Skip to content

GitLab

Unverified Commit 24af0144 authored by Po Yen Chen's avatar Po Yen Chen Committed by GitHub
Browse files

Merge branch 'develop' into gemm_layernorm_welford

parents 961f5e9e b79bbbc2
Hide whitespace changes
Inline Side-by-side
Showing with 602 additions and 161 deletions
example/24_batched_gemm/batched_gemm_xdl_fp16.cpp
View file @ 24af0144
...@@ -6,7 +6,7 @@ ...@@ -6,7 +6,7 @@
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d_xdl.hpp" #include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
......
example/24_batched_gemm/batched_gemm_xdl_fp32.cpp
View file @ 24af0144
...@@ -6,7 +6,7 @@ ...@@ -6,7 +6,7 @@
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d_xdl.hpp" #include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
......
example/24_batched_gemm/batched_gemm_xdl_int4.cpp
View file @ 24af0144
...@@ -6,7 +6,7 @@ ...@@ -6,7 +6,7 @@
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d_xdl.hpp" #include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
......
example/24_batched_gemm/batched_gemm_xdl_int8.cpp
View file @ 24af0144
...@@ -6,7 +6,7 @@ ...@@ -6,7 +6,7 @@
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp" #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_gemm_multi_d_xdl.hpp" #include "ck/tensor_operation/gpu/device/impl/device_batched_gemm_multi_d_xdl.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
......
example/24_batched_gemm/run_batched_gemm_example.inc
View file @ 24af0144
...@@ -55,15 +55,15 @@ bool run_batched_gemm(const ProblemSize& problem_size, const ExecutionConfig& co ...@@ -55,15 +55,15 @@ bool run_batched_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
std::size_t stride, std::size_t stride,
std::size_t batch_stride, std::size_t batch_stride,
auto layout) { auto layout) {
using namespace ck::literals;
if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value) if(std::is_same<decltype(layout), ck::tensor_layout::gemm::RowMajor>::value)
{ {
return HostTensorDescriptor(std::vector<std::size_t>({batch_count_, row, col}), return HostTensorDescriptor({batch_count_, row, col}, {batch_stride, stride, 1_uz});
std::vector<std::size_t>({batch_stride, stride, 1}));
} }
else else
{ {
return HostTensorDescriptor(std::vector<std::size_t>({batch_count_, row, col}), return HostTensorDescriptor({batch_count_, row, col}, {batch_stride, 1_uz, stride});
std::vector<std::size_t>({batch_stride, 1, stride}));
} }
}; };
...@@ -174,11 +174,11 @@ bool run_batched_gemm(const ProblemSize& problem_size, const ExecutionConfig& co ...@@ -174,11 +174,11 @@ bool run_batched_gemm(const ProblemSize& problem_size, const ExecutionConfig& co
#ifdef BUILD_INT4_EXAMPLE #ifdef BUILD_INT4_EXAMPLE
const Tensor<EDataType> e_device_result_converted(e_g_m_n_device_result); const Tensor<EDataType> e_device_result_converted(e_g_m_n_device_result);
pass &= ck::utils::check_err(e_device_result_converted.mData, e_g_m_n_host_result.mData); pass &= ck::utils::check_err(e_device_result_converted, e_g_m_n_host_result);
#else #else
pass = ck::utils::check_err( pass = ck::utils::check_err(
e_g_m_n_device_result.mData, e_g_m_n_host_result.mData, "Error: Incorrect results c"); e_g_m_n_device_result, e_g_m_n_host_result, "Error: Incorrect results c");
#endif #endif
} }
......
