Skip to content

GitLab

Commit 09ec28be authored by Chao Liu's avatar Chao Liu
Browse files

Merge remote-tracking branch 'origin/develop' into gelu

parents b9d3d277 85fc91c3
Show whitespace changes
Inline Side-by-side
Showing with 523 additions and 131 deletions
include/ck/tensor_operation/gpu/device/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp
View file @ 09ec28be
......@@ -417,6 +417,8 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
using BGridDesc_K0_N_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
using CGridDesc_M_N = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
using Block2CTileMap = BlockToCTileMap_M00_N0_M01<MPerBlock, NPerBlock, CGridDesc_M_N>;
// GridwiseGemm
using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1<
BlockSize,
......@@ -477,8 +479,6 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads,
ck::index_t M01,
ck::index_t N01,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op)
......@@ -490,8 +490,6 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
c_grid_desc_m_n_{},
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_{},
block_2_ctile_map_{},
M01_{M01},
N01_{N01},
in_element_op_{in_element_op},
wei_element_op_{wei_element_op},
out_element_op_{out_element_op},
......@@ -520,11 +518,10 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
a_grid_desc_k0_m_k1_ = descs[I0];
b_grid_desc_k0_n_k1_ = descs[I1];
block_2_ctile_map_ =
GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_, M01, N01);
c_grid_desc_m_n_ = descs[I2];
block_2_ctile_map_ = Block2CTileMap{c_grid_desc_m_n_};
if(GridwiseGemm::CheckValidity(a_grid_desc_k0_m_k1_,
b_grid_desc_k0_n_k1_,
c_grid_desc_m_n_,
......@@ -546,9 +543,7 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
typename GridwiseGemm::
CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_;
typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
index_t M01_;
index_t N01_;
Block2CTileMap block_2_ctile_map_;
InElementwiseOperation in_element_op_;
WeiElementwiseOperation wei_element_op_;
OutElementwiseOperation out_element_op_;
......@@ -661,7 +656,7 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
InElementwiseOperation,
WeiElementwiseOperation,
OutElementwiseOperation,
remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
Block2CTileMap,
true>;
ave_time = launch_and_time_kernel(
......@@ -695,7 +690,7 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
InElementwiseOperation,
WeiElementwiseOperation,
OutElementwiseOperation,
remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
Block2CTileMap,
false>;
ave_time = launch_and_time_kernel(
......@@ -814,8 +809,6 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
conv_filter_dilations,
input_left_pads,
input_right_pads,
1,
1,
in_element_op,
wei_element_op,
out_element_op};
......@@ -854,8 +847,6 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_W
conv_filter_dilations,
input_left_pads,
input_right_pads,
1,
1,
in_element_op,
wei_element_op,
out_element_op);
......
include/ck/tensor_operation/gpu/device/device_convnd_fwd_xdl_nhwc_kyxc_nhwk.hpp
View file @ 09ec28be
#ifndef DEVICE_CONVND_FWD_XDL_NHWC_KYXC_NHWK_HPP
#define DEVICE_CONVND_FWD_XDL_NHWC_KYXC_NHWK_HPP
#pragma once
#include <functional>
#include <iostream>
......@@ -8,6 +7,7 @@
#include <sstream>
#include "device.hpp"
#include "device_prop.hpp"
#include "device_base.hpp"
#include "device_conv_fwd.hpp"
#include "convolution_forward_specialization.hpp"
......@@ -607,6 +607,8 @@ struct DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
using BGridDesc_K0_N_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
using CGridDesc_M_N = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
using Block2CTileMap = BlockToCTileMap_M00_N0_M01<MPerBlock, NPerBlock, CGridDesc_M_N>;
// GridwiseGemm
using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<
BlockSize,
......@@ -664,8 +666,6 @@ struct DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads,
ck::index_t M01,
ck::index_t N01,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op)
......@@ -677,8 +677,6 @@ struct DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
c_grid_desc_m_n_{},
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_{},
block_2_ctile_map_{},
M01_{M01},
N01_{N01},
in_element_op_{in_element_op},
wei_element_op_{wei_element_op},
out_element_op_{out_element_op},
......@@ -705,8 +703,8 @@ struct DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
a_grid_desc_k0_m_k1_ = descs[I0];
b_grid_desc_k0_n_k1_ = descs[I1];
c_grid_desc_m_n_ = descs[I2];
block_2_ctile_map_ =
GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_, M01, N01);
block_2_ctile_map_ = Block2CTileMap{c_grid_desc_m_n_};
if(GridwiseGemm::CheckValidity(a_grid_desc_k0_m_k1_,
b_grid_desc_k0_n_k1_,
......@@ -727,9 +725,7 @@ struct DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
CGridDesc_M_N c_grid_desc_m_n_;
typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_;
typename GridwiseGemm::DefaultBlock2CTileMap block_2_ctile_map_;
index_t M01_;
index_t N01_;
Block2CTileMap block_2_ctile_map_;
InElementwiseOperation in_element_op_;
WeiElementwiseOperation wei_element_op_;
OutElementwiseOperation out_element_op_;
......@@ -793,7 +789,7 @@ struct DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
InElementwiseOperation,
WeiElementwiseOperation,
OutElementwiseOperation,
remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
Block2CTileMap,
true>;
ave_time = launch_and_time_kernel(stream_config,
......@@ -824,7 +820,7 @@ struct DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
InElementwiseOperation,
WeiElementwiseOperation,
OutElementwiseOperation,
remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
Block2CTileMap,
false>;
ave_time = launch_and_time_kernel(stream_config,
......@@ -862,6 +858,27 @@ struct DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
static bool IsSupportedArgument(const Argument& arg)
{
if(ck::get_device_name() == "gfx908")
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
is_same_v<AccDataType, int32_t>))
{
return false;
}
}
else if(ck::get_device_name() == "gfx90a")
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
is_same_v<AccDataType, int32_t> || is_same_v<AccDataType, double>))
{
return false;
}
}
else
{
return false;
}
// Input tensors can't be bigger than 2GB each.
