Skip to content

GitLab

"deploy/hubserving/ocr_cls/module.py" did not exist on "f09d1f730b71a9b220cceec14a765c24cd9c1d6f"
Commit b134b7d6 authored by carlushuang's avatar carlushuang
Browse files

Merge remote-tracking branch 'origin/develop' into cpu_avx2

parents 090ba885 9f71ff48
Hide whitespace changes
Inline Side-by-side
Showing with 417 additions and 379 deletions
example/09_convnd_fwd/convnd_fwd_xdl_int8.cpp
View file @ b134b7d6
......@@ -5,7 +5,7 @@
#include "check_err.hpp"
#include "config.hpp"
#include "conv_fwd_util.hpp"
#include "conv_util.hpp"
#include "device.hpp"
#include "device_tensor.hpp"
#include "device_convnd_fwd_xdl_nhwc_kyxc_nhwk.hpp"
......@@ -45,10 +45,10 @@ template <ck::index_t NumDimSpatial>
using DeviceConvNDFwdInstance = ck::tensor_operation::device::
DeviceConvNDFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<
// clang-format off
InDataType, //
InDataType, //
WeiDataType, //
OutDataType, //
AccDataType, //
AccDataType, //
InElementOp, // Input Elementwise Operation
WeiElementOp, // Weights Elementwise Operation
OutElementOp, // Output Elementwise Operation
......@@ -112,7 +112,7 @@ void print_use_msg()
{
std::cout << "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
<< "arg3: run kernel # of times (>1)\n"
<< "arg3: time kernel (0=n0, 1=yes)\n"
<< "arg4: N spatial dimensions (default 2)\n"
<< "Following arguments (depending on number of spatial dims):\n"
<< " N, K, C, \n"
......@@ -139,40 +139,40 @@ ck::utils::conv::ConvParams parse_conv_params(int num_dim_spatial, int argc, cha
ck::utils::conv::ConvParams params;
int arg_idx = 5;
params.num_dim_spatial = num_dim_spatial;
params.N = std::stoi(argv[arg_idx++]);
params.K = std::stoi(argv[arg_idx++]);
params.C = std::stoi(argv[arg_idx++]);
params.num_dim_spatial_ = num_dim_spatial;
params.N_ = std::stoi(argv[arg_idx++]);
params.K_ = std::stoi(argv[arg_idx++]);
params.C_ = std::stoi(argv[arg_idx++]);
params.filter_spatial_lengths.resize(num_dim_spatial);
params.filter_spatial_lengths_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.filter_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
params.filter_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
}
params.input_spatial_lengths.resize(num_dim_spatial);
params.input_spatial_lengths_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
params.input_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
}
params.conv_filter_strides.resize(num_dim_spatial);
params.conv_filter_strides_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.conv_filter_strides[i] = std::stoi(argv[arg_idx++]);
params.conv_filter_strides_[i] = std::stoi(argv[arg_idx++]);
}
params.conv_filter_dilations.resize(num_dim_spatial);
params.conv_filter_dilations_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.conv_filter_dilations[i] = std::stoi(argv[arg_idx++]);
params.conv_filter_dilations_[i] = std::stoi(argv[arg_idx++]);
}
params.input_left_pads.resize(num_dim_spatial);
params.input_left_pads_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_left_pads[i] = std::stoi(argv[arg_idx++]);
params.input_left_pads_[i] = std::stoi(argv[arg_idx++]);
}
params.input_right_pads.resize(num_dim_spatial);
params.input_right_pads_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_right_pads[i] = std::stoi(argv[arg_idx++]);
params.input_right_pads_[i] = std::stoi(argv[arg_idx++]);
}
return params;
......@@ -184,9 +184,9 @@ int main(int argc, char* argv[])
{
using namespace ck::utils::conv;
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
int num_dim_spatial = 2;
ck::utils::conv::ConvParams params;
......@@ -195,7 +195,7 @@ int main(int argc, char* argv[])
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
num_dim_spatial = std::stoi(argv[4]);
}
......@@ -204,21 +204,21 @@ int main(int argc, char* argv[])
params = parse_conv_params(num_dim_spatial, argc, argv);
}
std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N),
static_cast<std::size_t>(params.C)};
std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N_),
static_cast<std::size_t>(params.C_)};
input_dims.insert(std::end(input_dims),
std::begin(params.input_spatial_lengths),
std::end(params.input_spatial_lengths));
std::begin(params.input_spatial_lengths_),
std::end(params.input_spatial_lengths_));
std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K),
static_cast<std::size_t>(params.C)};
std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K_),
static_cast<std::size_t>(params.C_)};
filter_dims.insert(std::end(filter_dims),
std::begin(params.filter_spatial_lengths),
std::end(params.filter_spatial_lengths));
std::begin(params.filter_spatial_lengths_),
std::end(params.filter_spatial_lengths_));
const std::vector<ck::index_t>& output_spatial_lengths = params.GetOutputSpatialLengths();
std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N),
static_cast<std::size_t>(params.K)};
std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N_),
static_cast<std::size_t>(params.K_)};
output_dims.insert(std::end(output_dims),
std::begin(output_spatial_lengths),
std::end(output_spatial_lengths));
......@@ -258,16 +258,16 @@ int main(int argc, char* argv[])
conv->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
params.N,
params.K,
params.C,
params.input_spatial_lengths,
params.filter_spatial_lengths,
params.N_,
params.K_,
params.C_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
output_spatial_lengths,
params.conv_filter_strides,
params.conv_filter_dilations,
params.input_left_pads,
params.input_right_pads,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
InElementOp{},
WeiElementOp{},
OutElementOp{});
......@@ -279,16 +279,16 @@ int main(int argc, char* argv[])
"not support this Conv problem");
}
float ave_time = invoker->Run(argument.get(), nrepeat);
float ave_time = invoker->Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = get_flops(
params.N, params.C, params.K, params.filter_spatial_lengths, output_spatial_lengths);
params.N_, params.C_, params.K_, params.filter_spatial_lengths_, output_spatial_lengths);
std::size_t num_btype = get_btype<InDataType, WeiDataType, OutDataType>(
params.N,
params.C,
params.K,
params.input_spatial_lengths,
params.filter_spatial_lengths,
params.N_,
params.C_,
params.K_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
output_spatial_lengths);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
......@@ -304,18 +304,18 @@ int main(int argc, char* argv[])
auto ref_argument = ref_conv.MakeArgument(input,
weights,
host_output,
params.conv_filter_strides,
params.conv_filter_dilations,
params.input_left_pads,
params.input_right_pads,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
InElementOp{},
WeiElementOp{},
OutElementOp{});
ref_invoker.Run(ref_argument);