example/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m2n3k1_xdl_fp16.cpp
View file @ 24af0144
...@@ -8,7 +8,7 @@ ...@@ -8,7 +8,7 @@
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_contraction_multiple_d_xdl_cshuffle.hpp" #include "ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
...@@ -246,21 +246,11 @@ int main(int argc, char* argv[]) ...@@ -246,21 +246,11 @@ int main(int argc, char* argv[])
exit(0); exit(0);
} }
Tensor<ADataType> a_gs_ms_ks( Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
std::vector<std::size_t>(a_gs_ms_ks_lengths.begin(), a_gs_ms_ks_lengths.end()), Tensor<BDataType> b_gs_ns_ks(b_gs_ns_ks_lengths, b_gs_ns_ks_strides);
std::vector<std::size_t>(a_gs_ms_ks_strides.begin(), a_gs_ms_ks_strides.end())); Tensor<DDataType> d_gs_ms_ns(d_gs_ms_ns_lengths, d_gs_ms_ns_strides);
Tensor<BDataType> b_gs_ns_ks( Tensor<EDataType> e_gs_ms_ns_host_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
std::vector<std::size_t>(b_gs_ns_ks_lengths.begin(), b_gs_ns_ks_lengths.end()), Tensor<EDataType> e_gs_ms_ns_device_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
std::vector<std::size_t>(b_gs_ns_ks_strides.begin(), b_gs_ns_ks_strides.end()));
Tensor<DDataType> d_gs_ms_ns(
std::vector<std::size_t>(d_gs_ms_ns_lengths.begin(), d_gs_ms_ns_lengths.end()),
std::vector<std::size_t>(d_gs_ms_ns_strides.begin(), d_gs_ms_ns_strides.end()));
Tensor<EDataType> e_gs_ms_ns_host_result(
std::vector<std::size_t>(e_gs_ms_ns_lengths.begin(), e_gs_ms_ns_lengths.end()),
std::vector<std::size_t>(e_gs_ms_ns_strides.begin(), e_gs_ms_ns_strides.end()));
Tensor<EDataType> e_gs_ms_ns_device_result(
std::vector<std::size_t>(e_gs_ms_ns_lengths.begin(), e_gs_ms_ns_lengths.end()),
std::vector<std::size_t>(e_gs_ms_ns_strides.begin(), e_gs_ms_ns_strides.end()));
std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl; std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
std::cout << "b_gs_ns_ks: " << b_gs_ns_ks.mDesc << std::endl; std::cout << "b_gs_ns_ks: " << b_gs_ns_ks.mDesc << std::endl;
...@@ -357,9 +347,7 @@ int main(int argc, char* argv[]) ...@@ -357,9 +347,7 @@ int main(int argc, char* argv[])
if(do_verification) if(do_verification)
{ {
Tensor<CShuffleDataType> c_gs_ms_ns_host_result( Tensor<CShuffleDataType> c_gs_ms_ns_host_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
std::vector<std::size_t>(e_gs_ms_ns_lengths.begin(), e_gs_ms_ns_lengths.end()),
std::vector<std::size_t>(e_gs_ms_ns_strides.begin(), e_gs_ms_ns_strides.end()));
using ReferenceOpInstance = ReferenceContraction_G1_M2_N3_K1<NumDimM, using ReferenceOpInstance = ReferenceContraction_G1_M2_N3_K1<NumDimM,
NumDimN, NumDimN,
...@@ -407,9 +395,7 @@ int main(int argc, char* argv[]) ...@@ -407,9 +395,7 @@ int main(int argc, char* argv[])
} }
} }
return ck::utils::check_err(e_gs_ms_ns_device_result.mData, e_gs_ms_ns_host_result.mData) return ck::utils::check_err(e_gs_ms_ns_device_result, e_gs_ms_ns_host_result) ? 0 : 1;
? 0
: 1;
} }
return 0; return 0;
......
example/25_gemm_bias_e_permute/gemm_bias_e_permute_g1m3n2k1_xdl_fp16.cpp
View file @ 24af0144
...@@ -8,7 +8,7 @@ ...@@ -8,7 +8,7 @@
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_contraction_multiple_d_xdl_cshuffle.hpp" #include "ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
...@@ -246,21 +246,11 @@ int main(int argc, char* argv[]) ...@@ -246,21 +246,11 @@ int main(int argc, char* argv[])
exit(0); exit(0);
} }
Tensor<ADataType> a_gs_ms_ks( Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
std::vector<std::size_t>(a_gs_ms_ks_lengths.begin(), a_gs_ms_ks_lengths.end()), Tensor<BDataType> b_gs_ns_ks(b_gs_ns_ks_lengths, b_gs_ns_ks_strides);
std::vector<std::size_t>(a_gs_ms_ks_strides.begin(), a_gs_ms_ks_strides.end())); Tensor<DDataType> d_gs_ms_ns(d_gs_ms_ns_lengths, d_gs_ms_ns_strides);