constexpr ck::long_index_t GB2 = (ck::long_index_t{1} << 31);
......@@ -955,8 +972,6 @@ struct DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
conv_filter_dilations,
input_left_pads,
input_right_pads,
1,
1,
in_element_op,
wei_element_op,
out_element_op};
......@@ -995,8 +1010,6 @@ struct DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
conv_filter_dilations,
input_left_pads,
input_right_pads,
1,
1,
in_element_op,
wei_element_op,
out_element_op);
......@@ -1012,8 +1025,7 @@ struct DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
auto str = std::stringstream();
// clang-format off
str << "DeviceConv" << std::to_string(NumDimSpatial)
<< "DFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K"
str << "DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K"
<< "<"
<< BlockSize << ", "
<< MPerBlock << ", "
......@@ -1030,4 +1042,3 @@ struct DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
} // namespace device
} // namespace tensor_operation
} // namespace ck
#endif
include/ck/tensor_operation/gpu/device/device_gemm_dl.hpp
View file @ 09ec28be
......@@ -4,6 +4,7 @@
#include <sstream>
#include "device.hpp"
#include "device_prop.hpp"
#include "device_base.hpp"
#include "device_gemm.hpp"
#include "common_header.hpp"
......@@ -13,7 +14,6 @@
#include "gemm_specialization.hpp"
#include "element_wise_operation.hpp"
#include "gridwise_gemm_dl_v1r3.hpp"
#include "device_prop.hpp"
namespace ck {
namespace tensor_operation {
......
include/ck/tensor_operation/gpu/device/device_gemm_reduce.hpp
View file @ 09ec28be
......@@ -11,7 +11,7 @@ template <typename DPtrsGlobal,
typename BElementwiseOperation,
typename CElementwiseOperation,
typename DxsInElementwiseOperation,
typename DxsOutElementwiseOperation>
typename DxsAccElementwiseOperation>
struct DeviceGemmReduce : public BaseOperator
{
virtual std::unique_ptr<BaseArgument>
......@@ -29,7 +29,7 @@ struct DeviceGemmReduce : public BaseOperator
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
DxsInElementwiseOperation dxs_in_element_op,
DxsOutElementwiseOperation dxs_out_element_op,
DxsAccElementwiseOperation dxs_out_element_op,
ck::index_t BatchCount = 1) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
......@@ -40,13 +40,13 @@ template <typename DPtrsGlobal,
typename BElementwiseOperation,
typename CElementwiseOperation,
typename DxsInElementwiseOperation,
typename DxsOutElementwiseOperation>
typename DxsAccElementwiseOperation>
using DeviceGemmReducePtr = std::unique_ptr<DeviceGemmReduce<DPtrsGlobal,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
DxsInElementwiseOperation,
DxsOutElementwiseOperation>>;
DxsAccElementwiseOperation>>;
} // namespace device
} // namespace tensor_operation
......