out_device_buf.FromDevice(device_output.mData.data());
ck::utils::check_err(
host_output.mData, device_output.mData, "Error: incorrect results!", 1e-5f, 1e-4f);
return ck::utils::check_err(
host_output.mData, device_output.mData, "Error: incorrect results!", 1e-5f, 1e-4f) ? 0 : 1;
};
switch(num_dim_spatial)
......@@ -340,4 +340,5 @@ int main(int argc, char* argv[])
}
}
}
return 0;
}
example/10_conv2d_bwd_data/CMakeLists.txt
View file @ b134b7d6
add_example_executable(example_conv2d_bwd_data_xdl conv2d_bwd_data_xdl.cpp)
target_link_libraries(example_conv2d_bwd_data_xdl PRIVATE conv_fwd_util)
target_link_libraries(example_conv2d_bwd_data_xdl PRIVATE conv_util)
example/10_conv2d_bwd_data/conv2d_bwd_data_xdl.cpp
View file @ b134b7d6
......@@ -77,9 +77,9 @@ using ReferenceConvBwdInstance = ck::tensor_operation::host::ReferenceConvBwdDat
int main(int argc, char* argv[])
{
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// Conv shape
ck::index_t N = 128;
......@@ -102,13 +102,13 @@ int main(int argc, char* argv[])
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 19)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
N = std::stoi(argv[4]);
K = std::stoi(argv[5]);
......@@ -130,7 +130,7 @@ int main(int argc, char* argv[])
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: run kernel # of times (>1)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4 to 18: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
"RightPx\n");
exit(0);
......@@ -214,7 +214,7 @@ int main(int argc, char* argv[])
"not support this Conv problem");
}
float ave_time = invoker.Run(argument, nrepeat);
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
......@@ -249,6 +249,10 @@ int main(int argc, char* argv[])
in_device_buf.FromDevice(in_n_c_hi_wi_device_result.mData.data());
ck::utils::check_err(in_n_c_hi_wi_device_result.mData, in_n_c_hi_wi_host_result.mData);
return ck::utils::check_err(in_n_c_hi_wi_device_result.mData,
in_n_c_hi_wi_host_result.mData)
? 0
: 1;
}
return 0;
}
example/11_conv2d_bwd_weight/CMakeLists.txt
View file @ b134b7d6
add_example_executable(example_conv2d_bwd_weight_xdl conv2d_bwd_weight_xdl.cpp)
target_link_libraries(example_conv2d_bwd_weight_xdl PRIVATE conv_fwd_util)
target_link_libraries(example_conv2d_bwd_weight_xdl PRIVATE conv_util)
example/11_conv2d_bwd_weight/conv2d_bwd_weight_xdl.cpp
View file @ b134b7d6
......@@ -82,9 +82,9 @@ using ReferenceConvBwdWeightInstance =
int main(int argc, char* argv[])
{
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
int do_log = 0;
int split_k = 4;
......@@ -109,7 +109,7 @@ int main(int argc, char* argv[])
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
do_log = std::stoi(argv[4]);
split_k = std::stoi(argv[5]);
}
......@@ -117,7 +117,7 @@ int main(int argc, char* argv[])
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
do_log = std::stoi(argv[4]);
split_k = std::stoi(argv[5]);
......@@ -141,7 +141,7 @@ int main(int argc, char* argv[])
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: run kernel # of times (>1)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4: is show log (0=no, 1=yes)\n");
printf("arg5: split-k \n");
printf("arg6 to 19: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
......@@ -246,7 +246,7 @@ int main(int argc, char* argv[])
return 1;
}
float ave_time = invoker.Run(argument, nrepeat);
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
......@@ -291,6 +291,9 @@ int main(int argc, char* argv[])
LogRangeAsType<float>(std::cout << "wei_host : ", wei_k_c_y_x_host_result.mData, ",")
<< std::endl;
}
ck::utils::check_err(wei_k_c_y_x_device_result.mData, wei_k_c_y_x_host_result.mData);
return ck::utils::check_err(wei_k_c_y_x_device_result.mData, wei_k_c_y_x_host_result.mData)
? 0
: 1;
}
return 0;
}
example/12_reduce/CMakeLists.txt
View file @ b134b7d6
add_example_executable(example_reduce_blockwise reduce_blockwise.cpp)
add_example_executable(example_reduce_blockwise reduce_blockwise.cpp -D 16,64,32,960 -v 1 1 10)
example/12_reduce/reduce_blockwise.cpp
View file @ b134b7d6
......@@ -116,10 +116,9 @@ class SimpleAppArgs
std::vector<size_t> inLengths;
std::vector<float> scales;
bool do_verification = false;
int init_method = 1;
int nrepeat = 5;
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
public:
void show_usage(const char* cmd)
......@@ -135,12 +134,12 @@ class SimpleAppArgs
std::cout << "Arg1 -- init method (0=no init, 1=single integer value, 2=scope integer "
"value, 3=decimal value)"
<< std::endl;
std::cout << "Arg2 -- number of repeats to run the kernel" << std::endl;
std::cout << "Arg2 -- time kernel (0=n0, 1=yes)" << std::endl;
};
int processArgs(int argc, char* argv[])
{
unsigned int ch;
int ch;
while(1)
{
......@@ -182,7 +181,7 @@ class SimpleAppArgs
throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
init_method = std::atoi(argv[optind++]);
nrepeat = std::atoi(argv[optind]);
time_kernel = std::atoi(argv[optind]);
if(scales.empty())
{
......@@ -352,7 +351,7 @@ int main(int argc, char* argv[])
auto invoker_ptr = reduce.MakeInvokerPointer();
float avg_time = invoker_ptr->Run(argument_ptr.get(), args.nrepeat);
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, args.time_kernel});
std::size_t num_bytes = invariant_total_length * reduce_total_length * sizeof(InDataType) +
invariant_total_length * sizeof(OutDataType);
......@@ -362,16 +361,17 @@ int main(int argc, char* argv[])
std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, " << reduce_name
<< std::endl;
bool pass = true;
if(args.do_verification)
{
out_dev.FromDevice(out.mData.data());
ck::utils::check_err(out.mData, out_ref.mData);
pass &= ck::utils::check_err(out.mData, out_ref.mData);
if(NeedIndices)
{
out_indices_dev.FromDevice(out_indices.mData.data());
ck::utils::check_err(out_indices.mData, out_indices_ref.mData);
;
pass &= ck::utils::check_err(out_indices.mData, out_indices_ref.mData);
};
};
return pass ? 0 : 1;
}
example/13_pool2d_fwd/pool2d_fwd.cpp
View file @ b134b7d6
......@@ -80,8 +80,8 @@ static void pool_host_verify(const Tensor<InDataType>& in,
for(int x = 0; x < window_spatial_lengths[1]; ++x)
{
int wi = wo * window_strides[1] + x - in_left_pads[1];
if(hi >= 0 && hi < in.mDesc.GetLengths()[2] && wi >= 0 &&
wi < in.mDesc.GetLengths()[3])
if(hi >= 0 && hi < ck::type_convert<int>(in.mDesc.GetLengths()[2]) && wi >= 0 &&
wi < ck::type_convert<int>(in.mDesc.GetLengths()[3]))
{
AccDataType currVal = static_cast<AccDataType>(in(n, c, hi, wi));
......@@ -149,9 +149,9 @@ int main(int argc, char* argv[])
{
using namespace ck::host_reduce;
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// Pool shape
ck::index_t N = 128;
......@@ -171,13 +171,13 @@ int main(int argc, char* argv[])
