auto c_thread_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs =
blockwise_gemm.GetCThreadDescriptor_MRepeat_MWave_MSubGroup_NRepeat_NWave_NThreadPerSubGroup_MAccVgprs();
// TODO: hacky, fix it!
constexpr auto c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4 =
gemm1_blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
// This API Provide All dimension (size) you need
constexpr auto c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp =
blockwise_gemm.GetCBlockDescriptor_MRepeat_MWave_MSubGroup_NRepeat_NWave_NThreadPerSubGroup_MAccVgprs();
// TODO: hacky, fix it!
// c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp is only used to get lengths
constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp =
gemm1_blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_N2_N3_N4();
constexpr auto MWave = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I1);
constexpr auto MSubGroup = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I2);
constexpr auto NWave = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I4);
constexpr auto NThreadPerSubGroup = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I5);
constexpr auto MAccVgprs = c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs_tmp.GetLength(I6);
constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I0);
constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I1);
constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I2);
constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I3);
constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I4);
constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I5);
constexpr auto N3 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I6);
constexpr auto N4 = c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4_tmp.GetLength(I7);
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
// LDS descriptor, shuffle and write out in MRepeat x NRepeat times
constexpr auto c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat =
GetCShuffleBlockDescriptor_MShRepeat_MPerShRepeat_NShRepeat_NPerShRepeat();
auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<FloatCShuffle*>(p_shared),
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat.GetElementSpaceSize());
constexpr auto c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4 = transform_tensor_descriptor(
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
constexpr auto c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs = transform_tensor_descriptor(
c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat,
make_tuple(
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
M1, // M1 = MWave
M2)), // M2 = MPerXdl
Number<CShuffleMRepeatPerShuffle>{}, // MRepeat per shuffle repeat
MWave, // MWave
MSubGroup, // MSubGroup * MAccVgprs = MPerWmma
MAccVgprs)),
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
N1, // N1 = NWave
N2, // N2 * N3 * N4 = NPerXdl
N3,
N4))),
Number<CShuffleNRepeatPerShuffle>{}, // NRepeat per shuffle repeat
NWave, // NWave
NThreadPerSubGroup))), // NThreadPerSubGroup = NPerWmma
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(
Sequence<>{}, Sequence<0, 2, 4>{}, Sequence<>{}, Sequence<1, 3, 5, 6, 7>{}));
make_tuple(Sequence<>{}, Sequence<0, 1, 2, 6>{}, Sequence<>{}, Sequence<3, 4, 5>{}));
// calculate origin of thread output tensor on global memory
// blockwise GEMM c matrix starting index
const auto c_thread_mtx_on_block =
gemm1_blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
const auto c_thread_mtx_on_block = blockwise_gemm.CalculateCThreadOriginDataIndex(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_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(M0, M1, M2))),
make_tuple(Sequence<0, 1, 2>{}),
make_tuple(Sequence<0>{}));
const auto m_thread_data_on_block_idx =
m_thread_data_on_block_to_m0_m1_m2_adaptor.CalculateBottomIndex(
make_multi_index(m_thread_data_on_block));
const auto n_thread_data_on_block_to_n0_n1_n2_n3_n4_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(N0, N1, N2, N3, N4))),
make_tuple(Sequence<0, 1, 2, 3, 4>{}),
make_tuple(Sequence<0>{}));
const auto n_thread_data_on_block_idx =
n_thread_data_on_block_to_n0_n1_n2_n3_n4_adaptor.CalculateBottomIndex(
make_multi_index(n_thread_data_on_block));
const auto m_thread_data_on_block_to_mrepeat_mwave_msubgroup_maccvgprs_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(MRepeat, MWave, MSubGroup, MAccVgprs))),
make_tuple(Sequence<0, 1, 2, 3>{}),
make_tuple(Sequence<0>{}));
const auto n_thread_data_on_block_to_nrepeat_nwave_nthreadpersubgroup_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(NRepeat, NWave, NThreadPerSubGroup))),
make_tuple(Sequence<0, 1, 2>{}),
make_tuple(Sequence<0>{}));
const auto m_thread_data_on_block_idx = m_thread_data_on_block_to_mrepeat_mwave_msubgroup_maccvgprs_adaptor.CalculateBottomIndex(
make_multi_index(m_thread_data_on_block));
const auto n_thread_data_on_block_idx = n_thread_data_on_block_to_nrepeat_nwave_nthreadpersubgroup_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<FloatGemmAcc,
ThreadwiseTensorSliceTransfer_v1r3<FloatAcc,
FloatCShuffle,
decltype(c_thread_desc_m0_n0_m1_n1_m2_n2_n3_n4),
decltype(c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4),
tensor_operation::element_wise::PassThrough,
Sequence<CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
decltype(c_thread_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs),