include/ck/tensor_operation/gpu/device/device_gemm_reduce_xdl_cshuffle.hpp
View file @ 09ec28be
......@@ -32,7 +32,7 @@ template <typename ALayout,
typename CElementwiseOperation,
typename DxsReduceOperation,
typename DxsInElementwiseOperation,
typename DxsOutElementwiseOperation,
typename DxsAccElementwiseOperation,
typename DGlobalMemoryDataOperation,
GemmSpecialization GemmSpec,
index_t NumGemmKPrefetchStage,
......@@ -73,7 +73,7 @@ struct DeviceGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGlobal,
BElementwiseOperation,
CElementwiseOperation,
DxsInElementwiseOperation,
DxsOutElementwiseOperation>
DxsAccElementwiseOperation>
{
using DeviceOp = DeviceGemmReduce_Xdl_CShuffle;
......@@ -389,7 +389,7 @@ struct DeviceGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGlobal,
CElementwiseOperation,
DxsReduceOperation,
DxsInElementwiseOperation,
DxsOutElementwiseOperation,
DxsAccElementwiseOperation,
InMemoryDataOperationEnum::Set,
DGlobalMemoryDataOperation,
AGridDesc_AK0_M_AK1,
......@@ -449,7 +449,7 @@ struct DeviceGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGlobal,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
DxsInElementwiseOperation dxs_in_element_op,
DxsOutElementwiseOperation dxs_out_element_op)
DxsAccElementwiseOperation dxs_out_element_op)
: p_a_grid_{p_a_grid},
p_b_grid_{p_b_grid},
p_c_grid_{p_c_grid},
......@@ -498,7 +498,7 @@ struct DeviceGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGlobal,
BElementwiseOperation b_element_op_;
CElementwiseOperation c_element_op_;
DxsInElementwiseOperation dxs_in_element_op_;
DxsOutElementwiseOperation dxs_out_element_op_;
DxsAccElementwiseOperation dxs_out_element_op_;
};
// Invoker
......@@ -554,7 +554,7 @@ struct DeviceGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGlobal,
BElementwiseOperation,
CElementwiseOperation,
DxsInElementwiseOperation,
DxsOutElementwiseOperation,
DxsAccElementwiseOperation,
DeviceOp::AGridDesc_AK0_M_AK1,
DeviceOp::BGridDesc_BK0_N_BK1,
typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
......@@ -594,7 +594,7 @@ struct DeviceGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGlobal,
BElementwiseOperation,
CElementwiseOperation,
DxsInElementwiseOperation,
DxsOutElementwiseOperation,
DxsAccElementwiseOperation,
DeviceOp::AGridDesc_AK0_M_AK1,
DeviceOp::BGridDesc_BK0_N_BK1,
typename GridwiseGemm::CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
......@@ -669,7 +669,7 @@ struct DeviceGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGlobal,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
DxsInElementwiseOperation dxs_in_element_op,
DxsOutElementwiseOperation dxs_out_element_op)
DxsAccElementwiseOperation dxs_out_element_op)
{
return Argument{p_a,
p_b,
......@@ -705,7 +705,7 @@ struct DeviceGemmReduce_Xdl_CShuffle : public DeviceGemmReduce<DPtrsGlobal,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op,
DxsInElementwiseOperation dxs_in_element_op,
DxsOutElementwiseOperation dxs_out_element_op,
DxsAccElementwiseOperation dxs_out_element_op,
index_t /* KBatch */ = 1) override
{
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
......
include/ck/tensor_operation/gpu/device/device_gemm_xdl.hpp
View file @ 09ec28be
......@@ -3,6 +3,7 @@
#include <iostream>
#include <sstream>
#include "device.hpp"
#include "device_prop.hpp"
#include "device_base.hpp"
#include "device_gemm.hpp"
#include "common_header.hpp"
......@@ -11,7 +12,6 @@
#include "tensor_descriptor_helper.hpp"
#include "gridwise_gemm_xdlops_v2r3.hpp"
#include "gemm_specialization.hpp"
#include "device_prop.hpp"
namespace ck {
namespace tensor_operation {
......@@ -408,7 +408,23 @@ struct DeviceGemmXdl
static bool IsSupportedArgument(const Argument& arg)
{
if(!(ck::get_device_name() == "gfx908" || ck::get_device_name() == "gfx90a"))
if(ck::get_device_name() == "gfx908")
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
is_same_v<AccDataType, int32_t>))
{
return false;
}
}
else if(ck::get_device_name() == "gfx90a")
{
if constexpr(!(is_same_v<AccDataType, float> || is_same_v<AccDataType, float> ||
is_same_v<AccDataType, int32_t> || is_same_v<AccDataType, double>))
{
return false;
}
}
else
{
return false;
}
......
include/ck/tensor_operation/gpu/element/element_wise_operation.hpp
View file @ 09ec28be
......@@ -169,6 +169,24 @@ struct AddHardswishAdd
}
};
struct Normalize
{
Normalize(float epsilon = 1e-4) : epsilon_(epsilon) {}
__host__ __device__ constexpr void operator()(float& y,
const float& x,
const float& mean,
const float& mean_square,
const float& gamma,
const float& beta) const
{
float variance = mean_square - (mean * mean);
y = ((x - mean) / sqrtf(variance + epsilon_)) * gamma + beta;
}
float epsilon_;
};
// Unary operators are usually called element-wisely before/after the reduction is executed on the
// elements. They are needed for easy implementation of reduction types of AVG, NRM1, NRM2
......
