Skip to content

GitLab

Commit 65a317de authored by Po-Yen, Chen's avatar Po-Yen, Chen
Browse files

Merge branch 'feature/fix-imcomplete-support-of-amd-wave-read-first-lane' of... 

Merge branch 'feature/fix-imcomplete-support-of-amd-wave-read-first-lane' of github.com:ROCmSoftwarePlatform/composable_kernel into feature/fix-imcomplete-support-of-amd-wave-read-first-lane
parents 16ee6676 30ab4fb0
Expand all Show whitespace changes
Inline Side-by-side
Showing with 846 additions and 1168 deletions
example/17_convnd_bwd_data/convnd_bwd_data_common.hpp
View file @ 65a317de
......@@ -83,7 +83,8 @@ int run_conv_bwd_data(bool do_verification,
// do GEMM
auto conv = DeviceConvNdBwdDataInstance{};
auto invoker = conv.MakeInvoker();
auto argument = conv.MakeArgument(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
auto argument =
conv.MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
conv_param.N_,
......@@ -100,13 +101,13 @@ int run_conv_bwd_data(bool do_verification,
wei_element_op,
out_element_op);
if(!conv.IsSupportedArgument(argument))
if(!conv.IsSupportedArgument(argument.get()))
{
std::cout << "Not support,please check parameters or device";
return 0;
}
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
float ave_time = invoker.Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = conv_param.GetFlops();
std::size_t num_btype = conv_param.GetByte<InDataType, WeiDataType, OutDataType>();
......
include/ck/tensor_operation/gpu/device/device_grouped_gemm_softmax_gemm_permute_xdl_cshuffle.hpp
View file @ 65a317de
......@@ -611,10 +611,11 @@ struct DeviceGroupedGemmSoftmaxGemmPermute_Xdl_CShuffle
some_has_main_k_block_loop |= y;
}
hipGetErrorString(hipMemcpy(arg.p_workspace_,
hipGetErrorString(hipMemcpyWithStream(arg.p_workspace_,
arg.group_kernel_args_.data(),
arg.group_kernel_args_.size() * sizeof(GroupKernelArg),
hipMemcpyHostToDevice));
hipMemcpyHostToDevice,
stream_config.stream_id_));
float ave_time = 0;
......
include/ck/tensor_operation/gpu/device/impl/device_conv2d_bwd_data_xdl_nhwc_kyxc_nhwk.hpp
View file @ 65a317de
......@@ -379,9 +379,6 @@ struct DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
AccDataType,
CDataType,
InMemoryDataOperationEnum::Set,
AGridDesc_K0_M_K1,
BGridDesc_K0_N_K1,
CGridDesc_M_N,
InElementwiseOperation,
WeiElementwiseOperation,
OutElementwiseOperation,
......@@ -428,20 +425,10 @@ struct DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
std::vector<ck::index_t> conv_filter_strides,
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)
std::vector<ck::index_t> input_right_pads)
: p_a_grid_{p_out_grid},
p_b_grid_{p_wei_grid},
p_c_grid_{p_in_grid},
M01_{M01},
N01_{N01},
a_element_op_{out_element_op},
b_element_op_{wei_element_op},
c_element_op_{in_element_op},
Conv_N_{N},
Conv_K_{K},
Conv_C_{C},
......@@ -495,18 +482,6 @@ struct DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
a_grid_desc_k0_m_k1_container_.push_back(descs[I0]);
b_grid_desc_k0_n_k1_container_.push_back(descs[I1]);
c_grid_desc_m_n_container_.push_back(descs[I2]);
auto block_2_ctile_map =
GridwiseGemm::MakeDefaultBlock2CTileMap(descs[I2], M01, N01);
if(GridwiseGemm::CheckValidity(
descs[I0], descs[I1], descs[I2], block_2_ctile_map))
{
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_.push_back(
GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(descs[I2]));
block_2_ctile_map_container_.push_back(block_2_ctile_map);
}
}
}
}
......@@ -517,14 +492,6 @@ struct DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
std::vector<AGridDesc_K0_M_K1> a_grid_desc_k0_m_k1_container_;
std::vector<BGridDesc_K0_N_K1> b_grid_desc_k0_n_k1_container_;
std::vector<CGridDesc_M_N> c_grid_desc_m_n_container_;
std::vector<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_;
std::vector<typename GridwiseGemm::DefaultBlock2CTileMap> block_2_ctile_map_container_;
index_t M01_;
index_t N01_;
OutElementwiseOperation a_element_op_;
WeiElementwiseOperation b_element_op_;
InElementwiseOperation c_element_op_;
// for checking IsSupportedArgument()
index_t Conv_N_;
index_t Conv_K_;
......@@ -567,58 +534,37 @@ struct DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<< arg.c_grid_desc_m_n_container_[i].GetLength(I0) << ", "
<< arg.c_grid_desc_m_n_container_[i].GetLength(I1) << "}"
<< std::endl;
std::cout << "arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_( "
<< arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I0)
<< ", "
<< arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I1)
<< ", "
<< arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I2)
<< ", "
<< arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I3)
<< ", "
<< arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I4)
<< ", "
<< arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I5)
<< " ) " << std::endl;
}
#endif
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_container_[i],
arg.b_grid_desc_k0_n_k1_container_[i],
arg.c_grid_desc_m_n_container_[i],
arg.block_2_ctile_map_container_[i]))
arg.c_grid_desc_m_n_container_[i]))
{
throw std::runtime_error(
"wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v3r1 has invalid setting");
}
const index_t grid_size = arg.block_2_ctile_map_container_[i].CalculateGridSize(
arg.c_grid_desc_m_n_container_[i]);
const auto [gdx, gdy, gdz] =
GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_container_[i]);
const auto K = arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I0) *
arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I2);
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
{
const auto kernel = kernel_gemm_xdlops_v2r3<
GridwiseGemm,
const auto kernel =
kernel_gemm_xdlops_v2r3<GridwiseGemm,
ADataType, // TODO: distiguish A/B datatype
CDataType,
remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
remove_reference_t<
typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
OutElementwiseOperation,
WeiElementwiseOperation,
InElementwiseOperation,
remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
DeviceOp::AGridDesc_K0_M_K1,
DeviceOp::BGridDesc_K0_N_K1,
DeviceOp::CGridDesc_M_N,
true>;
ave_time += launch_and_time_kernel(
stream_config,
ave_time += launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(gdx, gdy, gdz),
dim3(BlockSize),
0,
arg.p_a_grid_,
......@@ -626,32 +572,22 @@ struct DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
arg.p_c_grid_,
arg.a_grid_desc_k0_m_k1_container_[i],
arg.b_grid_desc_k0_n_k1_container_[i],
arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i],
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_,
arg.block_2_ctile_map_container_[i]);
arg.c_grid_desc_m_n_container_[i]);
}
else
{
const auto kernel = kernel_gemm_xdlops_v2r3<
GridwiseGemm,
const auto kernel =
kernel_gemm_xdlops_v2r3<GridwiseGemm,
ADataType, // TODO: distiguish A/B datatype
CDataType,
remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
remove_reference_t<
typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
OutElementwiseOperation,
WeiElementwiseOperation,
InElementwiseOperation,
remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
DeviceOp::AGridDesc_K0_M_K1,
DeviceOp::BGridDesc_K0_N_K1,
DeviceOp::CGridDesc_M_N,
false>;
ave_time += launch_and_time_kernel(
stream_config,
ave_time += launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(gdx, gdy, gdz),
dim3(BlockSize),
0,
arg.p_a_grid_,
......@@ -659,11 +595,7 @@ struct DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
arg.p_c_grid_,
arg.a_grid_desc_k0_m_k1_container_[i],
arg.b_grid_desc_k0_n_k1_container_[i],
arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i],
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_,
arg.block_2_ctile_map_container_[i]);
arg.c_grid_desc_m_n_container_[i]);
}
}
return ave_time;
......@@ -716,8 +648,7 @@ struct DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
{
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_container_[i],
arg.b_grid_desc_k0_n_k1_container_[i],
arg.c_grid_desc_m_n_container_[i],
arg.block_2_ctile_map_container_[i]))
arg.c_grid_desc_m_n_container_[i]))
{
return false;
}
......@@ -742,10 +673,7 @@ struct DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op)
std::vector<ck::index_t> input_right_pads)
{
return Argument{p_in_grid,
p_wei_grid,
......@@ -759,12 +687,7 @@ struct DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
1,
1,
in_element_op,
wei_element_op,
out_element_op};
input_right_pads};
}
static auto MakeInvoker() { return Invoker{}; }
......@@ -783,9 +706,9 @@ struct DeviceConv2dBwdDataXdl_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,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op) override
InElementwiseOperation,
WeiElementwiseOperation,
OutElementwiseOperation) override
{
return std::make_unique<Argument>(static_cast<InDataType*>(p_in_grid),
static_cast<const WeiDataType*>(p_wei_grid),
......@@ -799,12 +722,7 @@ struct DeviceConv2dBwdDataXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
1,
1,
in_element_op,
wei_element_op,
out_element_op);
input_right_pads);
}
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
......