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 16)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
N = std::stoi(argv[4]);
C = std::stoi(argv[5]);
......@@ -196,7 +196,7 @@ int main(int argc, char* argv[])
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: run kernel # of times (>1)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, LeftPy, LeftPx, RightPy, "
"RightPx\n");
exit(0);
......@@ -271,7 +271,7 @@ int main(int argc, char* argv[])
"not support this problem");
}
float ave_time = invoker_ptr->Run(argument_ptr.get(), nrepeat);
float ave_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * N * C * Ho * Wo * Y * X;
......@@ -285,6 +285,7 @@ int main(int argc, char* argv[])
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
bool pass = true;
if(do_verification)
{
pool_host_verify<InDataType,
......@@ -302,14 +303,15 @@ int main(int argc, char* argv[])
out_device_buf.FromDevice(out_n_c_ho_wo_device.mData.data());
ck::utils::check_err(out_n_c_ho_wo_device.mData, out_n_c_ho_wo_host.mData);
pass &= ck::utils::check_err(out_n_c_ho_wo_device.mData, out_n_c_ho_wo_host.mData);
if constexpr(NeedIndices)
{
out_indices_device_buf.FromDevice(out_indices_n_c_ho_wo_device.mData.data());
// ck::utils::check_err(out_indices_n_c_ho_wo_device.mData,
// out_indices_n_c_ho_wo_host.mData);;
pass &= ck::utils::check_err(out_indices_n_c_ho_wo_device.mData,
out_indices_n_c_ho_wo_host.mData);
};
}
return pass ? 0 : 1;
}
example/14_gemm_xdl_requant_relu_requant/gemm_xdl_requant_relu_requant_int8.cpp
View file @ b134b7d6
......@@ -13,74 +13,91 @@
#include "host_tensor_generator.hpp"
#include "host_gemm.hpp"
#include "device_tensor.hpp"
#include "device_gemm_xdl.hpp"
#include "device_gemm_xdl_c_shuffle.hpp"
#include "device_gemm_xdl_cshuffle.hpp"
#include "element_wise_operation.hpp"
#include "reference_gemm.hpp"
#include "gemm_specialization.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
struct RequantReluRequant
{
// FIXME: We just need one scale for Relu / Leaky Relu / PRelu
RequantReluRequant(float scaleGemm, float scaleRelu)
: scaleGemm_(scaleGemm), scaleRelu_(scaleRelu)
{
}
using F32 = float;
__host__ __device__ constexpr void operator()(float& y, const float& x) const
{
float gemm_requant = scaleGemm_ * x;
float relu = gemm_requant > 0 ? gemm_requant : 0;
float relu_requant = scaleRelu_ * relu;
y = relu_requant > 127 ? 127 : relu_requant < -128 ? -128 : relu_requant;
}
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
float scaleGemm_;
float scaleRelu_;
};
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using RequantReluRequant = ck::tensor_operation::element_wise::RequantReluRequant;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ADataType = int8_t;
using BDataType = int8_t;
using CDataType = int8_t;
using AccDataType = int32_t;
using CShuffleDataType = int32_t;
using CShuffleDataType = float;
using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
using CLayout = ck::tensor_layout::gemm::RowMajor;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl_C_Shuffle<
ADataType, // ADataType
BDataType, // BDataType
CDataType, // CDataType
AccDataType, // AccDataType
CShuffleDataType, // CShuffleDataType
ALayout, // ALayout
BLayout, // BLayout
CLayout, // CLayout
PassThrough, // AElementwiseOperation
PassThrough, // BElementwiseOperation
RequantReluRequant, // CElementwiseOperation
256, // BlockSize
256, // MPerBlock
128, // NPerBlock
64, // KPerBlock
16, // AK1
16, // BK1
32, // MPerXDL
32, // NPerXDL
4, // MXdlPerWave
2, // NXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
16, // ABlockTransferSrcScalarPerVector
16, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
16, // BBlockTransferSrcScalarPerVector
16, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 1, 64, 1, 1, 4>, // CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
16>; // CBlockTransferScalarPerVector_NWaveNPerXdl
using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemm_Xdl_CShuffle<
ALayout, // typename ALayout,
BLayout, // typename BLayout,
CLayout, // typename CLayout,
ADataType, // typename ADataType,
BDataType, // typename BDataType,
CDataType, // typename CDataType,
AccDataType, // typename GemmAccDataType,
CShuffleDataType, // typename CShuffleDataType,
PassThrough, // typename AElementwiseOperation,
PassThrough, // typename BElementwiseOperation,
RequantReluRequant, // typename CElementwiseOperation,
GemmDefault, // GemmSpecialization GemmSpec,
1, // index_t NumGemmKPrefetchStage,
256, // index_t BlockSize,
256, // index_t MPerBlock,
128, // index_t NPerBlock,
64, // index_t KPerBlock,
16, // index_t AK1,
16, // index_t BK1,
32, // index_t MPerXDL,
32, // index_t NPerXDL,
4, // index_t MXdlPerWave,
2, // index_t NXdlPerWave,
S<4, 64, 1>, // typename ABlockTransferThreadClusterLengths_AK0_M_AK1,
S<1, 0, 2>, // typename ABlockTransferThreadClusterArrangeOrder,
S<1, 0, 2>, // typename ABlockTransferSrcAccessOrder,
2, // index_t ABlockTransferSrcVectorDim,
16, // index_t ABlockTransferSrcScalarPerVector,
16, // index_t ABlockTransferDstScalarPerVector_AK1,
1, // bool ABlockLdsExtraM,
S<4, 64, 1>, // typename BBlockTransferThreadClusterLengths_BK0_N_BK1,
S<1, 0, 2>, // typename BBlockTransferThreadClusterArrangeOrder,
S<1, 0, 2>, // typename BBlockTransferSrcAccessOrder,
2, // index_t BBlockTransferSrcVectorDim,
8, // index_t BBlockTransferSrcScalarPerVector,
8, // index_t BBlockTransferDstScalarPerVector_BK1,
1, // bool BBlockLdsExtraN,
1, // index_t CShuffleMXdlPerWavePerShuffle,
1, // index_t CShuffleNXdlPerWavePerShuffle,
S<1, 64, 1, 4>, // typename CShuffleBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock,
16>; // index_t CShuffleBlockTransferScalarPerVector_NPerBlock>
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::
......@@ -88,9 +105,9 @@ using ReferenceGemmInstance = ck::tensor_operation::host::
int main(int argc, char* argv[])
{
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 3840;
......@@ -108,13 +125,13 @@ int main(int argc, char* argv[])
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 10)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
......@@ -128,7 +145,7 @@ int main(int argc, char* argv[])
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: run kernel # of times (>1)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
exit(0);
}
......@@ -202,7 +219,7 @@ int main(int argc, char* argv[])
"not support this GEMM problem");
}
float ave_time = invoker.Run(argument, nrepeat);
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
......@@ -227,7 +244,7 @@ int main(int argc, char* argv[])
ref_invoker.Run(ref_argument);
ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData);
return ck::utils::check_err(c_m_n_device_result.mData, c_m_n_host_result.mData) ? 0 : 1;
}
return 0;
......
