Skip to content

GitLab

Unverified Commit 0d6da13c authored by rocking's avatar rocking Committed by GitHub
Browse files

Merge branch 'develop' into normalization/splitK

parents 27f8c64b b076a02a
Hide whitespace changes
Inline Side-by-side
Showing with 753 additions and 245 deletions
include/ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp
View file @ 0d6da13c
......@@ -54,7 +54,8 @@ __global__ void
const ReduceGridDescriptor_MBlock_MPerBlock reduce_grid_desc_mblock_mperblock,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_cshuffle_v1.hpp
View file @ 0d6da13c
......@@ -44,7 +44,8 @@ __global__ void
c_grid_desc_mblock_mperblock_nblock_nperblock,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdl_layernorm_cshuffle_v1.hpp
View file @ 0d6da13c
......@@ -57,7 +57,8 @@ __global__ void
const C0GridDescriptor_NBlock_NPerBlock c0_grid_desc_nblock_nperblock,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
// TODO ANT: separate into MMA + Epilogue
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_bwd_weight.hpp
View file @ 0d6da13c
......@@ -165,7 +165,8 @@ __global__ void
const CElementwiseOperation c_element_op,
const CBlockClusterAdaptor c_block_cluster_adaptor)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
......@@ -265,7 +266,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_bwd_weight
// we convert fp16->fp32->bf16 and execute bf16 mfma instruction
// when mfma if fixed, remove this section and update
// FloatABAdjusted -> FloatAB throughout this file
#if CK_WORKAROUND_DENORM_FIX && defined(__gfx90a__)
#if CK_WORKAROUND_DENORM_FIX
using FloatABAdjusted = conditional_t<is_same_v<FloatAB, ck::half_t>, ck::bhalf_t, FloatAB>;
#else
using FloatABAdjusted = FloatAB;
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_skip_b_lds_v1.hpp
View file @ 0d6da13c
......@@ -44,7 +44,8 @@ __global__ void
const CElementwiseOperation c_element_op,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainK0BlockLoop>(p_a_grid,
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp
View file @ 0d6da13c
......@@ -43,7 +43,8 @@ __global__ void
const CElementwiseOperation c_element_op,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
......@@ -135,7 +136,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
// we convert fp16->fp32->bf16 and execute bf16 mfma instruction
// when mfma if fixed, remove this section and update
// FloatABAdjusted -> FloatAB throughout this file
#if CK_WORKAROUND_DENORM_FIX && defined(__gfx90a__)
#if CK_WORKAROUND_DENORM_FIX
using FloatABAdjusted = conditional_t<is_same_v<FloatAB, ck::half_t>, ck::bhalf_t, FloatAB>;
#else
using FloatABAdjusted = FloatAB;
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4.hpp
View file @ 0d6da13c
......@@ -42,7 +42,8 @@ __global__ void
const CElementwiseOperation c_element_op,
const CBlockClusterAdaptor c_block_cluster_adaptor)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
constexpr index_t shared_block_size =
GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp
View file @ 0d6da13c
......@@ -15,63 +15,32 @@
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_selector.hpp"
namespace ck {
template <typename GridwiseGemm,
typename FloatAB,
typename FloatC,
typename AGridDesc_B_K0_M_K1,
typename BGridDesc_B_K0_N_K1,
typename CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
typename CBlockClusterAdaptor,
bool HasMainKBlockLoop>
bool HasMainKBlockLoop,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
typename Block2CTileMap>
__global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
kernel_gemm_xdlops_v2r4r2(const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid,
const AGridDesc_B_K0_M_K1 a_b_k0_m_k1_grid_desc,
const BGridDesc_B_K0_N_K1 b_b_k0_n_k1_grid_desc,
const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock
c_grid_desc_mblock_mperblock_nblock_nperblock,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const CElementwiseOperation c_element_op,
const CBlockClusterAdaptor c_block_cluster_adaptor)
kernel_gemm_xdlops_v2r4r2_simplified(typename GridwiseGemm::Argument karg,
const Block2CTileMap& b2c_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
constexpr index_t shared_block_size =
GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
__shared__ FloatAB p_shared_block[shared_block_size];
GridwiseGemm::template Run<HasMainKBlockLoop>(p_a_grid,
p_b_grid,
p_c_grid,
static_cast<void*>(p_shared_block),
a_b_k0_m_k1_grid_desc,
b_b_k0_n_k1_grid_desc,
c_grid_desc_mblock_mperblock_nblock_nperblock,
a_element_op,
b_element_op,
c_element_op,
c_block_cluster_adaptor);
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
constexpr index_t shared_size = GridwiseGemm::GetSharedMemoryNumberOfByte();
__shared__ uint8_t p_shared[shared_size];
GridwiseGemm::template Run<HasMainKBlockLoop, CGlobalMemoryDataOperation>(
karg, static_cast<void*>(p_shared), b2c_map);
#else
ignore = p_a_grid;
ignore = p_b_grid;
ignore = p_c_grid;
ignore = a_b_k0_m_k1_grid_desc;
ignore = b_b_k0_n_k1_grid_desc;
ignore = c_grid_desc_mblock_mperblock_nblock_nperblock;
ignore = a_element_op;
ignore = b_element_op;
ignore = c_element_op;
ignore = c_block_cluster_adaptor;
ignore = karg;
ignore = b2c_map;
#endif // end of if (defined(__gfx908__) || defined(__gfx90a__))
}
......@@ -79,13 +48,13 @@ template <index_t BlockSize,
typename FloatAB,
typename FloatAcc,
typename FloatC,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
typename AGridDesc_B_K0_M_K1,
typename BGridDesc_B_K0_N_K1,
typename CMNGridDesc,
typename ALayout,
typename BLayout,
typename CLayout,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
tensor_operation::device::GemmSpecialization GemmSpec,
index_t MPerBlock,
index_t NPerBlock,
index_t K0PerBlock,
......@@ -126,10 +95,238 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
