Skip to content

GitLab

Commit f0d63f25 authored by wangshaojie6's avatar wangshaojie6
Browse files

add some code

parent 7c7364a6
Show whitespace changes
Inline Side-by-side
Showing with 903 additions and 55 deletions
example/42_splitK_gemm_bias/run_splitK_gemm_bias_example.inc
View file @ f0d63f25
...@@ -30,7 +30,7 @@ bool run_splitK_gemm_bias(const ProblemSize& problem_size, const ExecutionConfig ...@@ -30,7 +30,7 @@ bool run_splitK_gemm_bias(const ProblemSize& problem_size, const ExecutionConfig
static_assert(sizeof(BDataType) == sizeof(KernelBDataType)); static_assert(sizeof(BDataType) == sizeof(KernelBDataType));
#endif #endif
auto& [M, N, K, StrideA, StrideB, StrideC, KBatch] = problem_size; auto& [M, N, K, StrideA, StrideB, StrideE, KBatch] = problem_size;
auto f_host_tensor_descriptor = auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) { [](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
...@@ -48,21 +48,36 @@ bool run_splitK_gemm_bias(const ProblemSize& problem_size, const ExecutionConfig ...@@ -48,21 +48,36 @@ bool run_splitK_gemm_bias(const ProblemSize& problem_size, const ExecutionConfig
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{})); Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{})); Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<CDataType> c_m_n_device_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{})); Tensor<DDataType> d_m_n(f_host_tensor_descriptor(M, N, 0, ELayout{}));
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
// A[M0, M1, K0, K1] Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
std::vector<ck::index_t> a_ms_ks_lengths{M, KBatch, 64};
std::vector<ck::index_t> a_ms_ks_strides{524288, 4096, 128, 1};
// B[N0, N1, K0, K1]
std::vector<ck::index_t> b_ns_ks_lengths{32, 64, 32, 64};
std::vector<ck::index_t> b_ns_ks_strides{524288, 4096, 128, 1};
// E[M0, M1, N0, N1]
std::vector<ck::index_t> e_ms_ns_lengths{30, 128, 32, 64};
std::vector<ck::index_t> e_ms_ns_strides{524288, 4096, 128, 1};
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl; std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl; std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "c_m_n: " << c_m_n_device_result.mDesc << std::endl; std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
auto f_tensor_length_stride_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout){
if (std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{
return {std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1})};
}
else
{
return {std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride})};
}
};
std::vector<ck::index_t> a_ms_ks_lengths = f_tensor_length_stride_descriptor(M, K, StrideA, ALayout{})[0];
std::vector<ck::index_t> a_ms_ks_strides = f_tensor_length_stride_descriptor(M, K, StrideA, ALayout{})[1];
std::vector<ck::index_t> b_ns_ks_lengths = f_tensor_length_stride_descriptor(N, K, StrideB, Row{})[0];
std::vector<ck::index_t> b_ns_ks_strides = f_tensor_length_stride_descriptor(N, K, StrideB, Row{})[1];
std::vector<ck::index_t> d_ms_ns_lengths = f_tensor_length_stride_descriptor(M, N, 0, Row{})[0];
std::vector<ck::index_t> d_ms_ns_strides = f_tensor_length_stride_descriptor(M, N, 0, Row{})[1];
std::vector<ck::index_t> e_ms_ns_lengths = f_tensor_length_stride_descriptor(M, N, StrideE, ELayout{})[0];
std::vector<ck::index_t> e_ms_ns_strides = f_tensor_length_stride_descriptor(M, N, StrideE, ELayout{})[1];
switch(config.init_method) switch(config.init_method)
{ {
...@@ -70,38 +85,45 @@ bool run_splitK_gemm_bias(const ProblemSize& problem_size, const ExecutionConfig ...@@ -70,38 +85,45 @@ bool run_splitK_gemm_bias(const ProblemSize& problem_size, const ExecutionConfig
case 1: case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5}); a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5}); b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d_m_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
break; break;
case 2: case 2:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0}); a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5}); b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d_m_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
break; break;
default: default:
a_m_k.GenerateTensorValue(GeneratorTensor_Sequential<0>{}); a_m_k.GenerateTensorValue(GeneratorTensor_Sequential<0>{});
b_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{}); b_k_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
d_m_n.GenerateTensorValue(GeneratorTensor_Sequential<1>{});
} }
DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize()); DeviceMem a_m_k_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpaceSize());
DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize()); DeviceMem b_k_n_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpaceSize());
DeviceMem c_m_n_device_buf(sizeof(CDataType) * c_m_n_device_result.mDesc.GetElementSpaceSize()); DeviceMem d_m_n_device_buf(sizeof(DDataType) * d_m_n.mDesc.GetElementSpaceSize());
DeviceMem e_m_n_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpaceSize());
#ifdef BUILD_INT4_EXAMPLE #ifdef BUILD_INT4_EXAMPLE
const Tensor<KernelADataType> a_m_k_converted(a_m_k); const Tensor<KernelADataType> a_m_k_converted(a_m_k);
const Tensor<KernelBDataType> b_k_n_converted(b_k_n); const Tensor<KernelBDataType> b_k_n_converted(b_k_n);
const Tensor<KernelDDataType> d_m_n_converted(d_m_n);
a_m_k_device_buf.ToDevice(a_m_k_converted.mData.data()); a_m_k_device_buf.ToDevice(a_m_k_converted.mData.data());
b_k_n_device_buf.ToDevice(b_k_n_converted.mData.data()); b_k_n_device_buf.ToDevice(b_k_n_converted.mData.data());
d_m_n_device_buf.ToDevice(d_m_n_converted.mData.data());
#else #else
a_m_k_device_buf.ToDevice(a_m_k.mData.data()); a_m_k_device_buf.ToDevice(a_m_k.mData.data());
b_k_n_device_buf.ToDevice(b_k_n.mData.data()); b_k_n_device_buf.ToDevice(b_k_n.mData.data());
d_m_n_device_buf.ToDevice(d_m_n.mData.data());
#endif #endif
c_m_n_device_buf.SetZero(); e_m_n_device_buf.SetZero();
auto a_element_op = AElementOp{}; auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{}; auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{}; auto cde_element_op = CDEElementOp{};
// do GEMM // do GEMM
auto gemm = DeviceGemmInstance{}; auto gemm = DeviceOpInstance{};
auto invoker = gemm.MakeInvoker(); auto invoker = gemm.MakeInvoker();
auto argument = gemm.MakeArgument( auto argument = gemm.MakeArgument(
#ifdef BUILD_INT4_EXAMPLE #ifdef BUILD_INT4_EXAMPLE
...@@ -110,17 +132,20 @@ bool run_splitK_gemm_bias(const ProblemSize& problem_size, const ExecutionConfig ...@@ -110,17 +132,20 @@ bool run_splitK_gemm_bias(const ProblemSize& problem_size, const ExecutionConfig
#else #else
static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()), static_cast<ADataType*>(a_m_k_device_buf.GetDeviceBuffer()),
static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()), static_cast<BDataType*>(b_k_n_device_buf.GetDeviceBuffer()),
std::array<const void*, 1>{static_cast<DDataType*>(d_m_n_device_buf.GetDeviceBuffer())},
#endif #endif
static_cast<CDataType*>(c_m_n_device_buf.GetDeviceBuffer()), static_cast<EDataType*>(e_m_n_device_buf.GetDeviceBuffer()),
M, a_ms_ks_lengths,
N, a_ms_ks_strides,
K, b_ns_ks_lengths,
StrideA, b_ns_ks_strides,
StrideB, std::array<std::vector<ck::index_t>, 1>{d_ms_ns_lengths},
StrideC, std::array<std::vector<ck::index_t>, 1>{d_ms_ns_strides},
e_ms_ns_lengths,
e_ms_ns_strides,
a_element_op, a_element_op,
b_element_op, b_element_op,
c_element_op, cde_element_op,
KBatch); KBatch);
if(!gemm.IsSupportedArgument(argument)) if(!gemm.IsSupportedArgument(argument))
...@@ -135,22 +160,23 @@ bool run_splitK_gemm_bias(const ProblemSize& problem_size, const ExecutionConfig ...@@ -135,22 +160,23 @@ bool run_splitK_gemm_bias(const ProblemSize& problem_size, const ExecutionConfig
if(config.do_verification) if(config.do_verification)
{ {
c_m_n_device_buf.FromDevice(c_m_n_device_result.mData.data()); e_m_n_device_buf.FromDevice(e_m_n_device_result.mData.data());
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType, using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemmBias2D<ADataType,
BDataType, BDataType,
CDataType, DDataType
EDataType,
AccDataType, AccDataType,
AElementOp, AElementOp,
BElementOp, BElementOp,
CElementOp>; CDEElementOp>;
auto ref_gemm = ReferenceGemmInstance{}; auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker(); auto ref_invoker = ref_gemm.MakeInvoker();
Tensor<CDataType> c_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{})); Tensor<CDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideC, CLayout{}));
auto ref_argument = ref_gemm.MakeArgument( auto ref_argument = ref_gemm.MakeArgument(
a_m_k, b_k_n, c_m_n_host_result, a_element_op, b_element_op, c_element_op); a_m_k, b_k_n, e_m_n_host_result, d_m_n, a_element_op, b_element_op, c_element_op);
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
...@@ -164,7 +190,7 @@ bool run_splitK_gemm_bias(const ProblemSize& problem_size, const ExecutionConfig ...@@ -164,7 +190,7 @@ bool run_splitK_gemm_bias(const ProblemSize& problem_size, const ExecutionConfig
} }
else else
{ {
pass &= ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData); pass &= ck::utils::check_err(e_m_n_device_result.mData, e_m_n_host_result.mData);
} }
} }
......