example/15_grouped_gemm/grouped_gemm_xdl_fp16.cpp
View file @ b134b7d6
......@@ -60,21 +60,21 @@ using ReferenceGemmInstance = ck::tensor_operation::host::
int main(int argc, char* argv[])
{
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: run kernel # of times (>1)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n");
exit(0);
}
......@@ -131,7 +131,7 @@ int main(int argc, char* argv[])
std::size_t flop = 0, num_btype = 0;
for(int i = 0; i < gemm_shapes.size(); i++)
for(std::size_t i = 0; i < gemm_shapes.size(); i++)
{
a_tensors.push_back(Tensor<ADataType>(f_host_tensor_descriptor(
gemm_shapes[i].M, gemm_shapes[i].K, gemm_shapes[i].StrideA, ALayout{})));
......@@ -168,7 +168,7 @@ int main(int argc, char* argv[])
}
}
for(int i = 0; i < gemm_shapes.size(); i++)
for(std::size_t i = 0; i < gemm_shapes.size(); i++)
{
a_tensors_device.emplace_back(
std::make_unique<DeviceMem>(sizeof(ADataType) * a_tensors[i].mDesc.GetElementSpace()));
......@@ -202,7 +202,7 @@ int main(int argc, char* argv[])
"not support this GEMM problem");
}
float ave_time = invoker.Run(argument, nrepeat);
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
......@@ -211,9 +211,10 @@ int main(int argc, char* argv[])
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< gemm.GetTypeString() << std::endl;
bool pass = true;
if(do_verification)
{
for(int i = 0; i < gemm_shapes.size(); i++)
for(std::size_t i = 0; i < gemm_shapes.size(); i++)
{
c_tensors_device[i]->FromDevice(c_device_tensors[i].mData.data());
auto ref_gemm = ReferenceGemmInstance{};
......@@ -227,9 +228,9 @@ int main(int argc, char* argv[])
c_element_op);
ref_invoker.Run(ref_argument);
ck::utils::check_err(c_device_tensors[i].mData, c_host_tensors[i].mData);
pass &= ck::utils::check_err(c_device_tensors[i].mData, c_host_tensors[i].mData);
}
}
return 0;
return pass ? 0 : 1;
}
example/16_gemm_reduce/gemm_reduce_xdl_fp16.cpp
View file @ b134b7d6
......@@ -4,6 +4,7 @@
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
......@@ -11,9 +12,10 @@
#include "device_tensor.hpp"
#include "device_gemm_reduce_xdl_cshuffle.hpp"
#include "element_wise_operation.hpp"
#include "reduction_operator.hpp"
#include "reference_gemm.hpp"
#include "gemm_specialization.hpp"
#include "element_wise_reduce_operation.hpp"
#include "reduction_operator.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
......@@ -33,22 +35,23 @@ using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
using CLayout = ck::tensor_layout::gemm::RowMajor;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::tensor_operation::element_wise::ReduceSum;
using D1ReduceOp = ck::tensor_operation::element_wise::ReduceSquareSum;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::reduce::Add<float>;
using D1ReduceOp = ck::reduce::Add<float>;
using D1ElementOp = ck::tensor_operation::element_wise::UnarySquare<float, float, false>;
static constexpr auto GemmSpecialization =
ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceGemmReduceInstance = ck::tensor_operation::device::DeviceGemmReduce_Xdl_CShuffle
//######| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| D0| D1| 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| D0| D1| D1EleOp| 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| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| 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|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, AElementOp, BElementOp, CElementOp, D0ReduceOp, D1ReduceOp, GemmSpecialization, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>;
< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, AElementOp, BElementOp, CElementOp, D0ReduceOp, D1ReduceOp, D1ElementOp, GemmSpecialization, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>;
// clang-format on
using ReferenceGemmInstance = ck::tensor_operation::host::
......@@ -56,9 +59,9 @@ using ReferenceGemmInstance = ck::tensor_operation::host::
int main(int argc, char* argv[])
{
bool do_verification = 1;
bool do_verification = true;
int init_method = 1;
int nrepeat = 5;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 3840;
......@@ -77,13 +80,13 @@ int main(int argc, char* argv[])
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 10)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
......@@ -97,7 +100,7 @@ int main(int argc, char* argv[])
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: run kernel # of times (>1)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4 to 9: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC\n");
exit(0);
}
......@@ -159,11 +162,10 @@ int main(int argc, char* argv[])
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
auto d0_reduce_op = D0ReduceOp{};
auto d1_reduce_op = D1ReduceOp{};
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
auto d1_element_op = D1ElementOp{};
// do GEMM
auto gemm = DeviceGemmReduceInstance{};
......@@ -182,8 +184,7 @@ int main(int argc, char* argv[])
a_element_op,
b_element_op,
c_element_op,
d0_reduce_op,
d1_reduce_op);
d1_element_op);
if(!gemm.IsSupportedArgument(argument))
{
......@@ -192,30 +193,13 @@ int main(int argc, char* argv[])
"not support this GEMM problem");
}
// warm up
invoker.Run(argument);
// init DO, D1 to 0
d0_device_buf.SetZero();
d1_device_buf.SetZero();
// timing
float total_time = 0;
for(int i = 0; i < nrepeat; ++i)
{
// init DO, D1 to 0
d0_device_buf.SetZero();
d1_device_buf.SetZero();
KernelTimer timer;
timer.Start();
invoker.Run(argument);
timer.End();
total_time += timer.GetElapsedTime();