include/ck/tensor_operation/gpu/device/impl/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp
View file @ 65a317de
......@@ -329,9 +329,6 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
AccDataType,
CDataType,
InMemoryDataOperationEnum::Set,
AGridDesc_K0_M_K1,
BGridDesc_K0_N_K1,
CGridDesc_M_N,
InElementwiseOperation,
WeiElementwiseOperation,
OutElementwiseOperation,
......@@ -378,25 +375,13 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
std::vector<ck::index_t> conv_filter_strides,
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)
std::vector<ck::index_t> input_right_pads)
: p_a_grid_{p_in_grid},
p_b_grid_{p_wei_grid},
p_c_grid_{p_out_grid},
a_grid_desc_k0_m_k1_{},
b_grid_desc_k0_n_k1_{},
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},
Conv_N_{N},
Conv_K_{K},
Conv_C_{C},
......@@ -420,17 +405,6 @@ struct DeviceConv2dFwdXdl_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);
if(GridwiseGemm::CheckValidity(a_grid_desc_k0_m_k1_,
b_grid_desc_k0_n_k1_,
c_grid_desc_m_n_,
block_2_ctile_map_))
{
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_ =
GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m_n_);
}
}
// private:
......@@ -440,14 +414,6 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
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_;
InElementwiseOperation in_element_op_;
WeiElementwiseOperation wei_element_op_;
OutElementwiseOperation out_element_op_;
// for checking IsSupportedArgument()
index_t Conv_N_;
index_t Conv_K_;
......@@ -479,17 +445,14 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
<< arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
}
#endif
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_m_n_,
arg.block_2_ctile_map_))
if(!GridwiseGemm::CheckValidity(
arg.a_grid_desc_k0_m_k1_, arg.b_grid_desc_k0_n_k1_, arg.c_grid_desc_m_n_))
{
throw std::runtime_error(
"wrong! GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3 has invalid setting");
}
const index_t grid_size =
arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_);
const auto [gdx, gdy, gdz] = GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_);
const auto K =
arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
......@@ -498,22 +461,18 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
{
const auto kernel = kernel_gemm_xdlops_v2r3<
GridwiseGemm,
const auto kernel =
kernel_gemm_xdlops_v2r3<GridwiseGemm,
ADataType, // TODO: distiguish A/B datatype
CDataType,
remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
InElementwiseOperation,
WeiElementwiseOperation,
OutElementwiseOperation,
remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
DeviceOp::AGridDesc_K0_M_K1,
DeviceOp::BGridDesc_K0_N_K1,
DeviceOp::CGridDesc_M_N,
true>;
ave_time = launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(gdx, gdy, gdz),
dim3(BlockSize),
0,
arg.p_a_grid_,
......@@ -521,30 +480,22 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
arg.p_c_grid_,
arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
arg.in_element_op_,
arg.wei_element_op_,
arg.out_element_op_,
arg.block_2_ctile_map_);
arg.c_grid_desc_m_n_);
}
else
{
const auto kernel = kernel_gemm_xdlops_v2r3<
GridwiseGemm,
const auto kernel =
kernel_gemm_xdlops_v2r3<GridwiseGemm,
ADataType, // TODO: distiguish A/B datatype
CDataType,
remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
InElementwiseOperation,
WeiElementwiseOperation,
OutElementwiseOperation,
remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
DeviceOp::AGridDesc_K0_M_K1,
DeviceOp::BGridDesc_K0_N_K1,
DeviceOp::CGridDesc_M_N,
false>;
ave_time = launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(gdx, gdy, gdz),
dim3(BlockSize),
0,
arg.p_a_grid_,
......@@ -552,11 +503,7 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
arg.p_c_grid_,
arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
arg.in_element_op_,
arg.wei_element_op_,
arg.out_element_op_,
arg.block_2_ctile_map_);
arg.c_grid_desc_m_n_);
}
return ave_time;
......@@ -616,10 +563,8 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
}
// Gridwise GEMM size
return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_m_n_,
arg.block_2_ctile_map_);
return GridwiseGemm::CheckValidity(
arg.a_grid_desc_k0_m_k1_, arg.b_grid_desc_k0_n_k1_, arg.c_grid_desc_m_n_);
}
bool IsSupportedArgument(const BaseArgument* p_arg) override
......@@ -639,10 +584,7 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op)
std::vector<ck::index_t> input_right_pads)
{
return Argument{p_in_grid,
p_wei_grid,
......@@ -656,12 +598,7 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
1,
1,
in_element_op,
wei_element_op,
out_element_op};
input_right_pads};
}
static auto MakeInvoker() { return Invoker{}; }
......@@ -680,9 +617,9 @@ struct DeviceConv2dFwdXdl_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,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op) override
InElementwiseOperation,
WeiElementwiseOperation,
OutElementwiseOperation) override
{
return std::make_unique<Argument>(static_cast<const InDataType*>(p_in_grid),
static_cast<const WeiDataType*>(p_wei_grid),
......@@ -696,12 +633,7 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
1,
1,
in_element_op,
wei_element_op,
out_element_op);
input_right_pads);
}
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
......