example/42_splitK_gemm_bias/splitK_gemm_bias_xdl_fp16.cpp
View file @ f0d63f25
...@@ -27,32 +27,43 @@ using F32 = float; ...@@ -27,32 +27,43 @@ using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor; using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor; using Col = ck::tensor_layout::gemm::ColumnMajor;
using ALayout = Row;
using BLayout = Col;
using DsLayout = Row;
using ELayout = Row;
using PassThrough = ck::tensor_operation::element_wise::PassThrough; using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using Add = ck::tensor_operation::element_wise::Add;
using ADataType = F16; using ADataType = F16;
using BDataType = F16; using BDataType = F16;
using AccDataType = F32; using AccDataType = F32;
using CDataType = F16; using CShuffleDataType = F16;
using DDataType = F16;
using ALayout = Row; using DsDataType = ck::Tuple<F16>;
using BLayout = Col; using EDataType = F16;
using CLayout = Row;
using AElementOp = PassThrough; using AElementOp = PassThrough;
using BElementOp = PassThrough; using BElementOp = PassThrough;
using CElementOp = PassThrough; using CDEElementOp = Add;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default; static constexpr ck::index_t NumDimM = 1;
static constexpr ck::index_t NumDimN = 1;
static constexpr ck::index_t NumDimK = 1;
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdlSplitKCShuffle static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
// clang-format off
//######| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| KPer| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| // clang-format off
//######| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector| using DeviceOpInstanceKKN = ck::tensor_operation::device::
//######| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl| //#####################################| NumDimM| NumDimN| NumDimK| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | //#####################################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
< ADataType, BDataType, CDataType, AccDataType, ALayout, BLayout, CLayout, AElementOp, BElementOp, CElementOp, GemmDefault, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<1, 4, 64, 1>, S<0, 2, 1, 3>, S<0, 2, 1, 3>, 3, 8, 8, true, S<1, 4, 64, 1>, S<0, 1, 3, 2>, S<0, 1, 3, 2>, 3, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8>; //#####################################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//#####################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceSplitKContractionMultipleD_Xdl_CShuffle< NumDimM, NumDimN, NumDimK, F32, F32, F32, F32, DsDataType, F32, AElementOp, BElementOp, CDEElementOp, GemmSpec, 1, 256, 256, 128, 16, 4, 4, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 4, 4, 1, 1, 1, S<1, 16, 1, 16>, 4>;
// clang-format on // clang-format on
using DeviceOpInstance = DeviceOpInstanceKKN;
#include "run_splitK_gemm_bias_example.inc" #include "run_splitK_gemm_bias_example.inc"
int main(int argc, char* argv[]) { return !run_splitK_gemm_bias_example(argc, argv); } int main(int argc, char* argv[]) { return !run_splitK_gemm_bias_example(argc, argv); }
include/ck/tensor_operation/gpu/device/device_batched_contraction_multiple_d.hpp
View file @ f0d63f25
...@@ -59,6 +59,43 @@ struct DeviceBatchedContractionMultipleD : public BaseOperator ...@@ -59,6 +59,43 @@ struct DeviceBatchedContractionMultipleD : public BaseOperator
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0; virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
}; };
template <index_t NumDimG,
index_t NumDimM,
index_t NumDimN,
index_t NumDimK,
typename ADataType,
typename BDataType,
typename DsDataType,
typename EDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation>
struct DeviceSplitKContractionMultipleD : public BaseOperator
{
static constexpr index_t NumDTensor = DsDataType::Size();
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_a,
const void* p_b,
std::array<const void*, NumDTensor> p_ds,
void* p_e,
const std::vector<index_t>& a_gs_ms_ns_lengths,
const std::vector<index_t>& a_gs_ms_ks_strides,
const std::vector<index_t>& b_gs_ns_ks_lengths,
const std::vector<index_t>& b_gs_ns_ks_strides,
const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_lengths,
const std::array<std::vector<index_t>, NumDTensor>& ds_gs_ms_ns_strides,
const std::vector<index_t>& e_gs_ms_ns_lengths,
const std::vector<index_t>& e_gs_ms_ns_strides,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op,
const index_t k_batch) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
} // namespace device } // namespace device
} // namespace tensor_operation } // namespace tensor_operation
} // namespace ck } // namespace ck
include/ck/tensor_operation/gpu/device/device_contraction_splitK_multiple_d_xdl_cshuffle.hpp
View file @ f0d63f25
...@@ -195,8 +195,8 @@ template <index_t NumDimG, ...@@ -195,8 +195,8 @@ template <index_t NumDimG,
typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock, typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CDEBlockTransferScalarPerVector_NPerBlock, index_t CDEBlockTransferScalarPerVector_NPerBlock,