}
float ave_time = total_time / nrepeat;
// if time_kernel == true, kernel will run multiple times. This kernel use atomic-add so result
// will not be correct. need to set time_kernel = false for correctness test
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
......@@ -228,6 +212,7 @@ int main(int argc, char* argv[])
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< gemm.GetTypeString() << std::endl;
bool pass = true;
if(do_verification)
{
c_device_buf.FromDevice(c_m_n_device_result.mData.data());
......@@ -242,25 +227,41 @@ int main(int argc, char* argv[])
ref_invoker.Run(ref_argument);
auto d0_reduce_op = D0ReduceOp{};
auto d1_reduce_op = D1ReduceOp{};
for(int m = 0; m < M; ++m)
{
float d0_acc = d0_reduce_op.GetReduceZeroValue();
float d1_acc = d1_reduce_op.GetReduceZeroValue();
float d0_acc = d0_reduce_op.GetReductionZeroVal();
float d1_acc = d1_reduce_op.GetReductionZeroVal();
for(int n = 0; n < N; ++n)
{
d0_reduce_op.Reduce(d0_acc, c_m_n_host_result(m, n));
d1_reduce_op.Reduce(d1_acc, c_m_n_host_result(m, n));
float d0_val = ck::type_convert<float>(c_m_n_host_result(m, n));
float d1_val;
d1_element_op(d1_val, d0_val);
d0_reduce_op(d0_acc, d0_val);
d1_reduce_op(d1_acc, d1_val);
}
d0_m_host_result(m) = d0_acc;
d1_m_host_result(m) = d1_acc;
d0_m_host_result(m) = ck::type_convert<DDataType>(d0_acc);
d1_m_host_result(m) = ck::type_convert<DDataType>(d1_acc);
}
check_error(c_m_n_host_result, c_m_n_device_result);
check_error(d0_m_host_result, d0_m_device_result);
check_error(d1_m_host_result, d1_m_device_result);
pass &= ck::utils::check_err(
c_m_n_device_result.mData, c_m_n_host_result.mData, "Error: Incorrect results c");
pass &= ck::utils::check_err(d0_m_device_result.mData,
d0_m_host_result.mData,
"Error: Incorrect results d0",
1e-3,
1e-3);
pass &= ck::utils::check_err(d1_m_device_result.mData,
d1_m_host_result.mData,
"Error: Incorrect results d1",
1e-3,
1e-3);
}
return 0;
return pass ? 0 : 1;
}
example/17_convnd_bwd_data_xdl/CMakeLists.txt
View file @ b134b7d6
add_example_executable(example_convnd_bwd_data_xdl convnd_bwd_data_xdl.cpp)
target_link_libraries(example_convnd_bwd_data_xdl PRIVATE conv_fwd_util)
target_link_libraries(example_convnd_bwd_data_xdl PRIVATE conv_util)
example/17_convnd_bwd_data_xdl/convnd_bwd_data_xdl.cpp
View file @ b134b7d6
......@@ -6,7 +6,7 @@
#include <half.hpp>
#include "config.hpp"
#include "conv_fwd_util.hpp"
#include "conv_util.hpp"
#include "print.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
......@@ -87,7 +87,7 @@ void print_use_msg()
{
std::cout << "arg1: verification (0=no, 1=yes)\n"
<< "arg2: initialization (0=no init, 1=random value, 2= init to 1 )\n"
<< "arg3: run kernel # of times (>1)\n"
<< "arg3: time kernel (0=n0, 1=yes)\n"
<< "arg4: N spatial dimensions (default 2)\n"
<< "Following arguments (depending on number of spatial dims):\n"
<< " N, K, C, \n"
......@@ -105,40 +105,40 @@ ck::utils::conv::ConvParams parse_conv_params(int num_dim_spatial, char* argv[])
ck::utils::conv::ConvParams params;
int arg_idx = 5;
params.num_dim_spatial = num_dim_spatial;
params.N = std::stoi(argv[arg_idx++]);
params.K = std::stoi(argv[arg_idx++]);
params.C = std::stoi(argv[arg_idx++]);
params.num_dim_spatial_ = num_dim_spatial;
params.N_ = std::stoi(argv[arg_idx++]);
params.K_ = std::stoi(argv[arg_idx++]);
params.C_ = std::stoi(argv[arg_idx++]);
params.filter_spatial_lengths.resize(num_dim_spatial);
params.filter_spatial_lengths_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.filter_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
params.filter_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
}
params.input_spatial_lengths.resize(num_dim_spatial);
params.input_spatial_lengths_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_spatial_lengths[i] = std::stoi(argv[arg_idx++]);
params.input_spatial_lengths_[i] = std::stoi(argv[arg_idx++]);
}
params.conv_filter_strides.resize(num_dim_spatial);
params.conv_filter_strides_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.conv_filter_strides[i] = std::stoi(argv[arg_idx++]);
params.conv_filter_strides_[i] = std::stoi(argv[arg_idx++]);
}
params.conv_filter_dilations.resize(num_dim_spatial);
params.conv_filter_dilations_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.conv_filter_dilations[i] = std::stoi(argv[arg_idx++]);
params.conv_filter_dilations_[i] = std::stoi(argv[arg_idx++]);
}
params.input_left_pads.resize(num_dim_spatial);
params.input_left_pads_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_left_pads[i] = std::stoi(argv[arg_idx++]);
params.input_left_pads_[i] = std::stoi(argv[arg_idx++]);
}
params.input_right_pads.resize(num_dim_spatial);
params.input_right_pads_.resize(num_dim_spatial);
for(int i = 0; i < num_dim_spatial; ++i)
{
params.input_right_pads[i] = std::stoi(argv[arg_idx++]);
params.input_right_pads_[i] = std::stoi(argv[arg_idx++]);
}
return params;
......@@ -165,25 +165,25 @@ DeviceConvBwdDataBasePtr get_conv_instance(int num_dim_spatial)
int main(int argc, char* argv[])
{
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
int num_dim_spatial = 2;
ck::utils::conv::ConvParams params;
params.C = 128;
params.C_ = 128;
if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
}
else if(argc > 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
num_dim_spatial = std::stoi(argv[4]);
// check args number
int conv_args = 3 + num_dim_spatial * 6;
......@@ -202,21 +202,21 @@ int main(int argc, char* argv[])
exit(1);
}
std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N),