Tensor<BDataType> b_gs_ns_ks( Tensor<EDataType> e_gs_ms_ns_host_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
std::vector<std::size_t>(b_gs_ns_ks_lengths.begin(), b_gs_ns_ks_lengths.end()), Tensor<EDataType> e_gs_ms_ns_device_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
std::vector<std::size_t>(b_gs_ns_ks_strides.begin(), b_gs_ns_ks_strides.end()));
Tensor<DDataType> d_gs_ms_ns(
std::vector<std::size_t>(d_gs_ms_ns_lengths.begin(), d_gs_ms_ns_lengths.end()),
std::vector<std::size_t>(d_gs_ms_ns_strides.begin(), d_gs_ms_ns_strides.end()));
Tensor<EDataType> e_gs_ms_ns_host_result(
std::vector<std::size_t>(e_gs_ms_ns_lengths.begin(), e_gs_ms_ns_lengths.end()),
std::vector<std::size_t>(e_gs_ms_ns_strides.begin(), e_gs_ms_ns_strides.end()));
Tensor<EDataType> e_gs_ms_ns_device_result(
std::vector<std::size_t>(e_gs_ms_ns_lengths.begin(), e_gs_ms_ns_lengths.end()),
std::vector<std::size_t>(e_gs_ms_ns_strides.begin(), e_gs_ms_ns_strides.end()));
std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl; std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
std::cout << "b_gs_ns_ks: " << b_gs_ns_ks.mDesc << std::endl; std::cout << "b_gs_ns_ks: " << b_gs_ns_ks.mDesc << std::endl;
...@@ -357,9 +347,7 @@ int main(int argc, char* argv[]) ...@@ -357,9 +347,7 @@ int main(int argc, char* argv[])
if(do_verification) if(do_verification)
{ {
Tensor<CShuffleDataType> c_gs_ms_ns_host_result( Tensor<CShuffleDataType> c_gs_ms_ns_host_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
std::vector<std::size_t>(e_gs_ms_ns_lengths.begin(), e_gs_ms_ns_lengths.end()),
std::vector<std::size_t>(e_gs_ms_ns_strides.begin(), e_gs_ms_ns_strides.end()));
using ReferenceOpInstance = ReferenceContraction_G1_M3_N2_K1<NumDimG, using ReferenceOpInstance = ReferenceContraction_G1_M3_N2_K1<NumDimG,
NumDimM, NumDimM,
...@@ -408,9 +396,7 @@ int main(int argc, char* argv[]) ...@@ -408,9 +396,7 @@ int main(int argc, char* argv[])
} }
} }
return ck::utils::check_err(e_gs_ms_ns_device_result.mData, e_gs_ms_ns_host_result.mData) return ck::utils::check_err(e_gs_ms_ns_device_result, e_gs_ms_ns_host_result) ? 0 : 1;
? 0
: 1;
} }
return 0; return 0;
......
example/26_contraction/contraction_bilinear_xdl_fp32.cpp
View file @ 24af0144
...@@ -8,7 +8,7 @@ ...@@ -8,7 +8,7 @@
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_contraction_multiple_d_xdl_cshuffle.hpp" #include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
...@@ -288,21 +288,11 @@ int main(int argc, char* argv[]) ...@@ -288,21 +288,11 @@ int main(int argc, char* argv[])
exit(0); exit(0);
} }
Tensor<ADataType> a_ms_ks( Tensor<ADataType> a_ms_ks(a_ms_ks_lengths, a_ms_ks_strides);
std::vector<std::size_t>(a_ms_ks_lengths.begin(), a_ms_ks_lengths.end()), Tensor<BDataType> b_ns_ks(b_ns_ks_lengths, b_ns_ks_strides);
std::vector<std::size_t>(a_ms_ks_strides.begin(), a_ms_ks_strides.end())); Tensor<EDataType> d_ms_ns(d_ms_ns_lengths, d_ms_ns_strides);
Tensor<BDataType> b_ns_ks( Tensor<EDataType> e_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
std::vector<std::size_t>(b_ns_ks_lengths.begin(), b_ns_ks_lengths.end()), Tensor<EDataType> e_ms_ns_device_result(e_ms_ns_lengths, e_ms_ns_strides);
std::vector<std::size_t>(b_ns_ks_strides.begin(), b_ns_ks_strides.end()));
Tensor<EDataType> d_ms_ns(
std::vector<std::size_t>(d_ms_ns_lengths.begin(), d_ms_ns_lengths.end()),
std::vector<std::size_t>(d_ms_ns_strides.begin(), d_ms_ns_strides.end()));
Tensor<EDataType> e_ms_ns_host_result(
std::vector<std::size_t>(e_ms_ns_lengths.begin(), e_ms_ns_lengths.end()),
std::vector<std::size_t>(e_ms_ns_strides.begin(), e_ms_ns_strides.end()));
Tensor<EDataType> e_ms_ns_device_result(
std::vector<std::size_t>(e_ms_ns_lengths.begin(), e_ms_ns_lengths.end()),