include/ck/tensor_operation/gpu/device/impl/device_convnd_bwd_data_nwc_kxc_nwk_xdl.hpp
View file @ 65a317de
......@@ -980,9 +980,6 @@ struct DeviceConvNdBwdDataNwcKxcNwk_Xdl
AccDataType,
CDataType,
InMemoryDataOperationEnum::Set,
AGridDesc_K0_M_K1,
BGridDesc_K0_N_K1,
CGridDesc_M_N,
InElementwiseOperation,
WeiElementwiseOperation,
OutElementwiseOperation,
......@@ -1029,20 +1026,10 @@ struct DeviceConvNdBwdDataNwcKxcNwk_Xdl
std::vector<ck::index_t> conv_filter_strides,
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)
std::vector<ck::index_t> input_right_pads)
: p_a_grid_{p_out_grid},
p_b_grid_{p_wei_grid},
p_c_grid_{p_in_grid},
M01_{M01},
N01_{N01},
a_element_op_{out_element_op},
b_element_op_{wei_element_op},
c_element_op_{in_element_op},
Conv_N_{N},
Conv_K_{K},
Conv_C_{C},
......@@ -1092,17 +1079,6 @@ struct DeviceConvNdBwdDataNwcKxcNwk_Xdl
a_grid_desc_k0_m_k1_container_.push_back(descs[I0]);
b_grid_desc_k0_n_k1_container_.push_back(descs[I1]);
c_grid_desc_m_n_container_.push_back(descs[I2]);
auto block_2_ctile_map =
GridwiseGemm::MakeDefaultBlock2CTileMap(descs[I2], M01_, N01_);
if(GridwiseGemm::CheckValidity(descs[I0], descs[I1], descs[I2], block_2_ctile_map))
{
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_.push_back(
GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(descs[I2]));
block_2_ctile_map_container_.push_back(block_2_ctile_map);
}
}
}
template <ck::index_t NDim, typename ck::enable_if<NDim == 2, bool>::type = false>
......@@ -1150,18 +1126,6 @@ struct DeviceConvNdBwdDataNwcKxcNwk_Xdl
a_grid_desc_k0_m_k1_container_.push_back(descs[I0]);
b_grid_desc_k0_n_k1_container_.push_back(descs[I1]);
c_grid_desc_m_n_container_.push_back(descs[I2]);
auto block_2_ctile_map =
GridwiseGemm::MakeDefaultBlock2CTileMap(descs[I2], M01_, N01_);
if(GridwiseGemm::CheckValidity(
descs[I0], descs[I1], descs[I2], block_2_ctile_map))
{
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_.push_back(
GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(descs[I2]));
block_2_ctile_map_container_.push_back(block_2_ctile_map);
}
}
}
}
......@@ -1218,19 +1182,6 @@ struct DeviceConvNdBwdDataNwcKxcNwk_Xdl
a_grid_desc_k0_m_k1_container_.push_back(descs[I0]);
b_grid_desc_k0_n_k1_container_.push_back(descs[I1]);
c_grid_desc_m_n_container_.push_back(descs[I2]);
auto block_2_ctile_map =
GridwiseGemm::MakeDefaultBlock2CTileMap(descs[I2], M01_, N01_);
if(GridwiseGemm::CheckValidity(
descs[I0], descs[I1], descs[I2], block_2_ctile_map))
{
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_.push_back(
GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(
descs[I2]));
block_2_ctile_map_container_.push_back(block_2_ctile_map);
}
}
}
}
......@@ -1242,11 +1193,6 @@ struct DeviceConvNdBwdDataNwcKxcNwk_Xdl
std::vector<AGridDesc_K0_M_K1> a_grid_desc_k0_m_k1_container_;
std::vector<BGridDesc_K0_N_K1> b_grid_desc_k0_n_k1_container_;
std::vector<CGridDesc_M_N> c_grid_desc_m_n_container_;
std::vector<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_;
std::vector<typename GridwiseGemm::DefaultBlock2CTileMap> block_2_ctile_map_container_;
index_t M01_;
index_t N01_;
OutElementwiseOperation a_element_op_;
WeiElementwiseOperation b_element_op_;
InElementwiseOperation c_element_op_;
......@@ -1276,78 +1222,53 @@ struct DeviceConvNdBwdDataNwcKxcNwk_Xdl
{
#if DEBUG_LOG
{
std::cout << "arg.a_grid_desc_k0_m_k1_container_{"
std::cout << "arg.a_grid_desc_k0_m_k1{"
<< arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I0) << ", "
<< arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I1) << ", "
<< arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I2) << "}"
<< std::endl;
std::cout << "arg.b_grid_desc_k0_n_k1_container_{"
std::cout << "arg.b_grid_desc_k0_n_k1{"
<< arg.b_grid_desc_k0_n_k1_container_[i].GetLength(I0) << ", "
<< arg.b_grid_desc_k0_n_k1_container_[i].GetLength(I1) << ", "
<< arg.b_grid_desc_k0_n_k1_container_[i].GetLength(I2) << "}"
<< std::endl;
std::cout << "arg.c_grid_desc_m_n_container_{ "
std::cout << "arg.c_grid_desc_m_n{"
<< arg.c_grid_desc_m_n_container_[i].GetLength(I0) << ", "
<< arg.c_grid_desc_m_n_container_[i].GetLength(I1) << "}"
<< std::endl;
std::cout << "arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_( "
<< arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I0)
<< ", "
<< arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I1)
<< ", "
<< arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I2)
<< ", "
<< arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I3)
<< ", "
<< arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I4)
<< ", "
<< arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I5)
<< ", "
<< arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I6)
<< ", "
<< arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i].GetLength(I7)
<< " ) " << std::endl;
}
#endif
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_container_[i],
arg.b_grid_desc_k0_n_k1_container_[i],
arg.c_grid_desc_m_n_container_[i],
arg.block_2_ctile_map_container_[i]))
arg.c_grid_desc_m_n_container_[i]))
{
throw std::runtime_error(
"wrong! GridwiseGemm_km_kn_m0m1n0n1_xdlops_v3r1 has invalid setting");
"wrong! GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3 has invalid setting");
}
const index_t grid_size = arg.block_2_ctile_map_container_[i].CalculateGridSize(
arg.c_grid_desc_m_n_container_[i]);
const auto [gdx, gdy, gdz] =
GridwiseGemm::CalculateGridSize(arg.c_grid_desc_m_n_container_[i]);
const auto K = arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I0) *
arg.a_grid_desc_k0_m_k1_container_[i].GetLength(I2);
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
{
const auto kernel = kernel_gemm_xdlops_v2r3<
GridwiseGemm,
const auto kernel =
kernel_gemm_xdlops_v2r3<GridwiseGemm,
ADataType, // TODO: distiguish A/B datatype
CDataType,
remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
remove_reference_t<
typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
OutElementwiseOperation,
WeiElementwiseOperation,
InElementwiseOperation,
remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
DeviceOp::AGridDesc_K0_M_K1,
DeviceOp::BGridDesc_K0_N_K1,
DeviceOp::CGridDesc_M_N,
true>;
ave_time += launch_and_time_kernel(
stream_config,
ave_time += launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(gdx, gdy, gdz),
dim3(BlockSize),
0,
arg.p_a_grid_,
......@@ -1355,32 +1276,22 @@ struct DeviceConvNdBwdDataNwcKxcNwk_Xdl
arg.p_c_grid_,
arg.a_grid_desc_k0_m_k1_container_[i],
arg.b_grid_desc_k0_n_k1_container_[i],
arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i],
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_,
arg.block_2_ctile_map_container_[i]);
arg.c_grid_desc_m_n_container_[i]);
}
else
{
const auto kernel = kernel_gemm_xdlops_v2r3<
GridwiseGemm,
const auto kernel =
kernel_gemm_xdlops_v2r3<GridwiseGemm,
ADataType, // TODO: distiguish A/B datatype
CDataType,
remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
remove_reference_t<
typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
OutElementwiseOperation,
WeiElementwiseOperation,
InElementwiseOperation,
remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
DeviceOp::AGridDesc_K0_M_K1,
DeviceOp::BGridDesc_K0_N_K1,
DeviceOp::CGridDesc_M_N,
false>;
ave_time += launch_and_time_kernel(
stream_config,
ave_time += launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(gdx, gdy, gdz),
dim3(BlockSize),
0,
arg.p_a_grid_,
......@@ -1388,11 +1299,7 @@ struct DeviceConvNdBwdDataNwcKxcNwk_Xdl
arg.p_c_grid_,
arg.a_grid_desc_k0_m_k1_container_[i],
arg.b_grid_desc_k0_n_k1_container_[i],
arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_container_[i],
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_,
arg.block_2_ctile_map_container_[i]);
arg.c_grid_desc_m_n_container_[i]);
}
}
return ave_time;
......@@ -1446,8 +1353,7 @@ struct DeviceConvNdBwdDataNwcKxcNwk_Xdl
{
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_container_[i],
arg.b_grid_desc_k0_n_k1_container_[i],
arg.c_grid_desc_m_n_container_[i],
arg.block_2_ctile_map_container_[i]))
arg.c_grid_desc_m_n_container_[i]))
{
return false;
}
......@@ -1472,10 +1378,7 @@ struct DeviceConvNdBwdDataNwcKxcNwk_Xdl
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op)
std::vector<ck::index_t> input_right_pads)
{
return Argument{p_in_grid,
p_wei_grid,
......@@ -1489,12 +1392,7 @@ struct DeviceConvNdBwdDataNwcKxcNwk_Xdl
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
1,
1,
in_element_op,
wei_element_op,
out_element_op};
input_right_pads};
}
static auto MakeInvoker() { return Invoker{}; }
......@@ -1513,9 +1411,9 @@ struct DeviceConvNdBwdDataNwcKxcNwk_Xdl
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op) override
InElementwiseOperation,
WeiElementwiseOperation,
OutElementwiseOperation) override
{
return std::make_unique<Argument>(static_cast<InDataType*>(p_in_grid),
static_cast<const WeiDataType*>(p_wei_grid),
......@@ -1529,12 +1427,7 @@ struct DeviceConvNdBwdDataNwcKxcNwk_Xdl
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
1,
1,
in_element_op,
wei_element_op,
out_element_op);
input_right_pads);
}
std::unique_ptr<BaseInvoker> MakeInvokerPointer() override
......