static_cast<std::size_t>(params.C)};
std::vector<std::size_t> input_dims{static_cast<std::size_t>(params.N_),
static_cast<std::size_t>(params.C_)};
input_dims.insert(std::end(input_dims),
std::begin(params.input_spatial_lengths),
std::end(params.input_spatial_lengths));
std::begin(params.input_spatial_lengths_),
std::end(params.input_spatial_lengths_));
std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K),
static_cast<std::size_t>(params.C)};
std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params.K_),
static_cast<std::size_t>(params.C_)};
filter_dims.insert(std::end(filter_dims),
std::begin(params.filter_spatial_lengths),
std::end(params.filter_spatial_lengths));
std::begin(params.filter_spatial_lengths_),
std::end(params.filter_spatial_lengths_));
const std::vector<ck::index_t>& output_spatial_lengths = params.GetOutputSpatialLengths();
std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N),
static_cast<std::size_t>(params.K)};
std::vector<std::size_t> output_dims{static_cast<std::size_t>(params.N_),
static_cast<std::size_t>(params.K_)};
output_dims.insert(std::end(output_dims),
std::begin(output_spatial_lengths),
std::end(output_spatial_lengths));
......@@ -263,16 +263,16 @@ int main(int argc, char* argv[])
conv->MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
params.N,
params.K,
params.C,
params.input_spatial_lengths,
params.filter_spatial_lengths,
params.N_,
params.K_,
params.C_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
output_spatial_lengths,
params.conv_filter_strides,
params.conv_filter_dilations,
params.input_left_pads,
params.input_right_pads,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
InElementOp{},
WeiElementOp{},
OutElementOp{});
......@@ -284,16 +284,16 @@ int main(int argc, char* argv[])
"not support this Conv problem");
}
float ave_time = invoker->Run(argument.get(), nrepeat);
float ave_time = invoker->Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = ck::utils::conv::get_flops(
params.N, params.C, params.K, params.filter_spatial_lengths, output_spatial_lengths);
params.N_, params.C_, params.K_, params.filter_spatial_lengths_, output_spatial_lengths);
std::size_t num_btype = ck::utils::conv::get_btype<InDataType, WeiDataType, OutDataType>(
params.N,
params.C,
params.K,
params.input_spatial_lengths,
params.filter_spatial_lengths,
params.N_,
params.C_,
params.K_,
params.input_spatial_lengths_,
params.filter_spatial_lengths_,
output_spatial_lengths);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
......@@ -310,10 +310,10 @@ int main(int argc, char* argv[])
auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi_host_result,
wei_k_c_y_x,
out_n_k_ho_wo,
params.conv_filter_strides,
params.conv_filter_dilations,
params.input_left_pads,
params.input_right_pads,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
InElementOp{},
WeiElementOp{},
OutElementOp{});
......@@ -322,7 +322,10 @@ int main(int argc, char* argv[])
in_device_buf.FromDevice(in_n_c_hi_wi_device_result.mData.data());
check_error(in_n_c_hi_wi_host_result, in_n_c_hi_wi_device_result);
return ck::utils::check_err(in_n_c_hi_wi_device_result.mData,
in_n_c_hi_wi_host_result.mData)
? 0
: 1;
};
switch(num_dim_spatial)
......@@ -347,4 +350,5 @@ int main(int argc, char* argv[])
}
}
}
return 0;
}
example/18_batched_gemm_reduce/batched_gemm_reduce_xdl_fp16.cpp
View file @ b134b7d6
......@@ -4,6 +4,7 @@
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
......@@ -11,9 +12,9 @@
#include "device_tensor.hpp"
#include "device_batched_gemm_reduce_xdl_cshuffle.hpp"
#include "element_wise_operation.hpp"
#include "reduction_operator.hpp"
#include "reference_batched_gemm.hpp"
#include "gemm_specialization.hpp"
#include "element_wise_reduce_operation.hpp"
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
......@@ -33,22 +34,23 @@ using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
using CLayout = ck::tensor_layout::gemm::RowMajor;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::tensor_operation::element_wise::ReduceSum;
using D1ReduceOp = ck::tensor_operation::element_wise::ReduceSquareSum;
using AElementOp = ck::tensor_operation::element_wise::PassThrough;
using BElementOp = ck::tensor_operation::element_wise::PassThrough;
using CElementOp = ck::tensor_operation::element_wise::PassThrough;
using D0ReduceOp = ck::reduce::Add<float>;
using D1ReduceOp = ck::reduce::Add<float>;
using D1ElementOp = ck::tensor_operation::element_wise::UnarySquare<float, float, false>;
static constexpr auto GemmSpecialization =
ck::tensor_operation::device::GemmSpecialization::Default;
// clang-format off
using DeviceBatchedGemmReduceInstance = ck::tensor_operation::device::DeviceBatchedGemmReduce_Xdl_CShuffle
//######| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| D0| D1| 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| D0| D1| D1EleOp| 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| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| ExtraN| 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|
//######| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, AElementOp, BElementOp, CElementOp, D0ReduceOp, D1ReduceOp, GemmSpecialization, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>;
< Row, Col, Row, F16, F16, F16, F32, F32, F32, F32, AElementOp, BElementOp, CElementOp, D0ReduceOp, D1ReduceOp, D1ElementOp, GemmSpecialization, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>;
// clang-format on