include/ck/tensor_operation/gpu/grid/gridwise_5ary_Elementwise_1d.hpp 0 → 100644
View file @ 09ec28be
#pragma once
#include "cluster_descriptor.hpp"
#include "data_type.hpp"
#include "element_wise_operation.hpp"
#include "threadwise_tensor_slice_transfer.hpp"
namespace ck {
template <typename Gridwise5AryEltwise,
typename ADataType,
typename BDataType,
typename CDataType,
typename DDataType,
typename EDataType,
typename FDataType,
typename AGridDesc_M,
typename BGridDesc_M,
typename CGridDesc_M,
typename DGridDesc_M,
typename EGridDesc_M,
typename FGridDesc_M,
typename ElementwiseFunctor>
__global__ void kernel_5ary_elementwise_1d(const ADataType* __restrict__ p_a_global,
const BDataType* __restrict__ p_b_global,
const CDataType* __restrict__ p_c_global,
const DDataType* __restrict__ p_d_global,
const EDataType* __restrict__ p_e_global,
FDataType* __restrict__ p_f_global,
const AGridDesc_M a_grid_desc_m,
const BGridDesc_M b_grid_desc_m,
const CGridDesc_M c_grid_desc_m,
const DGridDesc_M d_grid_desc_m,
const EGridDesc_M e_grid_desc_m,
const FGridDesc_M f_grid_desc_m,
const ElementwiseFunctor functor)
{
Gridwise5AryEltwise::Run(p_a_global,
p_b_global,
p_c_global,
p_d_global,
p_e_global,
p_f_global,
a_grid_desc_m,
b_grid_desc_m,
c_grid_desc_m,
d_grid_desc_m,
e_grid_desc_m,
f_grid_desc_m,
functor);
}
// TODO - implement n-ary Elemenetwise_1D, tuple of inputs and tuple of outputs
template <typename ADataType,
typename BDataType,
typename CDataType,
typename DDataType,
typename EDataType,
typename FDataType,
typename ComputeDataType,
typename AGridDesc_M,
typename BGridDesc_M,
typename CGridDesc_M,
typename DGridDesc_M,
typename EGridDesc_M,
typename FGridDesc_M,
typename ElementwiseFunctor,
index_t MPerThread,
index_t AScalarPerVector,
index_t BScalarPerVector,
index_t CScalarPerVector,
index_t DScalarPerVector,
index_t EScalarPerVector,
index_t FScalarPerVector>
struct Gridwise5AryElementwise_1D
{
static constexpr auto I0 = Number<0>{};
static constexpr auto thread_desc_m =
make_naive_tensor_descriptor_packed(make_tuple(Number<MPerThread>{}));
using PassThrough = tensor_operation::element_wise::PassThrough;
static __device__ auto CalculateElementwiseIndex()
{
const index_t global_thread_id = get_thread_global_1d_id();
return make_multi_index(global_thread_id * MPerThread);
}
__device__ static void Run(const ADataType* __restrict__ p_a_global,
const BDataType* __restrict__ p_b_global,
const CDataType* __restrict__ p_c_global,
const DDataType* __restrict__ p_d_global,
const EDataType* __restrict__ p_e_global,
FDataType* __restrict__ p_f_global,
const AGridDesc_M a_grid_desc_m,
const BGridDesc_M b_grid_desc_m,
const CGridDesc_M c_grid_desc_m,
const DGridDesc_M d_grid_desc_m,
const EGridDesc_M e_grid_desc_m,
const FGridDesc_M f_grid_desc_m,
const ElementwiseFunctor functor)
{
const auto a_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_global, a_grid_desc_m.GetElementSpaceSize());
const auto b_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_b_global, b_grid_desc_m.GetElementSpaceSize());
const auto c_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_c_global, c_grid_desc_m.GetElementSpaceSize());
const auto d_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_d_global, d_grid_desc_m.GetElementSpaceSize());
const auto e_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_e_global, e_grid_desc_m.GetElementSpaceSize());
auto f_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_f_global, f_grid_desc_m.GetElementSpaceSize());
StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MPerThread, true> a_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MPerThread, true> b_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MPerThread, true> c_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MPerThread, true> d_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MPerThread, true> e_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MPerThread, true> f_thread_buf;
const auto thread_store_global_offset = CalculateElementwiseIndex();
auto a_global_load =
ThreadwiseTensorSliceTransfer_v2<ADataType,
ComputeDataType,
AGridDesc_M,
decltype(thread_desc_m),
Sequence<MPerThread>, // SliceLengths
Sequence<0>, // DimAccessOrder
0, // SrcVectorDim
AScalarPerVector, // ScalarPerVector
1, // SrcScalarStrideInVector
false>{a_grid_desc_m, thread_store_global_offset};
auto b_global_load =
ThreadwiseTensorSliceTransfer_v2<BDataType,
ComputeDataType,
BGridDesc_M,
decltype(thread_desc_m),
Sequence<MPerThread>, // SliceLengths
Sequence<0>, // DimAccessOrder
0, // SrcVectorDim
BScalarPerVector, // ScalarPerVector
1, // SrcScalarStrideInVector
false>{b_grid_desc_m, thread_store_global_offset};