static constexpr auto I7 = Number<7>{};
// K1 should be Number<...>
static constexpr auto K1 = Number<K1Value>{};
static constexpr auto K1 = Number<K1Value>{};
static constexpr auto M01 = 1;
static constexpr auto N01 = 1;
using ThisThreadBlock = ThisThreadBlock<BlockSize>;
struct Argument : public ck::tensor_operation::device::BaseArgument
{
const FloatAB* p_a_grid;
const FloatAB* p_b_grid;
FloatC* p_c_grid;
index_t M;
index_t N;
index_t K;
index_t StrideA;
index_t StrideB;
index_t StrideC;
index_t MPadded;
index_t NPadded;
index_t KPadded;
index_t K0;
index_t k_batch;
Argument(const FloatAB* p_a_grid_,
const FloatAB* p_b_grid_,
FloatC* p_c_grid_,
index_t M_,
index_t N_,
index_t K_,
index_t StrideA_,
index_t StrideB_,
index_t StrideC_,
index_t MPadded_,
index_t NPadded_,
index_t KPadded_,
index_t K0_,
index_t k_batch_)
: p_a_grid(p_a_grid_),
p_b_grid(p_b_grid_),
p_c_grid(p_c_grid_),
M(M_),
N(N_),
K(K_),
StrideA(StrideA_),
StrideB(StrideB_),
StrideC(StrideC_),
MPadded(MPadded_),
NPadded(NPadded_),
KPadded(KPadded_),
K0(K0_),
k_batch(k_batch_)
{
}
void Print() const
{
std::cout << "arg {"
<< "M:" << M << ", "
<< "N:" << N << ", "
<< "K:" << K << ", "
<< "SA:" << StrideA << ", "
<< "SB:" << StrideB << ", "
<< "SC:" << StrideC << ", "
<< "MP:" << MPadded << ", "
<< "NP:" << NPadded << ", "
<< "KP:" << KPadded << ", "
<< "K0:" << K0 << ", "
<< "KB:" << k_batch << "}" << std::endl;
}
};
__host__ __device__ static auto CalculateGridSize(const Argument& karg)
{
return std::make_tuple(math::integer_divide_ceil(karg.N, NPerBlock),
math::integer_divide_ceil(karg.M, MPerBlock),
karg.k_batch);
}
// prefer this to be called on host
__host__ __device__ static auto CalculateMPadded(index_t M)
{
return (M + MPerBlock - 1) / MPerBlock * MPerBlock;
}
__host__ __device__ static auto CalculateNPadded(index_t N)
{
return (N + NPerBlock - 1) / NPerBlock * NPerBlock;
}
__host__ __device__ static auto CalculateK0(index_t K, index_t K_Batch = 1)
{
// k_batch * k0 * k0_per_block * k1
auto K_t = K_Batch * K0PerBlock * K1;
return (K + K_t - 1) / K_t * K0PerBlock;
}
__host__ __device__ static auto CalculateKPadded(index_t K, index_t K_Batch = 1)
{
auto K0 = CalculateK0(K, K_Batch);
return K_Batch * K0 * K1;
}
__host__ __device__ static auto MakeAGridDescriptor_KBatch_K0_M_K1(index_t M,
index_t MPad,
index_t K,
index_t StrideA,
index_t KBatch,
index_t K0,
index_t KPad)
{
const auto a_grid_desc_m_k = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
}
}();
const auto a_grid_desc_m_kpad = transform_tensor_descriptor(
a_grid_desc_m_k,
make_tuple(make_pass_through_transform(M), make_right_pad_transform(K, KPad - K)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)
{
// const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
return transform_tensor_descriptor(
a_grid_desc_m_kpad,
make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1)),
make_right_pad_transform(M, MPad - M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
}
else
{
return transform_tensor_descriptor(
a_grid_desc_m_kpad,
make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1)),
make_pass_through_transform(M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
}
}
__host__ __device__ static auto MakeBGridDescriptor_KBatch_K0_N_K1(index_t K,
index_t NPad,
index_t N,
index_t StrideB,
index_t KBatch,
index_t K0,
index_t KPad)
{
const auto b_grid_desc_k_n = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
}
}();
const auto b_grid_desc_kpad_n = transform_tensor_descriptor(
b_grid_desc_k_n,
make_tuple(make_right_pad_transform(K, KPad - K), make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding)
{
// const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
return transform_tensor_descriptor(
b_grid_desc_kpad_n,
make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1)),
make_right_pad_transform(N, NPad - N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
}
else
{
return transform_tensor_descriptor(
b_grid_desc_kpad_n,
make_tuple(make_unmerge_transform(make_tuple(KBatch, K0, K1)),
make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 3>{}, Sequence<2>{}));
}
}
__host__ __device__ static auto
MakeCGridDescriptor_M_N(index_t M, index_t N, index_t MPad, index_t NPad, index_t StrideC)
{
const auto c_grid_desc_m_n = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
}
}();
if constexpr(GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding)
{
return transform_tensor_descriptor(c_grid_desc_m_n,
make_tuple(make_right_pad_transform(M, MPad - M),
make_right_pad_transform(N, NPad - N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else
{
return transform_tensor_descriptor(
c_grid_desc_m_n,
make_tuple(make_pass_through_transform(M), make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
}
__host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
{
constexpr auto max_lds_align = K1;
......@@ -178,45 +375,68 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
c_block_size * sizeof(FloatC));
}
// 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_B_K0_M_K1& a_b_k0_m_k1_grid_desc,
const BGridDesc_B_K0_N_K1& b_b_k0_n_k1_grid_desc,
const CMNGridDesc& c_m_n_grid_desc,
const Block2CTileMap& block_2_ctile_map)
__host__ __device__ static constexpr bool CheckValidity(const Argument& karg)
{
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 % (MPerXDL * MRepeat) == 0) &&
(NPerBlock % (NRepeat * NPerXDL)) == 0,
"Invalid tuning param!");
const auto M = a_b_k0_m_k1_grid_desc.GetLength(I2);
const auto N = b_b_k0_n_k1_grid_desc.GetLength(I2);
const auto K0 = a_b_k0_m_k1_grid_desc.GetLength(I1);
const auto KBatch = a_b_k0_m_k1_grid_desc.GetLength(I0);
if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::MPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MKPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding))