using ReferenceBatchedGemmInstance = ck::tensor_operation::host::
......@@ -56,18 +58,18 @@ using ReferenceBatchedGemmInstance = ck::tensor_operation::host::
int main(int argc, char* argv[])
{
bool do_verification = 1;
bool do_verification = true;
int init_method = 1;
int nrepeat = 5;
bool time_kernel = false;
// GEMM shape
ck::index_t M = 3840;
ck::index_t N = 4096;
ck::index_t K = 4096;
ck::index_t M = 2048;
ck::index_t N = 1920;
ck::index_t K = 2048;
ck::index_t StrideA = 4096;
ck::index_t StrideB = 4096;
ck::index_t StrideC = 4096;
ck::index_t StrideA = 2048;
ck::index_t StrideB = 2048;
ck::index_t StrideC = 1920;
ck::index_t BatchCount = 4;
......@@ -79,13 +81,13 @@ int main(int argc, char* argv[])
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
}
else if(argc == 11)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
time_kernel = std::stoi(argv[3]);
M = std::stoi(argv[4]);
N = std::stoi(argv[5]);
......@@ -95,13 +97,13 @@ int main(int argc, char* argv[])
StrideB = std::stoi(argv[8]);
StrideC = std::stoi(argv[9]);
BatchCount = std::stoi(argv[9]);
BatchCount = std::stoi(argv[10]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: run kernel # of times (>1)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg4 to 10: M (256x), N(128x), K(32x), StrideA, StrideB, StrideC, BatchCount\n");
exit(0);
}
......@@ -168,11 +170,12 @@ int main(int argc, char* argv[])
a_device_buf.ToDevice(a_g_m_k.mData.data());
b_device_buf.ToDevice(b_g_k_n.mData.data());
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
auto d0_reduce_op = D0ReduceOp{};
auto d1_reduce_op = D1ReduceOp{};
auto a_element_op = AElementOp{};
auto b_element_op = BElementOp{};
auto c_element_op = CElementOp{};
auto d0_reduce_op = D0ReduceOp{};
auto d1_reduce_op = D1ReduceOp{};
auto d1_element_op = D1ElementOp{};
// do GEMM
auto batched_gemm = DeviceBatchedGemmReduceInstance{};
......@@ -192,8 +195,7 @@ int main(int argc, char* argv[])
a_element_op,
b_element_op,
c_element_op,
d0_reduce_op,
d1_reduce_op,
d1_element_op,
BatchCount);
if(!batched_gemm.IsSupportedArgument(argument))
......@@ -203,30 +205,13 @@ int main(int argc, char* argv[])
"not support this GEMM problem");
}
// warm up
invoker.Run(argument);
// init DO, D1 to 0
d0_device_buf.SetZero();
d1_device_buf.SetZero();
// timing
float total_time = 0;
for(int i = 0; i < nrepeat; ++i)
{
// init DO, D1 to 0
d0_device_buf.SetZero();
d1_device_buf.SetZero();
KernelTimer timer;
timer.Start();
invoker.Run(argument);
timer.End();
total_time += timer.GetElapsedTime();
}
float ave_time = total_time / nrepeat;
// if time_kernel == true, kernel will run multiple times. This kernel use atomic-add so result
// will not be correct. need to set time_kernel = false for correctness test
float ave_time = invoker.Run(argument, StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * BatchCount * M * N * K;
std::size_t num_btype = sizeof(ADataType) * BatchCount * M * K +
......@@ -240,6 +225,7 @@ int main(int argc, char* argv[])
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s, "
<< batched_gemm.GetTypeString() << std::endl;
bool pass = true;
if(do_verification)
{
c_device_buf.FromDevice(c_g_m_n_device_result.mData.data());
......@@ -258,24 +244,36 @@ int main(int argc, char* argv[])
{
for(int m = 0; m < M; ++m)
{
float d0_acc = d0_reduce_op.GetReduceZeroValue();
float d1_acc = d1_reduce_op.GetReduceZeroValue();
float d0_acc = d0_reduce_op.GetReductionZeroVal();
float d1_acc = d1_reduce_op.GetReductionZeroVal();
for(int n = 0; n < N; ++n)
{
d0_reduce_op.Reduce(d0_acc, c_g_m_n_host_result(batch, m, n));
d1_reduce_op.Reduce(d1_acc, c_g_m_n_host_result(batch, m, n));
float d0_val = ck::type_convert<float>(c_g_m_n_host_result(batch, m, n));
float d1_val;
d1_element_op(d1_val, d0_val);
d0_reduce_op(d0_acc, d0_val);
d1_reduce_op(d1_acc, d1_val);
}
d0_g_m_host_result(batch, m) = d0_acc;
d1_g_m_host_result(batch, m) = d1_acc;
d0_g_m_host_result(batch, m) = ck::type_convert<DDataType>(d0_acc);
d1_g_m_host_result(batch, m) = ck::type_convert<DDataType>(d1_acc);
}
}
check_error(c_g_m_n_host_result, c_g_m_n_device_result);
check_error(d0_g_m_host_result, d0_g_m_device_result);
check_error(d1_g_m_host_result, d1_g_m_device_result);
pass &= ck::utils::check_err(c_g_m_n_host_result.mData, c_g_m_n_device_result.mData);
pass &= ck::utils::check_err(d0_g_m_device_result.mData,
d0_g_m_host_result.mData,
"Error: Incorrect results! D0",
1e-3,
1e-3);
pass &= ck::utils::check_err(d1_g_m_device_result.mData,
d1_g_m_host_result.mData,
"Error: Incorrect results! D1",
1e-3,
1e-3);
}
return 0;
return pass ? 0 : 1;
}
example/CMakeLists.txt
View file @ b134b7d6
......@@ -19,9 +19,18 @@ include_directories(BEFORE
add_custom_target(examples)
function(add_example_executable EXAMPLE_NAME)
function(add_example_executable EXAMPLE_NAME FILE_NAME)
message("adding example ${EXAMPLE_NAME}")
add_executable(${EXAMPLE_NAME} ${ARGN})
add_executable(${EXAMPLE_NAME} ${FILE_NAME})
target_link_libraries(${EXAMPLE_NAME} PRIVATE host_tensor)
add_test(NAME ${EXAMPLE_NAME} COMMAND $<TARGET_FILE:${EXAMPLE_NAME}> ${ARGN})
add_dependencies(examples ${EXAMPLE_NAME})
add_dependencies(check ${EXAMPLE_NAME})
endfunction(add_example_executable EXAMPLE_NAME)
function(add_example_executable_no_testing EXAMPLE_NAME FILE_NAME)
message("adding example ${EXAMPLE_NAME}")
add_executable(${EXAMPLE_NAME} ${FILE_NAME})
target_link_libraries(${EXAMPLE_NAME} PRIVATE host_tensor)
add_dependencies(examples ${EXAMPLE_NAME})
endfunction(add_example_executable EXAMPLE_NAME)
......