auto c_global_load =
ThreadwiseTensorSliceTransfer_v2<CDataType,
ComputeDataType,
CGridDesc_M,
decltype(thread_desc_m),
Sequence<MPerThread>, // SliceLengths
Sequence<0>, // DimAccessOrder
0, // SrcVectorDim
CScalarPerVector, // ScalarPerVector
1, // SrcScalarStrideInVector
false>{c_grid_desc_m, thread_store_global_offset};
auto d_global_load =
ThreadwiseTensorSliceTransfer_v2<DDataType,
ComputeDataType,
DGridDesc_M,
decltype(thread_desc_m),
Sequence<MPerThread>, // SliceLengths
Sequence<0>, // DimAccessOrder
0, // SrcVectorDim
DScalarPerVector, // ScalarPerVector
1, // SrcScalarStrideInVector
false>{d_grid_desc_m, thread_store_global_offset};
auto e_global_load =
ThreadwiseTensorSliceTransfer_v2<EDataType,
ComputeDataType,
EGridDesc_M,
decltype(thread_desc_m),
Sequence<MPerThread>, // SliceLengths
Sequence<0>, // DimAccessOrder
0, // SrcVectorDim
EScalarPerVector, // ScalarPerVector
1, // SrcScalarStrideInVector
false>{e_grid_desc_m, thread_store_global_offset};
auto f_global_write =
ThreadwiseTensorSliceTransfer_v1r3<ComputeDataType,
FDataType,
decltype(thread_desc_m),
FGridDesc_M,
PassThrough,
Sequence<MPerThread>, // SliceLengths
Sequence<0>, // DimAccessOrder
0, // DstVectorDim
FScalarPerVector, // ScalarPerVector
InMemoryDataOperationEnum::Set,
1, // DstScalarStrideInVector
false>{
f_grid_desc_m, thread_store_global_offset, PassThrough{}};
const index_t blockSize = get_block_size();
const index_t blockPerGrid = get_grid_size();
const auto M = c_grid_desc_m.GetLength(I0);
const index_t loop_step = blockPerGrid * blockSize * MPerThread;
const auto loop_step_index = make_multi_index(loop_step);
index_t num_iter = M / (loop_step);
do
{
// read and process MPerThread elements
a_global_load.Run(
a_grid_desc_m, a_global_buf, thread_desc_m, make_tuple(I0), a_thread_buf);
b_global_load.Run(
b_grid_desc_m, b_global_buf, thread_desc_m, make_tuple(I0), b_thread_buf);
c_global_load.Run(
c_grid_desc_m, c_global_buf, thread_desc_m, make_tuple(I0), c_thread_buf);
d_global_load.Run(
d_grid_desc_m, d_global_buf, thread_desc_m, make_tuple(I0), d_thread_buf);
e_global_load.Run(
e_grid_desc_m, e_global_buf, thread_desc_m, make_tuple(I0), e_thread_buf);
static_for<0, MPerThread, 1>{}([&](auto m) {
constexpr auto offset = thread_desc_m.CalculateOffset(make_tuple(m));
functor(f_thread_buf(Number<offset>{}),
a_thread_buf(Number<offset>{}),
b_thread_buf(Number<offset>{}),
c_thread_buf(Number<offset>{}),
d_thread_buf(Number<offset>{}),
e_thread_buf(Number<offset>{}));
});
f_global_write.Run(thread_desc_m,
make_tuple(I0), // SrcSliceOriginIdx
f_thread_buf,
f_grid_desc_m,
f_global_buf);
a_global_load.MoveSrcSliceWindow(a_grid_desc_m, loop_step_index);
b_global_load.MoveSrcSliceWindow(b_grid_desc_m, loop_step_index);
c_global_load.MoveSrcSliceWindow(c_grid_desc_m, loop_step_index);
d_global_load.MoveSrcSliceWindow(d_grid_desc_m, loop_step_index);
e_global_load.MoveSrcSliceWindow(e_grid_desc_m, loop_step_index);
f_global_write.MoveDstSliceWindow(f_grid_desc_m, loop_step_index);
} while(--num_iter);
}
};
} // namespace ck
include/ck/tensor_operation/gpu/grid/gridwise_binary_elementwise_1d.hpp
View file @ 09ec28be
......@@ -11,138 +11,140 @@ template <typename GridwiseBinEltwise,
typename ADataType,
typename BDataType,
typename CDataType,
typename GridDesc_M0,
typename AGridDesc_M,
typename BGridDesc_M,
typename CGridDesc_M,
typename ElementwiseFunctor>
__global__ void kernel_binary_elementwise_1d(const ADataType* __restrict__ p_a_global,
const BDataType* __restrict__ p_b_global,
CDataType* __restrict__ p_c_global,
const GridDesc_M0 a_grid_desc_m0,
const GridDesc_M0 b_grid_desc_m0,
const GridDesc_M0 c_grid_desc_m0,
const AGridDesc_M a_grid_desc_m,
const BGridDesc_M b_grid_desc_m,
const CGridDesc_M c_grid_desc_m,
const ElementwiseFunctor functor)
{
GridwiseBinEltwise::Run(p_a_global,
p_b_global,
p_c_global,
a_grid_desc_m0,
b_grid_desc_m0,
c_grid_desc_m0,
functor);
GridwiseBinEltwise::Run(
p_a_global, p_b_global, p_c_global, a_grid_desc_m, b_grid_desc_m, c_grid_desc_m, functor);
}
template <typename ADataType,
typename BDataType,
typename CDataType,
typename ComputeDataType,
typename GridDesc_M0,
typename AGridDesc_M,
typename BGridDesc_M,
typename CGridDesc_M,
typename ElementwiseFunctor,
index_t ScalarPerVector>
index_t MPerThread,
index_t AScalarPerVector,
index_t BScalarPerVector,
index_t CScalarPerVector>
struct GridwiseBinaryElementwise_1D
{
static constexpr auto I0 = Number<0>{};
static constexpr auto thread_desc_m0 =
make_naive_tensor_descriptor_packed(make_tuple(Number<ScalarPerVector>{}));
static constexpr auto thread_desc_m =
make_naive_tensor_descriptor_packed(make_tuple(Number<MPerThread>{}));