{
if(!(karg.M % MPerBlock == 0))
return false;
}
if constexpr(!(GemmSpec == tensor_operation::device::GemmSpecialization::NPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::NKPadding ||
GemmSpec == tensor_operation::device::GemmSpecialization::MNKPadding))
{
if(!(karg.N % NPerBlock == 0))
return false;
}
if(!(M == c_m_n_grid_desc.GetLength(I0) && N == c_m_n_grid_desc.GetLength(I1) &&
K0 == b_b_k0_n_k1_grid_desc.GetLength(I1) &&
K1 == a_b_k0_m_k1_grid_desc.GetLength(I3) &&
K1 == b_b_k0_n_k1_grid_desc.GetLength(I3) &&
KBatch == b_b_k0_n_k1_grid_desc.GetLength(I0)))
return false;
if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
{
if(karg.K % ABlockTransferSrcScalarPerVector != 0)
return false;
}
else
{
if(karg.M % ABlockTransferSrcScalarPerVector != 0)
return false;
}
if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K0 % K0PerBlock == 0))
return false;
if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
{
if(karg.N % BBlockTransferSrcScalarPerVector != 0)
return false;
}
else
{
if(karg.K % BBlockTransferSrcScalarPerVector != 0)
return false;
}
if(!block_2_ctile_map.CheckValidity(c_m_n_grid_desc))
if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
{
return false;
if(karg.N % CBlockTransferScalarPerVector_NWaveNPerXDL != 0)
return false;
}
else
{
if(karg.M % CBlockTransferScalarPerVector_NWaveNPerXDL != 0)
return false;
}
// TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
return true;
}
__host__ __device__ static auto GetKPad(index_t K, index_t KBatch)
{
const index_t K0 = math::integer_divide_ceil(K, K1 * K0PerBlock * KBatch) * K0PerBlock;
const index_t KPad = KBatch * K0 * K1;
return KPad;
}
__host__ __device__ static constexpr bool CalculateHasMainK0BlockLoop(index_t K0)
{
const bool has_main_k0_block_loop = K0 > K0PerBlock;
......@@ -224,8 +444,9 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
return has_main_k0_block_loop;
}
template <typename CGridDesc>
__host__ __device__ static constexpr auto
MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(const CMNGridDesc& c_m_n_grid_desc)
MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(const CGridDesc& c_m_n_grid_desc)
{
const auto M = c_m_n_grid_desc.GetLength(I0);
const auto N = c_m_n_grid_desc.GetLength(I1);
......@@ -242,10 +463,11 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
}
// return block_id to C matrix tile idx (m0, n0) mapping
template <typename CGridDesc>
__host__ __device__ static constexpr auto MakeCBlockClusterAdaptor(
const CMNGridDesc& c_m_n_grid_desc, index_t /* M01 */, index_t /* N01 */, index_t KBatch)
const CGridDesc& c_m_n_grid_desc, index_t /* M01 */, index_t /* N01 */, index_t KBatch)
{
return BlockToCTileMap_KSplit_M00_N0_M01Adapt<MPerBlock, NPerBlock, CMNGridDesc>(
return BlockToCTileMap_KSplit_M00_N0_M01Adapt<MPerBlock, NPerBlock, CGridDesc>(
c_m_n_grid_desc, 8, KBatch);
}
......@@ -262,24 +484,38 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
Number<CShuffleNRepeatPerShuffle * NWave * NPerXDL>{}));
}
using CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock =
decltype(MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(CMNGridDesc{}));
using CBlockClusterAdaptor = decltype(MakeCBlockClusterAdaptor(CMNGridDesc{}, 1, 1, 1));
// return block_id to C matrix tile idx (m0, n0, k_split) mapping
__host__ __device__ static constexpr auto MakeDefaultBlock2CTileMap()
{
return BlockToCTileMap_3DGrid_KSplit<MPerBlock, NPerBlock>();
}
using CGridDesc_M_N = remove_cvref_t<decltype(MakeCGridDescriptor_M_N(1, 1, 1, 1, 1))>;
using DefaultBlock2CTileMap = remove_cvref_t<decltype(MakeDefaultBlock2CTileMap())>;
template <bool HasMainKBlockLoop>
__device__ static void Run(const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid,
template <bool HasMainKBlockLoop,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
typename Block2CTileMap>
__device__ static void Run(const Argument& karg,
void* __restrict__ p_shared_block,
const AGridDesc_B_K0_M_K1& a_b_k0_m_k1_grid_desc,
const BGridDesc_B_K0_N_K1& b_b_k0_n_k1_grid_desc,
const CGridDesc_MBlock_MPerBlock_NBlock_NPerBlock&
c_grid_desc_mblock_mperblock_nblock_nperblock,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const CElementwiseOperation& c_element_op,
const CBlockClusterAdaptor& c_block_cluster_adaptor)
const Block2CTileMap& block_2_ctile_map)
{
const FloatAB* p_a_grid = karg.p_a_grid;
const FloatAB* p_b_grid = karg.p_b_grid;
FloatC* p_c_grid = karg.p_c_grid;
const auto a_b_k0_m_k1_grid_desc = MakeAGridDescriptor_KBatch_K0_M_K1(
karg.M, karg.MPadded, karg.K, karg.StrideA, karg.k_batch, karg.K0, karg.KPadded);
const auto b_b_k0_n_k1_grid_desc = MakeBGridDescriptor_KBatch_K0_N_K1(
karg.K, karg.NPadded, karg.N, karg.StrideB, karg.k_batch, karg.K0, karg.KPadded);
const auto c_grid_desc_m_n =
MakeCGridDescriptor_M_N(karg.M, karg.N, karg.MPadded, karg.NPadded, karg.StrideC);
const auto c_grid_desc_mblock_mperblock_nblock_nperblock =
MakeCGridDesc_MBlock_MPerBlock_NBlock_NPerBlock(c_grid_desc_m_n);
const AElementwiseOperation a_element_op = AElementwiseOperation{};
const BElementwiseOperation b_element_op = BElementwiseOperation{};
const CElementwiseOperation c_element_op = CElementwiseOperation{};
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_grid, a_b_k0_m_k1_grid_desc.GetElementSpaceSize());
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
......@@ -287,28 +523,28 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
const auto K0 = a_b_k0_m_k1_grid_desc.GetLength(I1);
// divide block work by [M, N]
// divide block work by [KBatch, M, N]
const auto block_work_idx =
c_block_cluster_adaptor.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
block_2_ctile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
if(!c_block_cluster_adaptor.ValidCTileIndex(