std::vector<std::size_t>(e_ms_ns_strides.begin(), e_ms_ns_strides.end()));
std::cout << "a_ms_ks: " << a_ms_ks.mDesc << std::endl; std::cout << "a_ms_ks: " << a_ms_ks.mDesc << std::endl;
std::cout << "b_ns_ks: " << b_ns_ks.mDesc << std::endl; std::cout << "b_ns_ks: " << b_ns_ks.mDesc << std::endl;
...@@ -398,9 +388,7 @@ int main(int argc, char* argv[]) ...@@ -398,9 +388,7 @@ int main(int argc, char* argv[])
if(do_verification) if(do_verification)
{ {
Tensor<CShuffleDataType> c_ms_ns_host_result( Tensor<CShuffleDataType> c_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
std::vector<std::size_t>(e_ms_ns_lengths.begin(), e_ms_ns_lengths.end()),
std::vector<std::size_t>(e_ms_ns_strides.begin(), e_ms_ns_strides.end()));
using ReferenceOpInstance = ReferenceContraction_M2_N2_K2<NumDimM, using ReferenceOpInstance = ReferenceContraction_M2_N2_K2<NumDimM,
NumDimN, NumDimN,
...@@ -437,7 +425,7 @@ int main(int argc, char* argv[]) ...@@ -437,7 +425,7 @@ int main(int argc, char* argv[])
} }
} }
return ck::utils::check_err(e_ms_ns_device_result.mData, e_ms_ns_host_result.mData) ? 0 : 1; return ck::utils::check_err(e_ms_ns_device_result, e_ms_ns_host_result) ? 0 : 1;
} }
return 0; return 0;
......
example/26_contraction/contraction_scale_xdl_fp32.cpp
View file @ 24af0144
...@@ -8,7 +8,7 @@ ...@@ -8,7 +8,7 @@
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_contraction_multiple_d_xdl_cshuffle.hpp" #include "ck/tensor_operation/gpu/device/impl/device_contraction_multiple_d_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
...@@ -277,18 +277,10 @@ int main(int argc, char* argv[]) ...@@ -277,18 +277,10 @@ int main(int argc, char* argv[])
exit(0); exit(0);
} }
Tensor<ADataType> a_ms_ks( Tensor<ADataType> a_ms_ks(a_ms_ks_lengths, a_ms_ks_strides);
std::vector<std::size_t>(a_ms_ks_lengths.begin(), a_ms_ks_lengths.end()), Tensor<BDataType> b_ns_ks(b_ns_ks_lengths, b_ns_ks_strides);
std::vector<std::size_t>(a_ms_ks_strides.begin(), a_ms_ks_strides.end())); Tensor<EDataType> e_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
Tensor<BDataType> b_ns_ks( Tensor<EDataType> e_ms_ns_device_result(e_ms_ns_lengths, e_ms_ns_strides);
std::vector<std::size_t>(b_ns_ks_lengths.begin(), b_ns_ks_lengths.end()),
std::vector<std::size_t>(b_ns_ks_strides.begin(), b_ns_ks_strides.end()));
Tensor<EDataType> e_ms_ns_host_result(
std::vector<std::size_t>(e_ms_ns_lengths.begin(), e_ms_ns_lengths.end()),
std::vector<std::size_t>(e_ms_ns_strides.begin(), e_ms_ns_strides.end()));
Tensor<EDataType> e_ms_ns_device_result(
std::vector<std::size_t>(e_ms_ns_lengths.begin(), e_ms_ns_lengths.end()),
std::vector<std::size_t>(e_ms_ns_strides.begin(), e_ms_ns_strides.end()));
std::cout << "a_ms_ks: " << a_ms_ks.mDesc << std::endl; std::cout << "a_ms_ks: " << a_ms_ks.mDesc << std::endl;
std::cout << "b_ns_ks: " << b_ns_ks.mDesc << std::endl; std::cout << "b_ns_ks: " << b_ns_ks.mDesc << std::endl;
...@@ -379,9 +371,7 @@ int main(int argc, char* argv[]) ...@@ -379,9 +371,7 @@ int main(int argc, char* argv[])
if(do_verification) if(do_verification)
{ {
Tensor<CShuffleDataType> c_ms_ns_host_result( Tensor<CShuffleDataType> c_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
std::vector<std::size_t>(e_ms_ns_lengths.begin(), e_ms_ns_lengths.end()),
std::vector<std::size_t>(e_ms_ns_strides.begin(), e_ms_ns_strides.end()));
using ReferenceOpInstance = ReferenceContraction_M2_N2_K2<NumDimM, using ReferenceOpInstance = ReferenceContraction_M2_N2_K2<NumDimM,
NumDimN, NumDimN,
...@@ -417,7 +407,7 @@ int main(int argc, char* argv[]) ...@@ -417,7 +407,7 @@ int main(int argc, char* argv[])
} }
} }
return ck::utils::check_err(e_ms_ns_device_result.mData, e_ms_ns_host_result.mData) ? 0 : 1; return ck::utils::check_err(e_ms_ns_device_result, e_ms_ns_host_result) ? 0 : 1;
} }
return 0; return 0;
......