using PassThrough = tensor_operation::element_wise::PassThrough;
static __device__ auto CalculateElementwiseIndex()
{
const index_t global_thread_id = get_thread_global_1d_id();
return make_multi_index(global_thread_id * ScalarPerVector);
return make_multi_index(global_thread_id * MPerThread);
}
__device__ static void Run(const ADataType* __restrict__ p_a_global,
const BDataType* __restrict__ p_b_global,
CDataType* __restrict__ p_c_global,
const GridDesc_M0 a_grid_desc_m0,
const GridDesc_M0 b_grid_desc_m0,
const GridDesc_M0 c_grid_desc_m0,
const AGridDesc_M a_grid_desc_m,
const BGridDesc_M b_grid_desc_m,
const CGridDesc_M c_grid_desc_m,
const ElementwiseFunctor functor)
{
const auto a_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_global, a_grid_desc_m0.GetElementSpaceSize());
p_a_global, a_grid_desc_m.GetElementSpaceSize());
const auto b_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_b_global, b_grid_desc_m0.GetElementSpaceSize());
p_b_global, b_grid_desc_m.GetElementSpaceSize());
auto c_global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_c_global, c_grid_desc_m0.GetElementSpaceSize());
p_c_global, c_grid_desc_m.GetElementSpaceSize());
StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, ScalarPerVector, true> a_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, ScalarPerVector, true> b_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, ScalarPerVector, true> c_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MPerThread, true> a_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MPerThread, true> b_thread_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, ComputeDataType, MPerThread, true> c_thread_buf;
const auto thread_store_global_offset = CalculateElementwiseIndex();
auto a_global_load =
ThreadwiseTensorSliceTransfer_v2<ADataType,
ComputeDataType,
GridDesc_M0,
decltype(thread_desc_m0),
Sequence<ScalarPerVector>, // SliceLengths
AGridDesc_M,
decltype(thread_desc_m),
Sequence<MPerThread>, // SliceLengths
Sequence<0>, // DimAccessOrder
0, // SrcVectorDim
ScalarPerVector,
AScalarPerVector, // ScalarPerVector
1, // SrcScalarStrideInVector
false>{a_grid_desc_m0, thread_store_global_offset};
false>{a_grid_desc_m, thread_store_global_offset};
auto b_global_load =
ThreadwiseTensorSliceTransfer_v2<BDataType,
ComputeDataType,
GridDesc_M0,
decltype(thread_desc_m0),
Sequence<ScalarPerVector>, // SliceLengths
BGridDesc_M,
decltype(thread_desc_m),
Sequence<MPerThread>, // SliceLengths
Sequence<0>, // DimAccessOrder
0, // SrcVectorDim
ScalarPerVector,
BScalarPerVector, // ScalarPerVector
1, // SrcScalarStrideInVector
false>{b_grid_desc_m0, thread_store_global_offset};
false>{b_grid_desc_m, thread_store_global_offset};
auto c_global_write =
ThreadwiseTensorSliceTransfer_v1r3<ComputeDataType,
CDataType,
decltype(thread_desc_m0),
GridDesc_M0,
decltype(thread_desc_m),
CGridDesc_M,
PassThrough,
Sequence<ScalarPerVector>, // SliceLengths
Sequence<MPerThread>, // SliceLengths
Sequence<0>, // DimAccessOrder
0, // DstVectorDim
ScalarPerVector,
CScalarPerVector, // ScalarPerVector
InMemoryDataOperationEnum::Set,
1, // DstScalarStrideInVector
false>{
c_grid_desc_m0, thread_store_global_offset, PassThrough{}};
c_grid_desc_m, thread_store_global_offset, PassThrough{}};
const index_t blockSize = get_block_size();
const index_t blockPerGrid = get_grid_size();
const auto m0 = c_grid_desc_m0.GetLength(I0);
const index_t loop_step = blockPerGrid * blockSize * ScalarPerVector;
const auto M = c_grid_desc_m.GetLength(I0);
const index_t loop_step = blockPerGrid * blockSize * MPerThread;
const auto loop_step_index = make_multi_index(loop_step);
index_t num_iter = m0 / (loop_step);
index_t num_iter = M / (loop_step);
do
{
// read and process ScalarPerVector elements
// read and process MPerThread elements
a_global_load.Run(
a_grid_desc_m0, a_global_buf, thread_desc_m0, make_tuple(I0), a_thread_buf);
a_grid_desc_m, a_global_buf, thread_desc_m, make_tuple(I0), a_thread_buf);
b_global_load.Run(
b_grid_desc_m0, b_global_buf, thread_desc_m0, make_tuple(I0), b_thread_buf);
b_grid_desc_m, b_global_buf, thread_desc_m, make_tuple(I0), b_thread_buf);
static_for<0, ScalarPerVector, 1>{}([&](auto m) {
constexpr auto offset = thread_desc_m0.CalculateOffset(make_tuple(m));
static_for<0, MPerThread, 1>{}([&](auto m) {
constexpr auto offset = thread_desc_m.CalculateOffset(make_tuple(m));
functor(c_thread_buf(Number<offset>{}),
a_thread_buf(Number<offset>{}),
b_thread_buf(Number<offset>{}));
});
c_global_write.Run(thread_desc_m0,
c_global_write.Run(thread_desc_m,
make_tuple(I0), // SrcSliceOriginIdx
c_thread_buf,
c_grid_desc_m0,
c_grid_desc_m,