make_tuple(block_work_idx[I1], block_work_idx[I2]),
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;
}
const index_t k_batch_id = block_work_idx[I0];
const index_t block_m_id = __builtin_amdgcn_readfirstlane(block_work_idx[I1]);
const index_t block_n_id = __builtin_amdgcn_readfirstlane(block_work_idx[I2]);
const index_t k_batch_id = __builtin_amdgcn_readfirstlane(block_work_idx[I0]);
// HACK: this force m/n_block_data_idx_on_grid into SGPR
const index_t m_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_work_idx[I1] * MPerBlock);
__builtin_amdgcn_readfirstlane(block_m_id * MPerBlock);
const index_t n_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_work_idx[I2] * NPerBlock);
__builtin_amdgcn_readfirstlane(block_n_id * NPerBlock);
// lds max alignment
constexpr auto max_lds_align = K1;
......@@ -444,7 +680,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
// c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
// register
// sanity check
#if 1
auto blockwise_gemm =
BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<BlockSize,
FloatAB,
......@@ -456,6 +692,20 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
MRepeat,
NRepeat,
K1>{};
#else
auto blockwise_gemm = BlockwiseGemmXdlopsInterwave_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_v1<
BlockSize,
FloatAB,
FloatAcc,
decltype(a_k0_m_k1_block_desc),
decltype(b_k0_n_k1_block_desc),
MPerXDL,
NPerXDL,
MRepeat,
NRepeat,
K1>{};
#endif
auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
......@@ -474,6 +724,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
p_b_block, b_k0_n_k1_block_desc.GetElementSpaceSize());
#if 0
// preload data into LDS
{
a_blockwise_copy.RunRead(a_b_k0_m_k1_grid_desc, a_grid_buf);
......@@ -519,6 +770,31 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
}
#else
// gridwise GEMM pipeline
const auto gridwise_gemm_pipeline =
GridwiseGemmPipeline_Selector<PipelineVersion::v2, 1, LoopScheduler::Default>();
const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
(a_b_k0_m_k1_grid_desc.GetLength(I1) * a_b_k0_m_k1_grid_desc.GetLength(I3)) /
(K0PerBlock * K1));
gridwise_gemm_pipeline.template Run<HasMainKBlockLoop>(a_b_k0_m_k1_grid_desc,
a_b_k0_m_k1_block_desc,
a_blockwise_copy,
a_grid_buf,
a_block_buf,
a_block_slice_copy_step,
b_b_k0_n_k1_grid_desc,
b_b_k0_n_k1_block_desc,
b_blockwise_copy,
b_grid_buf,
b_block_buf,
b_block_slice_copy_step,
blockwise_gemm,
c_thread_buf,
num_k_block_main_loop);
#endif
// output: register to global memory
{
......@@ -647,7 +923,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
{c_block_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(0, 0, 0, 0),
c_grid_desc_mblock_mperblock_nblock_nperblock,
make_multi_index(block_work_idx[I1], 0, block_work_idx[I2], 0),
make_multi_index(block_m_id, 0, block_n_id, 0),
c_element_op};
constexpr auto mxdlperwave_forward_step =
......@@ -716,6 +992,48 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
});
}
}
template <typename Layout>
struct LStr
{
static std::string Get() { return ""; }
};
template <>
struct LStr<ck::tensor_layout::gemm::RowMajor>
{
static std::string Get() { return "R"; }
};
template <>
struct LStr<ck::tensor_layout::gemm::ColumnMajor>
{
static std::string Get() { return "C"; }
};
static std::string GetTypeString()
{
auto str = std::stringstream();
// clang-format off
str << "GemmXdlSplitKCShuffle_"
<< getGemmSpecializationString(GemmSpec) << "_"
<< std::string(ALayout::name)[0]
<< std::string(BLayout::name)[0]
<< std::string(CLayout::name)[0]
<< "_"
<< "B" << BlockSize << "_"
<< "Vec" << ABlockTransferSrcScalarPerVector << "x"
<< BBlockTransferSrcScalarPerVector << "x"
<< CBlockTransferScalarPerVector_NWaveNPerXDL << "_"
<< MPerBlock << "x"
<< NPerBlock << "x"
<< K0PerBlock << "x"
<< K1 ;
// clang-format on
return str.str();
}
};
} // namespace ck
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp
View file @ 0d6da13c
......@@ -46,7 +46,8 @@ __global__ void
const CElementwiseOperation c_element_op,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainK0BlockLoop>(
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp
View file @ 0d6da13c
......@@ -49,7 +49,8 @@ __global__ void
const CElementwiseOperation c_element_op,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp
View file @ 0d6da13c
......@@ -53,7 +53,8 @@ __global__ void
const CElementwiseOperation c_element_op,
const Block2CTileMap block_2_ctile_map)
{
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__))
#if(!defined(__HIP_DEVICE_COMPILE__) || defined(__gfx908__) || defined(__gfx90a__) || \
defined(__gfx940__))
__shared__ char p_shared[GridwiseGemm::GetSharedMemoryNumberOfByte()];
GridwiseGemm::template Run<HasMainKBlockLoop>(
......
include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp
View file @ 0d6da13c
......@@ -6,6 +6,7 @@
#include "ck/utility/common_header.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
#include "ck/tensor/static_tensor.hpp"
......@@ -207,15 +208,6 @@ struct ThreadwiseTensorSliceTransfer_v3r1
auto src_vector_container = src_vector_type{
src_buf.template Get<src_vector_t>(src_coord_.GetOffset(), is_src_valid)};
// apply SrcElementwiseOperation on src_vector_container
static_for<0, SrcScalarPerVector, 1>{}([&](auto i) {
SrcData src_v;
src_element_op_(src_v, src_vector_container.template AsType<SrcData>()[i]);
src_vector_container.template AsType<SrcData>()(i) = src_v;
});
// copy data from src_vector_container into src_thread_scratch_
src_thread_scratch_tuple_(thread_scratch_id)
.template SetAsType<src_vector_t>(
......@@ -318,7 +310,6 @@ struct ThreadwiseTensorSliceTransfer_v3r1
constexpr auto data_idx_seq = generate_sequence_v2(
[&](auto i) { return Number<data_idx[i]>{}; }, Number<nDim>{});
// TODO type_convert is not used yet!!!!!