decltype(c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs),
ck::tensor_operation::element_wise::PassThrough,
Sequence<CShuffleMRepeatPerShuffle,
I1,
I1,
CShuffleNRepeatPerShuffle,
I1,
N2,
I1,
N4>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
7,
1,
MAccVgprs>,
Sequence<0, 1, 2, 3, 4, 5, 6>,
6,
1, // vector write pixel
InMemoryDataOperationEnum::Set,
1,
true>{
c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4,
c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs,
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],
0,
n_thread_data_on_block_idx[I1],
n_thread_data_on_block_idx[I2],
n_thread_data_on_block_idx[I3],
n_thread_data_on_block_idx[I4]),
tensor_operation::element_wise::PassThrough{}};
m_thread_data_on_block_idx[I3]),
ck::tensor_operation::element_wise::PassThrough{}};
// shuffle: blockwise copy C from LDS to global
auto c_shuffle_block_copy_lds_to_global = ThreadGroupTensorSliceTransfer_v6r1<
......@@ -1048,47 +913,47 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
CElementwiseOperation, // ElementwiseOperation,
CGlobalMemoryDataOperation, // DstInMemOp,
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
CShuffleMRepeatPerShuffle * MWave * MPerWmma,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
CShuffleNRepeatPerShuffle * NWave * NPerWmma>, // BlockSliceLengths,
CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
FloatCShuffle, // typename SrcData,
FloatC, // typename DstData,
decltype(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
decltype(c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat),
decltype(c_grid_desc_mblock_mperblock_nblock_nperblock),
Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
3, // index_t VectorDim,
CShuffleBlockTransferScalarPerVector_NPerBlock, // index_t ScalarPerVector,
true, // bool ThreadTransferSrcResetCoordinateAfterRun,
false> // bool ThreadTransferDstResetCoordinateAfterRun>
{c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
{c_shuffle_block_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat,
make_multi_index(0, 0, 0, 0),
c_grid_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0),
c_element_op};
// space filling curve for local reg & global memory
// space filling curve for threadwise C in VGPR
constexpr auto sfc_c_vgpr =
SpaceFillingCurve<Sequence<MXdlPerWave, Gemm1NXdlPerWave, 1, 1, 1, N2, 1, N4>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
Sequence<CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
SpaceFillingCurve<Sequence<MRepeat, 1, 1, NRepeat, 1, 1, MAccVgprs>,
Sequence<0, 1, 2, 3, 4, 5, 6>,
Sequence<CShuffleMRepeatPerShuffle,
1,
1,
CShuffleNRepeatPerShuffle,
1,
N2,
1,
N4>>{};
MAccVgprs>>{};
// space filling curve for shuffled blockwise C in global mem
constexpr auto sfc_c_global =
SpaceFillingCurve<Sequence<1, MPerBlock, 1, Gemm1NPerBlock>,
SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
Sequence<0, 2, 1, 3>,
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
CShuffleMRepeatPerShuffle * MWave * MPerWmma,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
CShuffleNRepeatPerShuffle * NWave * NPerWmma>>{};
constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
......@@ -1099,10 +964,10 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
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_n2_n3_n4,
c_thread_copy_vgpr_to_lds.Run(c_thread_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs,
sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
c_thread_buf,
c_block_desc_m0_n0_m1_n1_m2_n2_n3_n4,
c_block_desc_mrepeat_mwave_msubgroup_nrepeat_nwave_nthreadpersubgroup_maccvgprs,
c_shuffle_block_buf);
// make sure it's safe to read from LDS
......@@ -1110,7 +975,7 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
// 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_desc_mshrepeat_mpershrepeat_nshrepeat_npershrepeat,
c_shuffle_block_buf,
c_grid_desc_mblock_mperblock_nblock_nperblock,
c_grid_buf);
......@@ -1118,13 +983,13 @@ struct GridwiseBatchedGemmSoftmaxGemm_Xdl_CShuffle
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(
c_grid_desc_mblock_mperblock_nblock_nperblock, c_global_step);
}
});
}
// clang-format on
}
};
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_wmma.hpp
View file @ a0a469e4
......@@ -414,7 +414,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_wmma
constexpr auto KPack = math::integer_least_multiple(K1, WmmaK);
auto blockwise_gemm =
BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle_FIFO<BlockSize,
BlockwiseGemmWMMA_k0mk1_k0nk1_m0m1m2n0n1n2m3_CShuffle<BlockSize,
FloatA,
FloatB,
FloatAcc,
......
include/ck/utility/amd_wmma.hpp
View file @ a0a469e4
......@@ -23,11 +23,11 @@ struct intrin_wmma_f32_16x16x16_f16_w32<16, 16>
{
// * Inline assembly need to elimate the duplicated data load, compiler won't help you
// delete them.
amd_assembly_wmma_f32_16x16x16_f16_w32(
reg_a, reg_b, reg_c.template AsType<float8_t>()(Number<0>{}));
// reg_c.template AsType<float8_t>()(Number<0>{}) =
// __builtin_amdgcn_wmma_f32_16x16x16_f16_w32( reg_a, reg_b, reg_c.template
// AsType<float8_t>()[Number<0>{}]);
// amd_assembly_wmma_f32_16x16x16_f16_w32(
// reg_a, reg_b, reg_c.template AsType<float8_t>()(Number<0>{}));
reg_c.template AsType<float8_t>()(Number<0>{}) =
__builtin_amdgcn_wmma_f32_16x16x16_f16_w32( reg_a, reg_b, reg_c.template
AsType<float8_t>()[Number<0>{}]);
}
};
......
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