LoopScheduler LoopSched = make_default_loop_scheduler()> LoopScheduler LoopSched = make_default_loop_scheduler()>
struct DeviceBatchedContractionMultipleD_Xdl_CShuffle struct DeviceSplitKContractionMultipleD_Xdl_CShuffle
: public DeviceBatchedContractionMultipleD<NumDimG, : public DeviceSplitKContractionMultipleD<NumDimG,
NumDimM, NumDimM,
NumDimN, NumDimN,
NumDimK, NumDimK,
...@@ -208,7 +208,7 @@ struct DeviceBatchedContractionMultipleD_Xdl_CShuffle ...@@ -208,7 +208,7 @@ struct DeviceBatchedContractionMultipleD_Xdl_CShuffle
BElementwiseOperation, BElementwiseOperation,
CDEElementwiseOperation> CDEElementwiseOperation>
{ {
using DeviceOp = DeviceBatchedContractionMultipleD_Xdl_CShuffle; using DeviceOp = DeviceSplitKContractionMultipleD_Xdl_CShuffle;
static constexpr index_t NumDTensor = DsDataType::Size(); static constexpr index_t NumDTensor = DsDataType::Size();
...@@ -658,7 +658,8 @@ struct DeviceBatchedContractionMultipleD_Xdl_CShuffle ...@@ -658,7 +658,8 @@ struct DeviceBatchedContractionMultipleD_Xdl_CShuffle
const std::vector<index_t>& e_gs_ms_ns_strides, const std::vector<index_t>& e_gs_ms_ns_strides,
AElementwiseOperation a_element_op, AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op, BElementwiseOperation b_element_op,
CDEElementwiseOperation cde_element_op) CDEElementwiseOperation cde_element_op,
const index_t KBatch)
: p_a_grid_{static_cast<const ADataType*>(p_a_grid)}, : p_a_grid_{static_cast<const ADataType*>(p_a_grid)},
p_b_grid_{static_cast<const BDataType*>(p_b_grid)}, p_b_grid_{static_cast<const BDataType*>(p_b_grid)},
p_ds_grid_{}, p_ds_grid_{},
...@@ -680,7 +681,7 @@ struct DeviceBatchedContractionMultipleD_Xdl_CShuffle ...@@ -680,7 +681,7 @@ struct DeviceBatchedContractionMultipleD_Xdl_CShuffle
GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k_)}, GridwiseGemm::MakeDefaultBGridDescriptor_BK0_N_BK1(b_grid_desc_n_k_)},
ds_grid_desc_mblock_mperblock_nblock_nperblock_{}, ds_grid_desc_mblock_mperblock_nblock_nperblock_{},
e_grid_desc_mblock_mperblock_nblock_nperblock_{}, e_grid_desc_mblock_mperblock_nblock_nperblock_{},
block_2_etile_map_{GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_)}, block_2_etile_map_{GridwiseGemm::MakeDefaultBlock2ETileMap(e_grid_desc_m_n_, KBatch)},
a_element_op_{a_element_op}, a_element_op_{a_element_op},
b_element_op_{b_element_op}, b_element_op_{b_element_op},
cde_element_op_{cde_element_op}, cde_element_op_{cde_element_op},
...@@ -1056,7 +1057,7 @@ struct DeviceBatchedContractionMultipleD_Xdl_CShuffle ...@@ -1056,7 +1057,7 @@ struct DeviceBatchedContractionMultipleD_Xdl_CShuffle
auto str = std::stringstream(); auto str = std::stringstream();
// clang-format off // clang-format off
str << "DeviceBatchedContractionMultipleD_Xdl_CShuffle" str << "DeviceSplitKContractionMultipleD_Xdl_CShuffle"
<< "<" << "<"
<< NumDimG << ", " << NumDimG << ", "
<< NumDimM << ", " << NumDimM << ", "
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_splitk_multiple_d_xdl_cshuffle.hpp 0 → 100644
View file @ f0d63f25
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/common_header.hpp"
#include "ck/tensor_description/multi_index_transform_helper.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_gemm_pipeline_v1.hpp"
#include "ck/tensor_operation/gpu/block/blockwise_gemm_xdlops.hpp"
#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v4r1.hpp"
#include "ck/tensor_operation/gpu/block/thread_group_tensor_slice_transfer_v7.hpp"
#include "ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
namespace ck {
// GEMM:
// input : A[M, K]
// input : B[N, K]
// input : D0[M, N], D1[M, N], ...
// output : E[M, N]
// C = a_op(A) * b_op(B)
// E = cde_op(C, D0, D1, ...)
// Assume:
// D0, D1, ... and E have the same layout
template <typename ABDataType, // FIXME: don't assume A/B have same datatype
typename AccDataType,
typename CShuffleDataType,
typename DsDataType,
typename EDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CDEElementwiseOperation,
InMemoryDataOperationEnum EGlobalMemoryDataOperation,
typename AGridDesc_M_K,
typename BGridDesc_N_K,
typename DsGridDesc_M_N,
typename EGridDesc_M_N,
index_t NumGemmKPrefetchStage,
index_t BlockSize,
index_t MPerBlock,
index_t NPerBlock,
index_t KPerBlock,
index_t AK1Value,
index_t BK1Value,
index_t MPerXdl,
index_t NPerXdl,
index_t MXdlPerWave,
index_t NXdlPerWave,
typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
index_t ABlockTransferSrcVectorDim,
index_t ABlockTransferSrcScalarPerVector,
index_t ABlockTransferDstScalarPerVector_AK1,
bool AThreadTransferSrcResetCoordinateAfterRun,
index_t ABlockLdsExtraM,
typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
typename BBlockTransferThreadClusterArrangeOrder,
typename BBlockTransferSrcAccessOrder,
index_t BBlockTransferSrcVectorDim,
index_t BBlockTransferSrcScalarPerVector,
index_t BBlockTransferDstScalarPerVector_BK1,
bool BThreadTransferSrcResetCoordinateAfterRun,
index_t BBlockLdsExtraN,
index_t CShuffleMXdlPerWavePerShuffle,
index_t CShuffleNXdlPerWavePerShuffle,
typename CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
index_t CDEShuffleBlockTransferScalarPerVector_NPerBlock,
LoopScheduler LoopSched>
struct GridwiseGemmSplitKMultipleD_xdl_cshuffle