using src_vector_t = vector_type_maker_t<SrcData, SrcScalarPerVector>;
using dst_vector_t = vector_type_maker_t<DstData, DstScalarPerVector>;
......@@ -342,19 +333,17 @@ struct ThreadwiseTensorSliceTransfer_v3r1
Number<num_dst_vector>{});
// do data transpose
// TODO type_convert is not used yet!!!!!
transpose_vectors<SrcData, DstScalarPerVector, SrcScalarPerVector>{}(
src_vector_refs, dst_vector_refs);
});
}
else
{
static_ford<SliceLengths>{}([&](auto idx) {
// convert from SrcData to DstData here
dst_thread_scratch_(idx) =
type_convert<DstData>(src_thread_scratch_tuple_[thread_scratch_id][idx]);
});
}
static_ford<SliceLengths>{}([&](auto idx) {
// apply the src elementwise op and convert to DstData under the hood if needed
DstData dst_v;
src_element_op_(dst_v, src_thread_scratch_tuple_[thread_scratch_id][idx]);
dst_thread_scratch_(idx) = dst_v;
});
#endif
}
......
include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp
View file @ 0d6da13c
......@@ -27,6 +27,8 @@ enum struct MfmaInstr
mfma_f32_16x16x8bf16,
mfma_i32_32x32x8i8,
mfma_i32_16x16x16i8,
mfma_i32_32x32x16i8,
mfma_i32_16x16x32i8,
mfma_f64_16x16x4f64
};
......@@ -386,6 +388,50 @@ struct mfma_type<MfmaInstr::mfma_i32_16x16x16i8>
}
};
template <>
struct mfma_type<MfmaInstr::mfma_i32_32x32x16i8>
{
static constexpr index_t group_size = 4;
static constexpr index_t num_groups_per_blk = 4;
static constexpr index_t num_regs_per_blk = 16;
static constexpr index_t num_threads_per_blk = 32;
static constexpr index_t wave_size = 64;
static constexpr index_t num_input_blks = 2;
static constexpr index_t num_output_blks = 1;
static constexpr index_t m_per_blk = 32;
static constexpr index_t n_per_blk = 32;
static constexpr index_t k_per_blk = 8;
static constexpr bool is_k_reduction = true;
template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
__device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
{
intrin_mfma_i32_32x32x16i8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
}
};
template <>
struct mfma_type<MfmaInstr::mfma_i32_16x16x32i8>
{
static constexpr index_t group_size = 4;
static constexpr index_t num_groups_per_blk = 1;
static constexpr index_t num_regs_per_blk = 4;
static constexpr index_t num_threads_per_blk = 16;
static constexpr index_t wave_size = 64;
static constexpr index_t num_input_blks = 4;
static constexpr index_t num_output_blks = 1;
static constexpr index_t m_per_blk = 16;
static constexpr index_t n_per_blk = 16;
static constexpr index_t k_per_blk = 8;
static constexpr bool is_k_reduction = true;
template <index_t MPerXdlops, index_t NPerXdlops, class FloatA, class FloatB, class FloatC>
__device__ void run(const FloatA& a, const FloatB& b, FloatC& reg_c) const
{
intrin_mfma_i32_16x16x32i8<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
}
};
template <>
struct mfma_type<MfmaInstr::mfma_f64_16x16x4f64>
{
......@@ -524,17 +570,29 @@ struct MfmaSelector
#endif
}
#if defined(CK_USE_AMD_MFMA_GFX940)
template <>
static constexpr auto GetMfma<int8_t, 32, 32>()
{
return MfmaInstr::mfma_i32_32x32x16i8;
}
template <>
static constexpr auto GetMfma<int8_t, 16, 16>()
{
return MfmaInstr::mfma_i32_16x16x32i8;
}
#else
template <>
static constexpr auto GetMfma<int8_t, 32, 32>()
{
return MfmaInstr::mfma_i32_32x32x8i8;
}
template <>
static constexpr auto GetMfma<int8_t, 16, 16>()
{
return MfmaInstr::mfma_i32_16x16x16i8;
}
#endif
static constexpr auto selected_mfma = mfma_type<GetMfma<base_type, MPerXdlops, NPerXdlops>()>{};
......
include/ck/utility/amd_xdlops.hpp
View file @ 0d6da13c
......@@ -297,6 +297,44 @@ struct intrin_mfma_i32_16x16x16i8<16, 16>
}
};
template <index_t MPerWave, index_t NPerWave>
struct intrin_mfma_i32_32x32x16i8;
template <>
struct intrin_mfma_i32_32x32x16i8<32, 32>
{
template <class FloatC>
__device__ static void Run(const int8x8_t& reg_a, const int8x8_t& reg_b, FloatC& reg_c)
{
reg_c.template AsType<int32x16_t>()(Number<0>{}) =
__builtin_amdgcn_mfma_i32_32x32x16_i8(bit_cast<int64_t>(reg_a),
bit_cast<int64_t>(reg_b),
reg_c.template AsType<int32x16_t>()[Number<0>{}],
0,
0,
0);
}
};
template <index_t MPerWave, index_t NPerWave>
struct intrin_mfma_i32_16x16x32i8;
template <>
struct intrin_mfma_i32_16x16x32i8<16, 16>
{
template <class FloatC>
__device__ static void Run(const int8x8_t& reg_a, const int8x8_t& reg_b, FloatC& reg_c)
{
reg_c.template AsType<int32x4_t>()(Number<0>{}) =
__builtin_amdgcn_mfma_i32_16x16x32i8(bit_cast<int64_t>(reg_a),
bit_cast<int64_t>(reg_b),
reg_c.template AsType<int32x4_t>()[Number<0>{}],
0,
0,
0);
}
};
template <index_t MPerWave, index_t NPerWave>
struct intrin_mfma_f64_16x16x4f64;
......@@ -306,7 +344,7 @@ struct intrin_mfma_f64_16x16x4f64<16, 16>
template <class FloatC>
__device__ static void Run(const double& reg_a, const double& reg_b, FloatC& reg_c)
{
#ifdef __gfx90a__
#if defined(__gfx90a__) || defined(__gfx940__)
reg_c.template AsType<double4_t>()(Number<0>{}) = __builtin_amdgcn_mfma_f64_16x16x4f64(
reg_a, reg_b, reg_c.template AsType<double4_t>()[Number<0>{}], 0, 0, 0);
#else
......