include/ck/tensor_operation/gpu/device/impl/device_gemm_xdl.hpp
View file @ 65a317de
......@@ -75,132 +75,20 @@ struct DeviceGemmXdl : public DeviceGemm<ALayout,
static constexpr auto K1Number = Number<K1>{};
static auto MakeAGridDescriptor_K0_M_K1(index_t M, index_t K, index_t StrideA)
{
const index_t K0 = K / K1;
const auto a_grid_desc_m_k = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, ALayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(StrideA, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, ALayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, K), make_tuple(I1, StrideA));
}
}();
if constexpr(GemmSpec == GemmSpecialization::MNPadding)
{
const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
return transform_tensor_descriptor(
a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
make_right_pad_transform(M, PadM)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
else
{
return transform_tensor_descriptor(
a_grid_desc_m_k,
make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
make_pass_through_transform(M)),
make_tuple(Sequence<1>{}, Sequence<0>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
}
static auto MakeBGridDescriptor_K0_N_K1(index_t K, index_t N, index_t StrideB)
{
const index_t K0 = K / K1;
const auto b_grid_desc_k_n = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(StrideB, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, BLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(K, N), make_tuple(I1, StrideB));
}
}();
if constexpr(GemmSpec == GemmSpecialization::MNPadding)
{
const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
return transform_tensor_descriptor(
b_grid_desc_k_n,
make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
make_right_pad_transform(N, PadN)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
else
{
return transform_tensor_descriptor(
b_grid_desc_k_n,
make_tuple(make_unmerge_transform(make_tuple(K0, K1Number)),
make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 2>{}, Sequence<1>{}));
}
}
static auto MakeCGridDescriptor_M_N(index_t M, index_t N, index_t StrideC)
{
const auto c_grid_desc_m_n = [&]() {
if constexpr(is_same<tensor_layout::gemm::RowMajor, CLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(StrideC, I1));
}
else if constexpr(is_same<tensor_layout::gemm::ColumnMajor, CLayout>::value)
{
return make_naive_tensor_descriptor(make_tuple(M, N), make_tuple(I1, StrideC));
}
}();
if constexpr(GemmSpec == GemmSpecialization::MNPadding)
{
const auto PadM = (MPerBlock - M % MPerBlock) % MPerBlock;
const auto PadN = (NPerBlock - N % NPerBlock) % NPerBlock;
return transform_tensor_descriptor(
c_grid_desc_m_n,
make_tuple(make_right_pad_transform(M, PadM), make_right_pad_transform(N, PadN)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
else
{
return transform_tensor_descriptor(
c_grid_desc_m_n,
make_tuple(make_pass_through_transform(M), make_pass_through_transform(N)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
}
using AGridDesc_K0_M_K1 = decltype(MakeAGridDescriptor_K0_M_K1(1, 1, 1));
using BGridDesc_K0_N_K1 = decltype(MakeBGridDescriptor_K0_N_K1(1, 1, 1));
using CGridDesc_M_N = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
// GridwiseGemm
using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<
using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3_ext<
BlockSize,
ADataType, // TODO: distinguish A/B datatype
AccDataType,
CDataType,
InMemoryDataOperationEnum::Set,
AGridDesc_K0_M_K1,
BGridDesc_K0_N_K1,
CGridDesc_M_N,
ALayout,
BLayout,
CLayout,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
GemmSpec,
MPerBlock,
NPerBlock,
K0PerBlock,
......@@ -232,173 +120,41 @@ struct DeviceGemmXdl : public DeviceGemm<ALayout,
LoopSched,
PipelineVer>;
// Argument
struct Argument : public BaseArgument
{
Argument(const ADataType* p_a_grid,
const BDataType* p_b_grid,
CDataType* p_c_grid,
index_t M,
index_t N,
index_t K,
index_t StrideA,
index_t StrideB,
index_t StrideC,
index_t M01,
index_t N01,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
: p_a_grid_{p_a_grid},
p_b_grid_{p_b_grid},
p_c_grid_{p_c_grid},
a_grid_desc_k0_m_k1_{},
b_grid_desc_k0_n_k1_{},
c_grid_desc_m_n_{},
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_{},
block_2_ctile_map_{},
M01_{M01},
N01_{N01},
a_element_op_{a_element_op},
b_element_op_{b_element_op},
c_element_op_{c_element_op},
kraw_{K}
{
a_grid_desc_k0_m_k1_ = DeviceGemmXdl::MakeAGridDescriptor_K0_M_K1(M, K, StrideA);
b_grid_desc_k0_n_k1_ = DeviceGemmXdl::MakeBGridDescriptor_K0_N_K1(K, N, StrideB);
c_grid_desc_m_n_ = DeviceGemmXdl::MakeCGridDescriptor_M_N(M, N, StrideC);
block_2_ctile_map_ =
GridwiseGemm::MakeDefaultBlock2CTileMap(c_grid_desc_m_n_, M01, N01);
if(GridwiseGemm::CheckValidity(a_grid_desc_k0_m_k1_,
b_grid_desc_k0_n_k1_,
c_grid_desc_m_n_,