{
static constexpr index_t NumDTensor = DsDataType::Size();
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
static constexpr auto I2 = Number<2>{};
static constexpr auto I3 = Number<3>{};
static constexpr auto I4 = Number<4>{};
static constexpr auto I5 = Number<5>{};
static constexpr auto I6 = Number<6>{};
static constexpr auto I7 = Number<7>{};
// K1 should be Number<...>
static constexpr auto AK1 = Number<AK1Value>{};
static constexpr auto BK1 = Number<BK1Value>{};
static constexpr auto AK0PerBlock = Number<KPerBlock / AK1Value>{};
static constexpr auto BK0PerBlock = Number<KPerBlock / BK1Value>{};
using ThisThreadBlock = ThisThreadBlock<BlockSize>;
using GridwiseGemmPipe = GridwiseGemmPipeline_v1<NumGemmKPrefetchStage>;
__host__ __device__ static constexpr auto GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1()
{
// A matrix in LDS memory, dst of blockwise copy
return make_naive_tensor_descriptor(
make_tuple(AK0PerBlock, Number<MPerBlock>{}, AK1),
make_tuple(Number<MPerBlock + ABlockLdsExtraM>{} * AK1, AK1, I1));
}
__host__ __device__ static constexpr auto GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1()
{
// B matrix in LDS memory, dst of blockwise copy
return make_naive_tensor_descriptor(
make_tuple(BK0PerBlock, Number<NPerBlock>{}, BK1),
make_tuple(Number<NPerBlock + BBlockLdsExtraN>{} * BK1, BK1, I1));
}
__host__ __device__ static constexpr auto
GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock()
{
constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
make_naive_tensor_descriptor_packed(
make_tuple(I1,
Number<CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl>{},
I1,
Number<CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>{}));
return c_shuffle_block_desc_mblock_mperblock_nblock_nperblock;
}
// ck::Tuple<const D0DataType*, const D1DataType*, ...>
static constexpr auto MakeDsGridPointer()
{
return generate_tuple(
[&](auto i) {
using DDataType = remove_cvref_t<tuple_element_t<i.value, DsDataType>>;
return static_cast<const DDataType*>(nullptr);
},
Number<NumDTensor>{});
}
__host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
{
// LDS allocation for A and B: be careful of alignment
constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
// lds max alignment
constexpr auto max_lds_align = math::lcm(AK1, BK1);
constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
constexpr auto b_block_space_size_aligned = math::integer_least_multiple(
b_block_desc_bk0_n_bk1.GetElementSpaceSize(), max_lds_align);
// LDS allocation for C shuffle in LDS
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
constexpr auto c_block_size =
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize();
return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
sizeof(ABDataType),
c_block_size * sizeof(CShuffleDataType));
}
// A desc for source in blockwise copy
__host__ __device__ static constexpr auto
MakeDefaultAGridDescriptor_AK0_M_AK1(const AGridDesc_M_K& a_grid_desc_m_k)
{
const auto M = a_grid_desc_m_k.GetLength(I0);
const auto K = a_grid_desc_m_k.GetLength(I1);
const auto AK0 = K / AK1;
return transform_tensor_descriptor(a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(AK0, AK1)),
make_pass_through_transform(M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
// B desc for source in blockwise copy
__host__ __device__ static constexpr auto
MakeDefaultBGridDescriptor_BK0_N_BK1(const BGridDesc_N_K& b_grid_desc_n_k)
{
const auto N = b_grid_desc_n_k.GetLength(I0);
const auto K = b_grid_desc_n_k.GetLength(I1);
const auto BK0 = K / BK1;
return transform_tensor_descriptor(b_grid_desc_n_k,
make_tuple(make_unmerge_transform(make_tuple(BK0, BK1)),
make_pass_through_transform(N)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
// E desc for destination in blockwise copy
template <typename EGridDescriptor_M_N>
__host__ __device__ static constexpr auto MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
const EGridDescriptor_M_N& e_grid_desc_m_n)
{
const auto M = e_grid_desc_m_n.GetLength(I0);
const auto N = e_grid_desc_m_n.GetLength(I1);
const auto MBlock = M / MPerBlock;
const auto NBlock = N / NPerBlock;
const auto e_grid_desc_mblock_mperblock_nblock_nperblock = transform_tensor_descriptor(
e_grid_desc_m_n,
make_tuple(make_unmerge_transform(make_tuple(MBlock, Number<MPerBlock>{})),
make_unmerge_transform(make_tuple(NBlock, Number<NPerBlock>{}))),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1>{}, Sequence<2, 3>{}));
return e_grid_desc_mblock_mperblock_nblock_nperblock;
}
// Ds desc for source in blockwise copy
template <typename DsGridDescriptor_M_N>
__host__ __device__ static constexpr auto
MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(
const DsGridDescriptor_M_N& ds_grid_desc_m_n)
{
return generate_tuple(
[&](auto i) {
return MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(ds_grid_desc_m_n[i]);
},
Number<NumDTensor>{});
}
// return block_id to E matrix tile idx (m0, n0) mapping
// __host__ __device__ static constexpr auto
// MakeDefaultBlock2ETileMap(const EGridDesc_M_N& e_grid_desc_m_n)
// {
// return BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, EGridDesc_M_N>(