c_global_buf);
a_global_load.MoveSrcSliceWindow(a_grid_desc_m0, loop_step_index);
b_global_load.MoveSrcSliceWindow(b_grid_desc_m0, loop_step_index);
c_global_write.MoveDstSliceWindow(c_grid_desc_m0, loop_step_index);
a_global_load.MoveSrcSliceWindow(a_grid_desc_m, loop_step_index);
b_global_load.MoveSrcSliceWindow(b_grid_desc_m, loop_step_index);
c_global_write.MoveDstSliceWindow(c_grid_desc_m, loop_step_index);
} while(--num_iter);
}
};
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_reduce_xdl_cshuffle_v1.hpp
View file @ 09ec28be
......@@ -21,7 +21,7 @@ template <typename GridwiseGemm,
typename BElementwiseOperation,
typename CElementwiseOperation,
typename DxsInElementwiseOperation,
typename DxsOutElementwiseOperation,
typename DxsAccElementwiseOperation,
typename AGridDesc_AK0_M_AK1,
typename BGridDesc_BK0_N_BK1,
typename CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock,
......@@ -41,7 +41,7 @@ __global__ void
const BElementwiseOperation b_element_op,
const CElementwiseOperation c_element_op,
const DxsInElementwiseOperation dxs_in_element_op,
const DxsOutElementwiseOperation dxs_out_element_op,
const DxsAccElementwiseOperation dxs_out_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 CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock
......@@ -96,7 +96,7 @@ template <typename FloatAB,
typename CElementwiseOperation,
typename DxsReduceOperation,
typename DxsInElementwiseOperation,
typename DxsOutElementwiseOperation,
typename DxsAccElementwiseOperation,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
typename DGlobalMemoryDataOperation,
typename AGridDesc_AK0_M_AK1,
......@@ -329,7 +329,7 @@ struct GridwiseGemmReduce_k0mk1_k0nk1_mn_xdl_cshuffle_v1
const BElementwiseOperation& b_element_op,
const CElementwiseOperation& c_element_op,
const DxsInElementwiseOperation& dxs_in_element_op,
const DxsOutElementwiseOperation& dxs_out_element_op,
const DxsAccElementwiseOperation& dxs_out_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 CGridDescriptor_MBlock_MPerBlock_NBlock_NPerBlock&
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp
View file @ 09ec28be
......@@ -314,7 +314,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
using DefaultBlock2CTileMap =
remove_cvref_t<decltype(MakeDefaultBlock2CTileMap(CGridDesc_M_N{}, 1, 1))>;
template <bool HasMainK0BlockLoop, typename Block2CTileMap = DefaultBlock2CTileMap>
template <bool HasMainK0BlockLoop, typename Block2CTileMap>
__device__ static void
Run(const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
......
include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp
View file @ 09ec28be
......@@ -25,6 +25,7 @@ enum struct MfmaInstr
mfma_f32_16x16x8bf16,
mfma_i32_32x32x8i8,
mfma_i32_16x16x16i8,
mfma_f64_16x16x4f64
};
template <MfmaInstr instr>
......@@ -383,12 +384,40 @@ struct mfma_type<MfmaInstr::mfma_i32_16x16x16i8>
}
};
template <>
struct mfma_type<MfmaInstr::mfma_f64_16x16x4f64>
{
static constexpr index_t group_size = 1;
static constexpr index_t num_groups_per_blk = 4;
static constexpr index_t num_regs_per_blk = 4; // group_size * num_groups_per_blk;
static constexpr index_t num_threads_per_blk = 16;
static constexpr index_t wave_size = 64;
static constexpr index_t num_input_blks = 4; // wave_size / num_threads_per_blk;
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 = 1;
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_f64_16x16x4f64<MPerXdlops, NPerXdlops>::Run(a, b, reg_c);
}
};
template <typename base_type, index_t MPerXdlops, index_t NPerXdlops>
struct MfmaSelector
{
template <typename base_type_, index_t MPerXdlops_, index_t NPerXdlops_>
static constexpr auto GetMfma();
template <>
static constexpr auto GetMfma<double, 16, 16>()
{
return MfmaInstr::mfma_f64_16x16x4f64;
}
template <>
static constexpr auto GetMfma<float, 64, 64>()
{
......@@ -661,9 +690,10 @@ struct XdlopsGemm
template <class FloatA, class FloatB, class FloatC>
__device__ void Run(const FloatA& p_a_wave, const FloatB& p_b_wave, FloatC& p_c_thread) const
{
static_assert(is_same<base_type, float>::value || is_same<base_type, half_t>::value ||
is_same<base_type, bhalf_t>::value || is_same<base_type, int8_t>::value,
"base base_type must be float, half, bfloat16, and int8_t!");
static_assert(is_same<base_type, double>::value || is_same<base_type, float>::value ||
is_same<base_type, half_t>::value || is_same<base_type, bhalf_t>::value ||
is_same<base_type, int8_t>::value,
"base base_type must be double, float, half, bfloat16, and int8_t!");
static_for<0, KPack / mfma_instr.k_per_blk, 1>{}([&](auto k) {
mfma_instr.template run<MPerXdlops, NPerXdlops>(p_a_wave[k], p_b_wave[k], p_c_thread);
......