block_2_ctile_map_))
{
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_ =
GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(c_grid_desc_m_n_);
}
}
// private:
const ADataType* p_a_grid_;
const BDataType* p_b_grid_;
CDataType* p_c_grid_;
AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
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_;
AElementwiseOperation a_element_op_;
BElementwiseOperation b_element_op_;
CElementwiseOperation c_element_op_;
index_t kraw_;
};
using Argument = typename GridwiseGemm::Argument;
// Invoker
struct Invoker : public BaseInvoker
{
using Argument = DeviceGemmXdl::Argument;
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
float Run(const Argument& karg, const StreamConfig& stream_config = StreamConfig{})
{
#if DEBUG_LOG
if(stream_config.log_level_ > 0)
{
std::cout << "arg.a_grid_desc_k0_m_k1_{" << arg.a_grid_desc_k0_m_k1_.GetLength(I0)
<< ", " << arg.a_grid_desc_k0_m_k1_.GetLength(I1) << ", "
<< arg.a_grid_desc_k0_m_k1_.GetLength(I2) << "}" << std::endl;
std::cout << "arg.b_grid_desc_k0_n_k1_{" << arg.b_grid_desc_k0_n_k1_.GetLength(I0)
<< ", " << arg.b_grid_desc_k0_n_k1_.GetLength(I1) << ", "
<< arg.b_grid_desc_k0_n_k1_.GetLength(I2) << "}" << std::endl;
std::cout << "arg.c_grid_desc_m_n_{ " << arg.c_grid_desc_m_n_.GetLength(I0) << ", "
<< arg.c_grid_desc_m_n_.GetLength(I1) << "}" << std::endl;
karg.Print();
}
#endif
if(!GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_m_n_,
arg.block_2_ctile_map_))
if(!GridwiseGemm::CheckValidity(karg))
{
throw std::runtime_error(
"wrong! GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3 has invalid setting");
"wrong! GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3_ext has invalid setting");
}
const index_t grid_size =
arg.block_2_ctile_map_.CalculateGridSize(arg.c_grid_desc_m_n_);
const auto K =
arg.a_grid_desc_k0_m_k1_.GetLength(I0) * arg.a_grid_desc_k0_m_k1_.GetLength(I2);
const auto [gdx, gdy, gdz] = GridwiseGemm::CalculateGridSize(karg.M, karg.N);
float ave_time = 0;
if(GridwiseGemm::CalculateHasMainKBlockLoop(K))
if(GridwiseGemm::CalculateHasMainKBlockLoop(karg.K))
{
const auto kernel = kernel_gemm_xdlops_v2r3<
GridwiseGemm,
ADataType, // TODO: distiguish A/B datatype
CDataType,
remove_reference_t<DeviceGemmXdl::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceGemmXdl::BGridDesc_K0_N_K1>,
remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
true>;
ave_time = launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_,
arg.block_2_ctile_map_);
const auto kernel = kernel_gemm_xdlops_v2r3<GridwiseGemm, true>;
ave_time = launch_and_time_kernel(
stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, karg);
}
else
{
const auto kernel = kernel_gemm_xdlops_v2r3<
GridwiseGemm,
ADataType, // TODO: distiguish A/B datatype
CDataType,
remove_reference_t<DeviceGemmXdl::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceGemmXdl::BGridDesc_K0_N_K1>,
remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
remove_reference_t<typename GridwiseGemm::DefaultBlock2CTileMap>,
false>;
ave_time = launch_and_time_kernel(stream_config,
kernel,
dim3(grid_size),
dim3(BlockSize),
0,
arg.p_a_grid_,
arg.p_b_grid_,
arg.p_c_grid_,
arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_,
arg.a_element_op_,
arg.b_element_op_,
arg.c_element_op_,
arg.block_2_ctile_map_);
const auto kernel = kernel_gemm_xdlops_v2r3<GridwiseGemm, false>;
ave_time = launch_and_time_kernel(
stream_config, kernel, dim3(gdx, gdy, gdz), dim3(BlockSize), 0, karg);
}
return ave_time;
......@@ -418,7 +174,7 @@ struct DeviceGemmXdl : public DeviceGemm<ALayout,
return true;
}
static bool IsSupportedArgument(const Argument& arg)
static bool IsSupportedArgument(const Argument& karg)
{
if(ck::get_device_name() == "gfx908")
{
......@@ -441,15 +197,12 @@ struct DeviceGemmXdl : public DeviceGemm<ALayout,
return false;
}
if(arg.kraw_ % K1 != 0)
if(karg.K % K1 != 0)
{
return false;
}
return GridwiseGemm::CheckValidity(arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_m_n_,
arg.block_2_ctile_map_);
return GridwiseGemm::CheckValidity(karg);
}
// polymorphic
......@@ -467,24 +220,11 @@ struct DeviceGemmXdl : public DeviceGemm<ALayout,
index_t StrideA,
index_t StrideB,
index_t StrideC,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op)
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation)
{
return Argument{p_a,
p_b,
p_c,
M,
N,
K,
StrideA,
StrideB,
StrideC,
1,
1,
a_element_op,
b_element_op,
c_element_op};
return Argument{p_a, p_b, p_c, M, N, K, StrideA, StrideB, StrideC};
}
static auto MakeInvoker() { return Invoker{}; }
......@@ -499,9 +239,9 @@ struct DeviceGemmXdl : public DeviceGemm<ALayout,
index_t StrideA,
index_t StrideB,
index_t StrideC,
AElementwiseOperation a_element_op,
BElementwiseOperation b_element_op,