include/ck/utility/data_type.hpp
View file @ 0d6da13c
......@@ -898,6 +898,8 @@ struct vector_type<T, 256>
}
};
using int64_t = long;
// fp64
using double2_t = typename vector_type<double, 2>::type;
using double4_t = typename vector_type<double, 4>::type;
......@@ -974,37 +976,6 @@ inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, float>(float
uint32_t int32;
} u = {x};
// When the exponent bits are not all 1s, then the value is zero, normal,
// or subnormal. We round the bfloat16 mantissa up by adding 0x7FFF, plus
// 1 if the least significant bit of the bfloat16 mantissa is 1 (odd).
// This causes the bfloat16's mantissa to be incremented by 1 if the 16
// least significant bits of the float mantissa are greater than 0x8000,
// or if they are equal to 0x8000 and the least significant bit of the
// bfloat16 mantissa is 1 (odd). This causes it to be rounded to even when
// the lower 16 bits are exactly 0x8000. If the bfloat16 mantissa already
// has the value 0x7f, then incrementing it causes it to become 0x00 and
// the exponent is incremented by one, which is the next higher FP value
// to the unrounded bfloat16 value. When the bfloat16 value is subnormal
// with an exponent of 0x00 and a mantissa of 0x7f, it may be rounded up
// to a normal value with an exponent of 0x01 and a mantissa of 0x00.
// When the bfloat16 value has an exponent of 0xFE and a mantissa of 0x7F,
// incrementing it causes it to become an exponent of 0xFF and a mantissa
// of 0x00, which is Inf, the next higher value to the unrounded value.
bool flag0 = ~u.int32 & 0x7f800000;
// When all of the exponent bits are 1, the value is Inf or NaN.
// Inf is indicated by a zero mantissa. NaN is indicated by any nonzero
// mantissa bit. Quiet NaN is indicated by the most significant mantissa
// bit being 1. Signaling NaN is indicated by the most significant
// mantissa bit being 0 but some other bit(s) being 1. If any of the
// lower 16 bits of the mantissa are 1, we set the least significant bit
// of the bfloat16 mantissa, in order to preserve signaling NaN in case
// the bfloat16's mantissa bits are all 0.
bool flag1 = !flag0 && (u.int32 & 0xffff);
u.int32 += flag0 ? 0x7fff + ((u.int32 >> 16) & 1) : 0; // Round to nearest, round to even
u.int32 |= flag1 ? 0x10000 : 0x0; // Preserve signaling NaN
return uint16_t(u.int32 >> 16);
}
......@@ -1062,6 +1033,63 @@ inline __host__ __device__ constexpr bhalf_t type_convert<bhalf_t, int8_t>(int8_
return type_convert<bhalf_t>(x_fp32);
}
// Declare a template function for bf16 conversion using RTN
template <typename Y, typename X>
__host__ __device__ constexpr Y bf16_convert_rtn(X x);
// Convert fp32 to bf16 with RTN if higher precision is needed
template <>
inline __host__ __device__ constexpr bhalf_t bf16_convert_rtn<bhalf_t, float>(float x)
{
union
{
float fp32;
uint32_t int32;
} u = {x};
// When the exponent bits are not all 1s, then the value is zero, normal,
// or subnormal. We round the bfloat16 mantissa up by adding 0x7FFF, plus
// 1 if the least significant bit of the bfloat16 mantissa is 1 (odd).
// This causes the bfloat16's mantissa to be incremented by 1 if the 16
// least significant bits of the float mantissa are greater than 0x8000,
// or if they are equal to 0x8000 and the least significant bit of the
// bfloat16 mantissa is 1 (odd). This causes it to be rounded to even when
// the lower 16 bits are exactly 0x8000. If the bfloat16 mantissa already
// has the value 0x7f, then incrementing it causes it to become 0x00 and
// the exponent is incremented by one, which is the next higher FP value
// to the unrounded bfloat16 value. When the bfloat16 value is subnormal
// with an exponent of 0x00 and a mantissa of 0x7f, it may be rounded up
// to a normal value with an exponent of 0x01 and a mantissa of 0x00.
// When the bfloat16 value has an exponent of 0xFE and a mantissa of 0x7F,
// incrementing it causes it to become an exponent of 0xFF and a mantissa
// of 0x00, which is Inf, the next higher value to the unrounded value.
bool flag0 = ~u.int32 & 0x7f800000;
// When all of the exponent bits are 1, the value is Inf or NaN.
// Inf is indicated by a zero mantissa. NaN is indicated by any nonzero
// mantissa bit. Quiet NaN is indicated by the most significant mantissa
// bit being 1. Signaling NaN is indicated by the most significant
// mantissa bit being 0 but some other bit(s) being 1. If any of the
// lower 16 bits of the mantissa are 1, we set the least significant bit
// of the bfloat16 mantissa, in order to preserve signaling NaN in case
// the bfloat16's mantissa bits are all 0.
bool flag1 = !flag0 && (u.int32 & 0xffff);
u.int32 += flag0 ? 0x7fff + ((u.int32 >> 16) & 1) : 0; // Round to nearest, round to even
u.int32 |= flag1 ? 0x10000 : 0x0; // Preserve signaling NaN
return uint16_t(u.int32 >> 16);
}
// convert fp16 to bfp16 via fp32 with RTN if higher precision is needed
template <>
inline __host__ __device__ constexpr bhalf_t bf16_convert_rtn<bhalf_t, half_t>(half_t x)
{
float x_fp32 = static_cast<float>(x);
return bf16_convert_rtn<bhalf_t>(x_fp32);
}
template <typename T>
struct NumericLimits
{
......