// e_grid_desc_m_n);
// }
// return block_id to C matrix tile idx (m0, n0) mapping
__host__ __device__ static constexpr auto MakeDefaultBlock2ETileMap(
const EGridDesc_M_N& c_m_n_grid_desc, index_t /* M01 */, index_t /* N01 */, index_t KBatch = 1)
{
return BlockToCTileMap_KSplit_M00_N0_M01Adapt<MPerBlock, NPerBlock, EGridDesc_M_N>(
c_m_n_grid_desc, 8, KBatch);
}
// block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
template <typename Block2ETileMap>
__host__ __device__ static constexpr bool CheckValidity(const AGridDesc_M_K& a_grid_desc_m_k,
const BGridDesc_N_K& b_grid_desc_n_k,
const DsGridDesc_M_N& ds_grid_desc_m_n,
const EGridDesc_M_N& e_grid_desc_m_n,
const Block2ETileMap& block_2_etile_map)
{
static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
(NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
"Invalid tuning param!");
const auto M = a_grid_desc_m_k.GetLength(I0);
const auto N = b_grid_desc_n_k.GetLength(I0);
const auto K = a_grid_desc_m_k.GetLength(I1);
// check consistency of desc
if(!(M == e_grid_desc_m_n.GetLength(I0) && N == e_grid_desc_m_n.GetLength(I1)))
{
return false;
}
bool valid = true;
static_for<0, NumDTensor, 1>{}([&](auto i) {
valid = valid && (M == ds_grid_desc_m_n[i].GetLength(I0) &&
N == ds_grid_desc_m_n[i].GetLength(I1));
});
if(!valid)
{
return false;
}
// check tile size
if(!(M % MPerBlock == 0 && N % NPerBlock == 0 && K % KPerBlock == 0))
{
return false;
}
// check gridwise gemm pipeline
const auto num_k_loop = K / KPerBlock;
if(!GridwiseGemmPipe::IsSupported(num_k_loop))
{
return false;
}
// check block-to-E-tile
if(!block_2_etile_map.CheckValidity(e_grid_desc_m_n))
{
return false;
}
// TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
// check tensor size: cannot be larger than 2GB each
constexpr long_index_t TwoGB = (long_index_t{1} << 31);
if(!(a_grid_desc_m_k.GetElementSpaceSize() * sizeof(ABDataType) <= TwoGB &&
b_grid_desc_n_k.GetElementSpaceSize() * sizeof(ABDataType) <= TwoGB &&
e_grid_desc_m_n.GetElementSpaceSize() * sizeof(EDataType) <= TwoGB))
{
return false;
}
return true;
}
__host__ __device__ static constexpr bool CalculateHasMainKBlockLoop(index_t K)
{
const index_t num_loop = K / KPerBlock;
return GridwiseGemmPipe::CalculateHasMainLoop(num_loop);
}
using DefaultAGridDesc_AK0_M_AK1 =
remove_cvref_t<decltype(MakeDefaultAGridDescriptor_AK0_M_AK1(AGridDesc_M_K{}))>;
using DefaultBGridDesc_BK0_N_BK1 =
remove_cvref_t<decltype(MakeDefaultBGridDescriptor_BK0_N_BK1(BGridDesc_N_K{}))>;
using EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
MakeEGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(EGridDesc_M_N{}))>;
using DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock = remove_cvref_t<decltype(
MakeDsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock(DsGridDesc_M_N{}))>;
using DefaultBlock2ETileMap =
remove_cvref_t<decltype(MakeDefaultBlock2ETileMap(EGridDesc_M_N{}, KBatch))>;
using DsGridPointer = decltype(MakeDsGridPointer());
template <bool HasMainKBlockLoop,
typename AGridDesc_AK0_M_AK1,
typename BGridDesc_BK0_N_BK1,
typename Block2ETileMap>
__device__ static void Run(const ABDataType* __restrict__ p_a_grid,
const ABDataType* __restrict__ p_b_grid,
DsGridPointer p_ds_grid,
EDataType* __restrict__ p_e_grid,
void* __restrict__ p_shared,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const CDEElementwiseOperation& cde_element_op,
const AGridDesc_AK0_M_AK1& a_grid_desc_ak0_m_ak1,
const BGridDesc_BK0_N_BK1& b_grid_desc_bk0_n_bk1,
const DsGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
ds_grid_desc_mblock_mperblock_nblock_nperblock,
const EGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
e_grid_desc_mblock_mperblock_nblock_nperblock,
const Block2ETileMap& block_2_etile_map)
{
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
const auto ds_grid_buf = generate_tuple(
[&](auto i) {
return make_dynamic_buffer<AddressSpaceEnum::Global>(
p_ds_grid[i],
ds_grid_desc_mblock_mperblock_nblock_nperblock[i].GetElementSpaceSize());
},
Number<NumDTensor>{});
auto e_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_e_grid, e_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
// divide block work by [M, N]
const auto block_work_idx =
block_2_etile_map.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
if(!block_2_etile_map.ValidCTileIndex(
block_work_idx,
make_tuple(e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I0),
e_grid_desc_mblock_mperblock_nblock_nperblock.GetLength(I2))))
{
return;
}
// HACK: this force m/n_block_data_idx_on_grid into SGPR
const index_t 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);
// lds max alignment
constexpr auto max_lds_align = math::lcm(AK1, BK1);
// A matrix in LDS memory, dst of blockwise copy
constexpr auto a_block_desc_ak0_m_ak1 = GetABlockDescriptor_AK0PerBlock_MPerBlock_AK1();
// B matrix in LDS memory, dst of blockwise copy
constexpr auto b_block_desc_bk0_n_bk1 = GetBBlockDescriptor_BK0PerBlock_NPerBlock_BK1();
// A matrix blockwise copy
auto a_blockwise_copy =
ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
AElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum::Set,