CElementwiseOperation c_element_op) override
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation) override
{
return std::make_unique<Argument>(static_cast<const ADataType*>(p_a),
static_cast<const BDataType*>(p_b),
......@@ -511,12 +251,7 @@ struct DeviceGemmXdl : public DeviceGemm<ALayout,
K,
StrideA,
StrideB,
StrideC,
1,
1,
a_element_op,
b_element_op,
c_element_op);
StrideC);
}
// polymorphic
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_contraction_multiple_d_xdl_cshuffle.hpp
View file @ 65a317de
......@@ -652,11 +652,12 @@ struct DeviceGroupedContractionMultipleD_Xdl_CShuffle
}
}
hipGetErrorString(hipMemcpy(arg.p_workspace_,
hipGetErrorString(hipMemcpyWithStream(arg.p_workspace_,
arg.contraction_multi_d_kernel_args_.data(),
arg.contraction_multi_d_kernel_args_.size() *
sizeof(ContractionMultiDKernelArg),
hipMemcpyHostToDevice));
hipMemcpyHostToDevice,
stream_config.stream_id_));
float ave_time = 0;
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_multiple_d_dl.hpp
View file @ 65a317de
......@@ -597,10 +597,11 @@ struct DeviceGroupedGemmMultipleD_Dl : public DeviceGroupedGemm<ALayout,
}
}
hipGetErrorString(hipMemcpy(arg.p_workspace_,
hipGetErrorString(hipMemcpyWithStream(arg.p_workspace_,
arg.gemm_desc_kernel_arg_.data(),
arg.gemm_desc_kernel_arg_.size() * sizeof(GemmKernelArg),
hipMemcpyHostToDevice));
hipMemcpyHostToDevice,
stream_config.stream_id_));
auto launch_kernel = [&](auto has_main_k_block_loop,
auto has_double_tail_k_block_loop) {
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl.hpp
View file @ 65a317de
......@@ -549,10 +549,11 @@ struct DeviceGroupedGemm_Xdl : public DeviceGroupedGemm<ALayout,
}
hipGetErrorString(
hipMemcpy(arg.p_workspace_,
hipMemcpyWithStream(arg.p_workspace_,
arg.gemm_desc_kernel_arg_.data(),
arg.gemm_desc_kernel_arg_.size() * sizeof(GemmBiasTransKernelArg),
hipMemcpyHostToDevice));
hipMemcpyHostToDevice,
stream_config.stream_id_));
float ave_time = 0;
......
include/ck/tensor_operation/gpu/device/impl/device_grouped_gemm_xdl_splitk_cshuffle.hpp
View file @ 65a317de
......@@ -406,10 +406,11 @@ struct DeviceGroupedGemmXdlSplitKCShuffle : public DeviceGroupedGemmSplitK<ALayo
}
}
hip_check_error(hipMemcpy(arg.p_workspace_,
hip_check_error(hipMemcpyWithStream(arg.p_workspace_,
arg.gemm_kernel_args_.data(),
arg.gemm_kernel_args_.size() * sizeof(GemmTransKernelArg),
hipMemcpyHostToDevice));
hipMemcpyHostToDevice,
stream_config.stream_id_));
float ave_time = 0;
......
include/ck/tensor_operation/gpu/grid/block_to_ctile_map.hpp
View file @ 65a317de
......@@ -134,6 +134,14 @@ struct BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, void>
{
}
template <typename CGridDesc_M_N>
__host__ __device__ BlockToCTileMap_M00_N0_M01Adapt(const CGridDesc_M_N& c_grid_desc_m_n,
index_t M01 = 8)
: BlockToCTileMap_M00_N0_M01Adapt(
c_grid_desc_m_n.GetLength(I0), c_grid_desc_m_n.GetLength(I1), M01)
{
}
__host__ static constexpr index_t CalculateGridSize(index_t M, index_t N)
{
const auto M0 = math::integer_divide_ceil(M, MPerBlock);
......@@ -142,6 +150,18 @@ struct BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, void>
return M0 * N0;
}
template <typename CGridDesc_M_N>
__host__ static constexpr index_t CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n)
{
return CalculateGridSize(c_grid_desc_m_n.GetLength(I0), c_grid_desc_m_n.GetLength(I1));
}
template <typename CGridDesc_M_N>
__host__ bool CheckValidity(const CGridDesc_M_N& /* c_grid_desc_m_n */) const
{
return true;
}
template <typename TopIdx>
__host__ __device__ constexpr auto CalculateBottomIndex(const TopIdx& idx_top) const
{
......@@ -222,30 +242,12 @@ struct BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, void>
index_t M01_;
};
// keep the redundant type argument for backward compatibility
template <index_t MPerBlock, index_t NPerBlock, typename CGridDesc_M_N>
struct BlockToCTileMap_M00_N0_M01Adapt : BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, void>
{
using Parent = BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, void>;
using Parent::I0;
using Parent::I1;
using Parent::Parent;
using Parent::operator=;
__host__ __device__ BlockToCTileMap_M00_N0_M01Adapt(const CGridDesc_M_N& c_grid_desc_m_n,
index_t M01 = 8)
: Parent(c_grid_desc_m_n.GetLength(I0), c_grid_desc_m_n.GetLength(I1), M01)
{
}
__host__ static constexpr index_t CalculateGridSize(const CGridDesc_M_N& c_grid_desc_m_n)
{
return Parent::CalculateGridSize(c_grid_desc_m_n.GetLength(I0),
c_grid_desc_m_n.GetLength(I1));
}
__host__ bool CheckValidity(const CGridDesc_M_N& /* c_grid_desc_m_n */) const { return true; }
using BlockToCTileMap_M00_N0_M01Adapt<MPerBlock, NPerBlock, void>::
BlockToCTileMap_M00_N0_M01Adapt;
};
// 2D slices of column-vectors in 3D space
......
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