library/include/ck/library/reference_tensor_operation/cpu/reference_gemm.hpp
View file @ 0d6da13c
......@@ -6,6 +6,7 @@
#include <iostream>
#include <sstream>
#include "ck/tensor_operation/gpu/element/unary_element_wise_operation.hpp"
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/library/utility/host_tensor.hpp"
......@@ -66,8 +67,26 @@ struct ReferenceGemm : public device::BaseOperator
ADataType v_a;
BDataType v_b;
arg.a_element_op_(v_a, arg.a_m_k_(m, k));
arg.b_element_op_(v_b, arg.b_k_n_(k, n));
// use PassThrough instead of ConvertBF16RTN for reference calculation
if constexpr(is_same_v<AElementwiseOperation,
ck::tensor_operation::element_wise::ConvertBF16RTN>)
{
ck::tensor_operation::element_wise::PassThrough{}(v_a, arg.a_m_k_(m, k));
}
else
{
arg.a_element_op_(v_a, arg.a_m_k_(m, k));
}
// same for B matrix
if constexpr(is_same_v<BElementwiseOperation,
ck::tensor_operation::element_wise::ConvertBF16RTN>)
{
ck::tensor_operation::element_wise::PassThrough{}(v_b, arg.b_k_n_(k, n));
}
else
{
arg.b_element_op_(v_b, arg.b_k_n_(k, n));
}
v_acc +=
ck::type_convert<AccDataType>(v_a) * ck::type_convert<AccDataType>(v_b);
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_convolution_forward.hpp
View file @ 0d6da13c
......@@ -117,20 +117,6 @@ void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_f32_instances(
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GNHWC,
GKYXC,
Empty_Tuple,
GNHWK,
int8_t,
int8_t,
Empty_Tuple,
int8_t,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GNHWC,
......@@ -159,20 +145,21 @@ void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_f32_instances(
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instances(
// grouped conv2d forward, NHWGC/GKYXC/NHWGK
void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GNHWC,
NHWGC,
GKYXC,
Empty_Tuple,
GNHWK,
int8_t,
int8_t,
NHWGK,
BF16,
BF16,
Empty_Tuple,
int8_t,
BF16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
// grouped conv2d forward, NHWGC/GKYXC/NHWGK
void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
NHWGC,
......@@ -187,6 +174,20 @@ void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f16_instances(
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
NHWGC,
GKYXC,
Empty_Tuple,
NHWGK,
F32,
F32,
Empty_Tuple,
F32,
PassThrough,
PassThrough,
PassThrough>>>& instances);
// grouped conv3d forward, GNDHWC/GKZYXC/GNDHWK
void add_device_grouped_conv3d_fwd_xdl_gndhwc_gkzyxc_gndhwk_bf16_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<3,
......@@ -385,12 +386,6 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{
add_device_grouped_conv1d_fwd_xdl_gnhwc_gkyxc_gnhwk_bf16_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>)
{
add_device_grouped_conv2d_fwd_xdl_gnhwc_gkyxc_gnhwk_int8_instances(op_ptrs);
add_device_grouped_conv2d_fwd_dl_gnhwc_gkyxc_gnhwk_int8_instances(op_ptrs);
}
}
else if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> &&
is_same_v<WeiLayout, GKYXC> && is_same_v<OutLayout, NHWGK>)
......@@ -398,7 +393,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
if constexpr(is_same_v<InDataType, float> && is_same_v<WeiDataType, float> &&
is_same_v<OutDataType, float>)
{
// no instance
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_f32_instances(op_ptrs);
}
else if constexpr(is_same_v<InDataType, half_t> && is_same_v<WeiDataType, half_t> &&
is_same_v<OutDataType, half_t>)
......@@ -409,12 +404,7 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<WeiDataType, ck::bhalf_t> &&
is_same_v<OutDataType, ck::bhalf_t>)
{
// no instance
}
else if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<OutDataType, int8_t>)
{
// no instance
add_device_grouped_conv2d_fwd_xdl_nhwgc_gkyxc_nhwgk_bf16_instances(op_ptrs);
}
}
else if constexpr(NumDimSpatial == 3 && is_same_v<InLayout, GNDHWC> &&
......
library/include/ck/library/tensor_operation_instance/gpu/grouped_gemm.hpp
View file @ 0d6da13c
......@@ -68,6 +68,58 @@ void add_device_grouped_gemm_xdl_f16_f16_f16_km_nk_mn_instances(
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
Col,
Empty_Tuple,
Row,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
Row,
Empty_Tuple,
Row,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instances(
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
Col,
Empty_Tuple,
Row,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
void add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instances(
std::vector<std::unique_ptr<DeviceGroupedGemm<Row,
Row,
Empty_Tuple,
Row,
F16,
F16,
Empty_Tuple,
F16,
PassThrough,
PassThrough,
PassThrough>>>& instances);
template <typename ALayout,
typename BLayout,
typename ELayout,
......@@ -109,11 +161,17 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
is_same_v<ELayout, Row>)
{
add_device_grouped_gemm_xdl_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_instances(op_ptrs);
add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_kn_mn_irregular_instances(
op_ptrs);
}
else if constexpr(is_same_v<ALayout, Row> && is_same_v<BLayout, Col> &&
is_same_v<ELayout, Row>)
{
add_device_grouped_gemm_xdl_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instances(op_ptrs);
add_device_grouped_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_irregular_instances(
op_ptrs);
}
else if constexpr(is_same_v<ALayout, Col> && is_same_v<BLayout, Row> &&
is_same_v<ELayout, Row>)
......