Sequence<AK0PerBlock, MPerBlock, AK1>,
ABlockTransferThreadClusterLengths_AK0_M_AK1,
ABlockTransferThreadClusterArrangeOrder,
ABDataType,
ABDataType,
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,
AThreadTransferSrcResetCoordinateAfterRun,
true,
NumGemmKPrefetchStage>(
a_grid_desc_ak0_m_ak1,
make_multi_index(0, m_block_data_idx_on_grid, 0),
a_element_op,
a_block_desc_ak0_m_ak1,
make_multi_index(0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
// B matrix blockwise copy
auto b_blockwise_copy =
ThreadGroupTensorSliceTransfer_v4r1<ThisThreadBlock,
BElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum::Set,
Sequence<BK0PerBlock, NPerBlock, BK1>,
BBlockTransferThreadClusterLengths_BK0_N_BK1,
BBlockTransferThreadClusterArrangeOrder,
ABDataType,
ABDataType,
decltype(b_grid_desc_bk0_n_bk1),
decltype(b_block_desc_bk0_n_bk1),
BBlockTransferSrcAccessOrder,
Sequence<1, 0, 2>,
BBlockTransferSrcVectorDim,
2,
BBlockTransferSrcScalarPerVector,
BBlockTransferDstScalarPerVector_BK1,
1,
1,
BThreadTransferSrcResetCoordinateAfterRun,
true,
NumGemmKPrefetchStage>(
b_grid_desc_bk0_n_bk1,
make_multi_index(0, n_block_data_idx_on_grid, 0),
b_element_op,
b_block_desc_bk0_n_bk1,
make_multi_index(0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
// GEMM definition
// c_mtx += transpose(a_mtx) * b_mtx
// a_mtx[K0PerBlock, MPerBlock] is in LDS
// b_mtx[K0PerBlock, NPerBlock] is in LDS
// c_mtx[MPerBlock, NPerBlock] is distributed among threads, and saved in
// register
// sanity check
constexpr index_t KPack =
math::max(math::lcm(AK1, BK1),
MfmaSelector<ABDataType, MPerXdl, NPerXdl>::selected_mfma.k_per_blk);
auto blockwise_gemm = BlockwiseGemmXdlops_k0mk1_k0nk1_m0n0m1n1m2m3m4n2_Selector<
BlockSize,
ABDataType,
AccDataType,
decltype(a_block_desc_ak0_m_ak1),
decltype(b_block_desc_bk0_n_bk1),
MPerXdl,
NPerXdl,
MXdlPerWave,
NXdlPerWave,
KPack,
LoopSched>();
auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
// LDS allocation for A and B: be careful of alignment
constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<ABDataType*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<ABDataType*>(p_shared) + a_block_space_size_aligned,
b_block_desc_bk0_n_bk1.GetElementSpaceSize());
constexpr auto a_block_slice_copy_step = make_multi_index(KPerBlock / AK1, 0, 0);
constexpr auto b_block_slice_copy_step = make_multi_index(KPerBlock / BK1, 0, 0);
// gridwise GEMM pipeline
const auto gridwise_gemm_pipeline =
GridwiseGemmPipeline_v1_Selector<NumGemmKPrefetchStage, LoopSched>();
const index_t num_k_block_main_loop = __builtin_amdgcn_readfirstlane(
(a_grid_desc_ak0_m_ak1.GetLength(I0) * a_grid_desc_ak0_m_ak1.GetLength(I2)) /
KPerBlock);
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,
c_thread_buf,
num_k_block_main_loop);
// shuffle C and write out
{
static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
"wrong!");
constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
// TODO: hacky, fix it!
constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
blockwise_gemm.GetCThreadDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
// TODO: hacky, fix it!
// c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp is only used to get lengths
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp =
blockwise_gemm.GetCBlockDescriptor_M0_N0_M1_N1_M2_M3_M4_N2();
constexpr auto M0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I0);
constexpr auto N0 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I1);
constexpr auto M1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I2);
constexpr auto N1 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I3);
constexpr auto M2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I4);
constexpr auto M3 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I5);
constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
constexpr auto c_shuffle_block_desc_mblock_mperblock_nblock_nperblock =
GetCShuffleBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<CShuffleDataType*>(p_shared),
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
c_shuffle_block_desc_mblock_mperblock_nblock_nperblock,
make_tuple(
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleMXdlPerWavePerShuffle>{}, // M0 (MXdlPerWave) per shuffle
M1, // M1 = MWave
M2, // M2 * M3 * M4 = MPerXdl
M3,
M4)),
make_freeze_transform(I0),
make_unmerge_transform(make_tuple(
Number<CShuffleNXdlPerWavePerShuffle>{}, // N0 (NXdlPerWave) per shuffle
N1, // N1 = NWave
N2))), // N2 = NPerXdl
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(
Sequence<>{}, Sequence<0, 2, 4, 5, 6>{}, Sequence<>{}, Sequence<1, 3, 7>{}));
// calculate origin of thread output tensor on global memory
// blockwise GEMM c matrix starting index
const auto c_thread_mtx_on_block =
blockwise_gemm.CalculateCThreadOriginDataIndex(I0, I0, I0, I0);
const index_t m_thread_data_on_block = c_thread_mtx_on_block[I0];
const index_t n_thread_data_on_block = c_thread_mtx_on_block[I1];
const auto m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
make_tuple(Sequence<0, 1, 2, 3, 4>{}),
make_tuple(Sequence<0>{}));
const auto m_thread_data_on_block_idx =
m_thread_data_on_block_to_m0_m1_m2_m3_m4_adaptor.CalculateBottomIndex(