include/ck/config.hpp
View file @ b134b7d6
......@@ -26,17 +26,14 @@
#endif
#endif
// buffer resourse, wave size
// buffer resource
#ifndef __HIP_DEVICE_COMPILE__ // for host code
#define CK_BUFFER_RESOURCE_3RD_DWORD -1
#define CK_GPU_WAVE_SIZE -1
#elif defined(__gfx803__) || defined(__gfx900__) || defined(__gfx906__) || defined(__gfx908__) || \
defined(__gfx90a__) // for GPU code
#define CK_BUFFER_RESOURCE_3RD_DWORD 0x00020000
#define CK_GPU_WAVE_SIZE 64
#elif defined(__gfx1030__) // for GPU code
#define CK_BUFFER_RESOURCE_3RD_DWORD 0x31014000
#define CK_GPU_WAVE_SIZE 32
#endif
// FMA instruction
......@@ -112,6 +109,10 @@
// experimental feature: use __builtin_memcpy instead of union to do bit_cast
#define CK_EXPERIMENTAL_USE_MEMCPY_FOR_BIT_CAST 1
// experimental feature: optimize for inter-wave scheduling policy
#define CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING 0
#define CK_EXPERIMENTAL_INTER_WAVE_SCHEDULING_MAC_CLUSTERS 1
// hack: have underlying assumption that need to be satsified, otherwise it's a bug
// hack for forcing register to keep idx_diff_low_const in SGPR. idx_diff_low_const must be
// thread-invariant, otherwise it's a bug
......
include/ck/hip_version.hpp.in deleted 100644 → 0
View file @ 090ba885
#pragma once
// "_PACKAGE_" to avoid name contentions: the macros like
// HIP_VERSION_MAJOR are defined in HIP_VERSION.h.
// clang-format off
#define CK_HIP_PACKAGE_VERSION_MAJOR @CK_HIP_VERSION_MAJOR@
#define CK_HIP_PACKAGE_VERSION_MINOR @CK_HIP_VERSION_MINOR@
#define CK_HIP_PACKAGE_VERSION_PATCH @CK_HIP_VERSION_PATCH@
// clang-format on
#ifndef CK_HIP_PACKAGE_VERSION_MAJOR
#define CK_HIP_PACKAGE_VERSION_MAJOR 0
#endif
#ifndef CK_HIP_PACKAGE_VERSION_MINOR
#define CK_HIP_PACKAGE_VERSION_MINOR 0
#endif
#ifndef CK_HIP_PACKAGE_VERSION_PATCH
#define CK_HIP_PACKAGE_VERSION_PATCH 0
#endif
// 3 decimal digits for major and minor, 6 digits for patch number.
// Max number is 999,999,999999 == 0xE8,D4A5,0FFF that fits into 64-bit math.
#if CK_HIP_PACKAGE_VERSION_MAJOR > 999 || CK_HIP_PACKAGE_VERSION_MAJOR > 999 || \
CK_HIP_PACKAGE_VERSION_PATCH > 999999
#error "Too big HIP version number(s)"
#endif
#define CK_HIP_PACKAGE_VERSION_FLAT \
((CK_HIP_PACKAGE_VERSION_MAJOR * 1000ULL + CK_HIP_PACKAGE_VERSION_MINOR) * 1000000 + \
CK_HIP_PACKAGE_VERSION_PATCH)
include/ck/options.hpp.in 0 → 100644
View file @ b134b7d6
#pragma once
#cmakedefine01 CK_TIME_KERNEL
include/ck/stream_config.hpp 0 → 100644
View file @ b134b7d6
#pragma once
#include <hip/hip_runtime.h>
#include <hip/hip_fp16.h>
struct StreamConfig
{
hipStream_t stream_id_ = nullptr;
bool time_kernel_ = false;
};
include/ck/tensor_description/tensor_descriptor_helper.hpp
View file @ b134b7d6
#ifndef CK_TENSOR_DESCRIPTOR_HELPER_HPP
#define CK_TENSOR_DESCRIPTOR_HELPER_HPP
#pragma once
#include "common_header.hpp"
#include "tensor_descriptor.hpp"
#include "multi_index_transform_helper.hpp"
......@@ -35,6 +33,12 @@ __host__ __device__ constexpr auto calculate_element_space_size_impl(const Lengt
}
#endif
// Lengths..., Strides... could be:
// 1) index_t, which is known at run-time, or
// 2) Number<>, which is known at compile-time
// element_space_size could be:
// 1) long_index_t, or
// 2) LongNumber<>
template <typename... Lengths,
typename... Strides,
typename enable_if<sizeof...(Lengths) == sizeof...(Strides), bool>::type = false>
......@@ -68,10 +72,10 @@ __host__ __device__ constexpr auto make_naive_tensor_descriptor(const Tuple<Leng
}
};
const auto element_space_size = f(f, Number<0>{}, Number<1>{});
const auto element_space_size = f(f, Number<0>{}, LongNumber<1>{});
#else
const auto element_space_size =
calculate_element_space_size_impl(lengths, strides, Number<0>{}, Number<1>{});
calculate_element_space_size_impl(lengths, strides, Number<0>{}, LongNumber<1>{});
#endif
return TensorDescriptor<remove_cv_t<decltype(transforms)>,
......@@ -82,9 +86,12 @@ __host__ __device__ constexpr auto make_naive_tensor_descriptor(const Tuple<Leng
element_space_size};
}
// Lengths... can be:
// 1) index_t, which is known at run-time
// Lengths... could be:
// 1) index_t, which is known at run-time, or
// 2) Number<>, which is known at compile-time
// element_space_size could be:
// 1) long_index_t, or
// 2) LongNumber<>
template <typename... Lengths>
__host__ __device__ constexpr auto
make_naive_tensor_descriptor_packed(const Tuple<Lengths...>& lengths)
......@@ -100,7 +107,7 @@ make_naive_tensor_descriptor_packed(const Tuple<Lengths...>& lengths)
constexpr auto visible_dim_hidden_ids = typename arithmetic_sequence_gen<1, N + 1, 1>::type{};
const auto element_space_size = container_reduce(lengths, math::multiplies{}, Number<1>{});
const auto element_space_size = container_reduce(lengths, math::multiplies{}, LongNumber<1>{});
return TensorDescriptor<remove_cv_t<decltype(transforms)>,
remove_cv_t<decltype(low_dim_hidden_idss)>,
......@@ -110,6 +117,12 @@ make_naive_tensor_descriptor_packed(const Tuple<Lengths...>& lengths)
element_space_size};
}
// Lengths... could be:
// 1) index_t, which is known at run-time, or
// 2) Number<>, which is known at compile-time
// align could be:
// 1) index_t, or
// 2) Number<>
template <typename... Lengths, typename Align>
__host__ __device__ constexpr auto
make_naive_tensor_descriptor_aligned(const Tuple<Lengths...>& lengths, Align align)
......@@ -146,4 +159,3 @@ make_naive_tensor_descriptor_aligned(const Tuple<Lengths...>& lengths, Align ali
}
} // namespace ck
#endif
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