library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perchannel_quantization.hpp
View file @ 0d6da13c
......@@ -17,14 +17,14 @@ namespace tensor_operation {
namespace device {
namespace instance {
// grouped conv2d forward, GNHWC/GKYXC/GNHWK
// grouped conv2d forward, NHWGC/GKYXC/NHWGK
void add_device_conv2d_dl_bias_perchannel_quantization_int8_instances(
std::vector<
std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GNHWC,
NHWGC,
GKYXC,
GK_GK_Tuple,
GNHWK,
NHWGK,
int8_t,
int8_t,
I32_F32_Tuple,
......@@ -36,10 +36,10 @@ void add_device_conv2d_dl_bias_perchannel_quantization_int8_instances(
void add_device_conv2d_dl_bias_relu_perchannel_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GNHWC,
NHWGC,
GKYXC,
GK_GK_Tuple,
GNHWK,
NHWGK,
int8_t,
int8_t,
I32_F32_Tuple,
......@@ -52,10 +52,10 @@ void add_device_conv2d_dl_bias_relu_perchannel_quantization_int8_instances(
void add_device_conv2d_dl_bias_tanh_perchannel_quantization_int8_instances(
std::vector<
std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GNHWC,
NHWGC,
GKYXC,
GK_GK_Tuple,
GNHWK,
NHWGK,
int8_t,
int8_t,
I32_F32_Tuple,
......@@ -68,10 +68,10 @@ void add_device_conv2d_dl_bias_tanh_perchannel_quantization_int8_instances(
void add_device_conv2d_xdl_bias_perchannel_quantization_int8_instances(
std::vector<
std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GNHWC,
NHWGC,
GKYXC,
GK_GK_Tuple,
GNHWK,
NHWGK,
int8_t,
int8_t,
I32_F32_Tuple,
......@@ -83,10 +83,10 @@ void add_device_conv2d_xdl_bias_perchannel_quantization_int8_instances(
void add_device_conv2d_xdl_bias_relu_perchannel_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GNHWC,
NHWGC,
GKYXC,
GK_GK_Tuple,
GNHWK,
NHWGK,
int8_t,
int8_t,
I32_F32_Tuple,
......@@ -99,10 +99,10 @@ void add_device_conv2d_xdl_bias_relu_perchannel_quantization_int8_instances(
void add_device_conv2d_xdl_bias_tanh_perchannel_quantization_int8_instances(
std::vector<
std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GNHWC,
NHWGC,
GKYXC,
GK_GK_Tuple,
GNHWK,
NHWGK,
int8_t,
int8_t,
I32_F32_Tuple,
......@@ -154,9 +154,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> &&
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> &&
is_same_v<WeiLayout, GKYXC> && is_same_v<DsLayout, GK_GK_Tuple> &&
is_same_v<OutLayout, GNHWK>)
is_same_v<OutLayout, NHWGK>)
{
if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<DsDataType, I32_F32_Tuple> && is_same_v<OutDataType, int8_t>)
......@@ -220,9 +220,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> &&
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> &&
is_same_v<WeiLayout, GKYXC> && is_same_v<DsLayout, GK_GK_Tuple> &&
is_same_v<OutLayout, GNHWK>)
is_same_v<OutLayout, NHWGK>)
{
if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<DsDataType, I32_F32_Tuple> && is_same_v<OutDataType, int8_t>)
......
library/include/ck/library/tensor_operation_instance/gpu/quantization/grouped_convolution_bias_forward_perlayer_quantization.hpp
View file @ 0d6da13c
......@@ -17,14 +17,14 @@ namespace tensor_operation {
namespace device {
namespace instance {
// grouped conv2d forward, GNHWC/GKYXC/GNHWK
// grouped conv2d forward, NHWGC/GKYXC/NHWGK
void add_device_conv2d_dl_bias_perlayer_quantization_int8_instances(
std::vector<
std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GNHWC,
NHWGC,
GKYXC,
GK_Tuple,
GNHWK,
NHWGK,
int8_t,
int8_t,
I32_Tuple,
......@@ -36,10 +36,10 @@ void add_device_conv2d_dl_bias_perlayer_quantization_int8_instances(
void add_device_conv2d_dl_bias_relu_perlayer_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GNHWC,
NHWGC,
GKYXC,
GK_Tuple,
GNHWK,
NHWGK,
int8_t,
int8_t,
I32_Tuple,
......@@ -51,10 +51,10 @@ void add_device_conv2d_dl_bias_relu_perlayer_quantization_int8_instances(
void add_device_conv2d_dl_bias_tanh_perlayer_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GNHWC,
NHWGC,
GKYXC,
GK_Tuple,
GNHWK,
NHWGK,
int8_t,
int8_t,
I32_Tuple,
......@@ -67,10 +67,10 @@ void add_device_conv2d_dl_bias_tanh_perlayer_quantization_int8_instances(
void add_device_conv2d_xdl_bias_perlayer_quantization_int8_instances(
std::vector<
std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GNHWC,
NHWGC,
GKYXC,
GK_Tuple,
GNHWK,
NHWGK,
int8_t,
int8_t,
I32_Tuple,
......@@ -82,10 +82,10 @@ void add_device_conv2d_xdl_bias_perlayer_quantization_int8_instances(
void add_device_conv2d_xdl_bias_relu_perlayer_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GNHWC,
NHWGC,
GKYXC,
GK_Tuple,
GNHWK,
NHWGK,
int8_t,
int8_t,
I32_Tuple,
......@@ -97,10 +97,10 @@ void add_device_conv2d_xdl_bias_relu_perlayer_quantization_int8_instances(
void add_device_conv2d_xdl_bias_tanh_perlayer_quantization_int8_instances(
std::vector<std::unique_ptr<DeviceGroupedConvFwdMultipleD<2,
GNHWC,
NHWGC,
GKYXC,
GK_Tuple,
GNHWK,
NHWGK,
int8_t,
int8_t,
I32_Tuple,
......@@ -152,9 +152,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> &&
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> &&
is_same_v<WeiLayout, GKYXC> && is_same_v<DsLayout, GK_Tuple> &&
is_same_v<OutLayout, GNHWK>)
is_same_v<OutLayout, NHWGK>)
{
if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<DsDataType, I32_Tuple> && is_same_v<OutDataType, int8_t>)
......@@ -218,9 +218,9 @@ struct DeviceOperationInstanceFactory<ck::tensor_operation::device::DeviceGroupe
{
std::vector<std::unique_ptr<DeviceOp>> op_ptrs;
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, GNHWC> &&
if constexpr(NumDimSpatial == 2 && is_same_v<InLayout, NHWGC> &&
is_same_v<WeiLayout, GKYXC> && is_same_v<DsLayout, GK_Tuple> &&
is_same_v<OutLayout, GNHWK>)
is_same_v<OutLayout, NHWGK>)
{
if constexpr(is_same_v<InDataType, int8_t> && is_same_v<WeiDataType, int8_t> &&
is_same_v<DsDataType, I32_Tuple> && is_same_v<OutDataType, int8_t>)
......
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