make_multi_index(m_thread_data_on_block));
const auto n_thread_data_on_block_to_n0_n1_n2_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(N0, N1, N2))),
make_tuple(Sequence<0, 1, 2>{}),
make_tuple(Sequence<0>{}));
const auto n_thread_data_on_block_idx =
n_thread_data_on_block_to_n0_n1_n2_adaptor.CalculateBottomIndex(
make_multi_index(n_thread_data_on_block));
// shuffle: threadwise copy C from VGPR to LDS
auto c_thread_copy_vgpr_to_lds =
ThreadwiseTensorSliceTransfer_v1r3<AccDataType,
CShuffleDataType,
decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
ck::tensor_operation::element_wise::PassThrough,
Sequence<CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
I1,
I1,
M2,
I1,
M4,
I1>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
7,
1,
InMemoryDataOperationEnum::Set,
1,
true>{
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
make_multi_index(0,
0,
m_thread_data_on_block_idx[I1],
n_thread_data_on_block_idx[I1],
m_thread_data_on_block_idx[I2],
m_thread_data_on_block_idx[I3],
m_thread_data_on_block_idx[I4],
n_thread_data_on_block_idx[I2]),
ck::tensor_operation::element_wise::PassThrough{}};
// tuple of reference to C/Ds tensor descriptors
const auto c_ds_desc_refs = concat_tuple_of_reference(
tie(c_shuffle_block_desc_mblock_mperblock_nblock_nperblock),
generate_tie(
[&](auto i) -> const auto& // return type should be reference
{ return ds_grid_desc_mblock_mperblock_nblock_nperblock[i]; },
Number<NumDTensor>{}));
// tuple of reference to C/Ds tensor descriptors
const auto c_ds_buf_refs = concat_tuple_of_reference(
tie(c_shuffle_block_buf),
generate_tie(
[&](auto i) -> const auto& // return type should be reference
{ return ds_grid_buf[i]; },
Number<NumDTensor>{}));
// tuple of starting index of C/Ds blockwise copy
const auto idx_c_ds_block_begin = container_concat(
make_tuple(make_multi_index(0, 0, 0, 0)),
generate_tuple(
[&](auto) {
return make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0);
},
Number<NumDTensor>{}));
// only do bias at the 1st atomic add position
if constexpr(EGlobalMemoryDataOperation == InMemoryDataOperationEnum::AtomicAdd)
{
}
// blockwise copy C/D/E between LDS and global
auto cde_block_copy_lds_and_global = ThreadGroupTensorSliceTransfer_v7<
ThisThreadBlock,
decltype(container_concat(make_tuple(CShuffleDataType{}), DsDataType{})),
Tuple<EDataType>,
decltype(c_ds_desc_refs),
decltype(tie(e_grid_desc_mblock_mperblock_nblock_nperblock)),
CDEElementwiseOperation,
Sequence<static_cast<index_t>(EGlobalMemoryDataOperation)>, // FIXME: make Sequence
// support arbitray type
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>, // BlockSliceLengths,
CDEBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
Sequence<0, 1, 2, 3>, // typename ThreadClusterArrangeOrder,
Sequence<0, 1, 2, 3>, // typename DimAccessOrder,
3, // index_t VectorDim,
CDEShuffleBlockTransferScalarPerVector_NPerBlock,
sequence_merge_t<
Sequence<true>,
uniform_sequence_gen_t<NumDTensor,
false>>, // ThreadTransferSrcResetCoordinateAfterRunFlags
Sequence<false>> // ThreadTransferDstResetCoordinateAfterRunFlags
{c_ds_desc_refs,
idx_c_ds_block_begin,
tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
make_tuple(make_multi_index(block_work_idx[I0], 0, block_work_idx[I1], 0)),
cde_element_op};
// space filling curve for threadwise C in VGPR before shuffle
constexpr auto sfc_c_vgpr =
SpaceFillingCurve<Sequence<MXdlPerWave, NXdlPerWave, 1, 1, M2, 1, M4, 1>,
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
Sequence<CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
1,
1,
M2,
1,
M4,
1>>{};
// space filling curve for shuffled blockwise C/D/E
constexpr auto sfc_cde_block =
SpaceFillingCurve<Sequence<1, MPerBlock, 1, NPerBlock>,
Sequence<0, 2, 1, 3>,
Sequence<1,
CShuffleMXdlPerWavePerShuffle * MWave * MPerXdl,
1,
CShuffleNXdlPerWavePerShuffle * NWave * NPerXdl>>{};
constexpr index_t num_access = sfc_c_vgpr.GetNumOfAccess();
static_assert(num_access == sfc_cde_block.GetNumOfAccess(), "wrong!");
static_for<0, num_access, 1>{}([&](auto access_id) {
// make sure it's safe to write to LDS
block_sync_lds();
// each thread write its data from VGPR to LDS
c_thread_copy_vgpr_to_lds.Run(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
sfc_c_vgpr.GetIndexTupleOfNumber(access_id),
c_thread_buf,
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
c_shuffle_block_buf);
// make sure it's safe to read from LDS
block_sync_lds();
// each block copy its data from LDS to global
cde_block_copy_lds_and_global.Run(
c_ds_desc_refs,
c_ds_buf_refs,
tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
tie(e_grid_buf));
if constexpr(access_id < num_access - 1)
{
constexpr auto cde_lds_and_global_step =
sfc_cde_block.GetForwardStep(access_id);
// move on Ds
static_for<0, NumDTensor, 1>{}([&](auto i) {
cde_block_copy_lds_and_global.MoveSrcSliceWindow(
c_ds_desc_refs, i + I1, cde_lds_and_global_step);
});
// move on E
cde_block_copy_lds_and_global.MoveDstSliceWindow(
tie(e_grid_desc_mblock_mperblock_nblock_nperblock),
I0,
cde_lds_and_global_step);
}
});
}
}
};
} // namespace ck
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