include/ck/utility/amd_xdlops.hpp
View file @ 09ec28be
......@@ -294,5 +294,24 @@ struct intrin_mfma_i32_16x16x16i8<16, 16>
}
};
template <index_t MPerWave, index_t NPerWave>
struct intrin_mfma_f64_16x16x4f64;
template <>
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__
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
ignore = reg_a;
ignore = reg_b;
ignore = reg_c;
#endif
}
};
} // namespace ck
#endif
library/include/ck/library/reference_tensor_operation/cpu/reference_gemm.hpp
View file @ 09ec28be
......@@ -11,6 +11,7 @@ namespace host {
template <typename ADataType,
typename BDataType,
typename CDataType,
typename AccDataType,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation>
......@@ -53,20 +54,20 @@ struct ReferenceGemm : public device::BaseOperator
auto f_mk_kn_mn = [&](auto m, auto n) {
const int K = arg.a_m_k_.mDesc.GetLengths()[1];
float v_acc = 0;
AccDataType v_acc = 0;
for(int k = 0; k < K; ++k)
{
float v_a;
float v_b;
AccDataType v_a;
AccDataType v_b;
arg.a_element_op_(v_a, static_cast<const float>(arg.a_m_k_(m, k)));
arg.b_element_op_(v_b, static_cast<const float>(arg.b_k_n_(k, n)));
arg.a_element_op_(v_a, static_cast<const AccDataType>(arg.a_m_k_(m, k)));
arg.b_element_op_(v_b, static_cast<const AccDataType>(arg.b_k_n_(k, n)));
v_acc += v_a * v_b;
}
float v_c;
AccDataType v_c;
arg.c_element_op_(v_c, v_acc);
......
library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp
View file @ 09ec28be
......@@ -38,7 +38,7 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa
using device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instances =
std::tuple<
// clang-format off
//##################################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//##################################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsAccEleOp| D| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//##################################| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData|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_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//##################################| | | | | | | | | | | Operation| Operation| 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_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
......
library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp
View file @ 09ec28be
......@@ -38,7 +38,7 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa
using device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instances =
std::tuple<
// clang-format off
//##################################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//##################################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsAccEleOp| D| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//##################################| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData|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_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//##################################| | | | | | | | | | | Operation| Operation| 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_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
......
library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp
View file @ 09ec28be
......@@ -38,7 +38,7 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa
using device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances =
std::tuple<
// clang-format off
//##################################| ALayout| BLayout| CLayout| AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//##################################| ALayout| BLayout| CLayout| AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsAccEleOp| D| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//##################################| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData|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_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//##################################| | | | | | | | | | | Operation| Operation| 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_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
......
library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp
View file @ 09ec28be
......@@ -38,7 +38,7 @@ static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecializa
using device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instances =
std::tuple<
// clang-format off
//##################################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsOutEleOp| D| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//##################################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsAccEleOp| D| 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| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//##################################| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData|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_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//##################################| | | | | | | | | | | Operation| Operation| 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_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
......
library/src/tensor_operation_instance/gpu/gemm/CMakeLists.txt
View file @ 09ec28be
set(DEVICE_GEMM_INSTANCE_SOURCE
device_gemm_xdl_f64_f64_f64_mk_kn_mn_instance.cpp;
device_gemm_xdl_f64_f64_f64_mk_nk_mn_instance.cpp;
device_gemm_xdl_f64_f64_f64_km_kn_mn_instance.cpp;
device_gemm_xdl_f64_f64_f64_km_nk_mn_instance.cpp;
device_gemm_xdl_f32_f32_f32_mk_kn_mn_instance.cpp;
device_gemm_xdl_f32_f32_f32_mk_nk_mn_instance.cpp;
device_gemm_xdl_f32_f32_f32_km_kn_mn_instance.cpp;
......
library/src/tensor_operation_instance/gpu/gemm/device_gemm_xdl_f64_f64_f64_km_kn_mn_instance.cpp 0 → 100644
View file @ 09ec28be
#include <stdlib.h>
#include "config.hpp"
#include "device_gemm_xdl.hpp"
#include "element_wise_operation.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F64 = double;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
using device_gemm_xdl_f64_f64_f64_km_kn_mn_instances =
std::tuple<
// clang-format off
//##########| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CThreadTransfer| CThreadTransfer|
//##########| 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| SrcDstVectorDim| DstScalar|
//##########| | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | | PerVector|
//##########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmXdl< F64, F64, F64, F64, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 4, 2, 16, 16, 4, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, true, 7, 1>,
DeviceGemmXdl< F64, F64, F64, F64, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 128, 64, 4, 2, 16, 16, 4, 4, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, true, 7, 1>,
DeviceGemmXdl< F64, F64, F64, F64, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 128, 64, 128, 4, 2, 16, 16, 4, 4, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, true, 7, 1>,
DeviceGemmXdl< F64, F64, F64, F64, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 64, 4, 2, 16, 16, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, true, 7, 1>,
DeviceGemmXdl< F64, F64, F64, F64, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 64, 128, 4, 2, 16, 16, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 2, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 2, true, 7, 1>
// clang-format on
>;
void add_device_gemm_xdl_f64_f64_f64_km_kn_mn_instances(
std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
{
add_device_operation_instances(instances, device_gemm_xdl_f64_f64_f64_km_kn_mn_instances{});
}
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // 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