example/27_layernorm/CMakeLists.txt
View file @ 24af0144
add_example_executable(example_layernorm_blockwise layernorm_blockwise.cpp) add_example_executable(example_layernorm_blockwise layernorm_blockwise.cpp)
\ No newline at end of file
example/27_layernorm/layernorm_blockwise.cpp
View file @ 24af0144
...@@ -9,7 +9,7 @@ ...@@ -9,7 +9,7 @@
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/utility/reduction_enums.hpp" #include "ck/utility/reduction_enums.hpp"
#include "ck/tensor_operation/gpu/device/device_layernorm_impl.hpp" #include "ck/tensor_operation/gpu/device/impl/device_normalization_impl.hpp"
#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp" #include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
...@@ -17,6 +17,7 @@ ...@@ -17,6 +17,7 @@
#include "ck/library/utility/host_common_util.hpp" #include "ck/library/utility/host_common_util.hpp"
#include "ck/library/utility/host_tensor.hpp" #include "ck/library/utility/host_tensor.hpp"
#include "ck/library/utility/host_tensor_generator.hpp" #include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/literals.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp" #include "ck/library/reference_tensor_operation/cpu/reference_layernorm.hpp"
using XDataType = ck::half_t; using XDataType = ck::half_t;
...@@ -30,26 +31,26 @@ constexpr int Rank = 2; ...@@ -30,26 +31,26 @@ constexpr int Rank = 2;
constexpr int NumReduceDim = 1; constexpr int NumReduceDim = 1;
using DeviceInstance = using DeviceInstance =
ck::tensor_operation::device::DeviceLayernormImpl<XDataType, ck::tensor_operation::device::DeviceNormalizationImpl<XDataType,
GammaDataType, GammaDataType,
BetaDataType, BetaDataType,
AccDataType, AccDataType,
YDataType, YDataType,
PassThrough, PassThrough,
Rank, Rank,
NumReduceDim, NumReduceDim,
256, // BlockSize 256, // BlockSize
8, // ClusterM 8, // ClusterM
32, // ClusterK 32, // ClusterK
1, // SliceM 1, // SliceM
8, // SliceK 8, // SliceK
1, // SrcVecDim (0=M, 1=K) 1, // SrcVecDim (0=M, 1=K)
8, // SrcScalarPerVector 8, // SrcScalarPerVector
1, // GammaVecDim (0=M, 1=K) 1, // GammaVecDim (0=M, 1=K)
8, // GammaScalarPerVector 8, // GammaScalarPerVector
1, // BetaVecDim (0=M, 1=K) 1, // BetaVecDim (0=M, 1=K)
8, // BetaScalarPerVector 8, // BetaScalarPerVector
8>; // OutScalarPerVector 8>; // OutScalarPerVector
int main() int main()
{ {
...@@ -60,13 +61,13 @@ int main() ...@@ -60,13 +61,13 @@ int main()
ck::index_t Stride = N; ck::index_t Stride = N;
auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) { auto f_host_tensor_descriptor1d = [](std::size_t len, std::size_t stride) {
return HostTensorDescriptor(std::vector<std::size_t>({len}), return HostTensorDescriptor({len}, {stride});
std::vector<std::size_t>({stride}));
}; };
auto f_host_tensor_descriptor2d = [](std::size_t row, std::size_t col, std::size_t stride) { auto f_host_tensor_descriptor2d = [](std::size_t row, std::size_t col, std::size_t stride) {
return HostTensorDescriptor(std::vector<std::size_t>({row, col}), using namespace ck::literals;
std::vector<std::size_t>({stride, 1}));
return HostTensorDescriptor({row, col}, {stride, 1_uz});
}; };
Tensor<XDataType> x(f_host_tensor_descriptor2d(M, N, Stride)); Tensor<XDataType> x(f_host_tensor_descriptor2d(M, N, Stride));
...@@ -100,6 +101,8 @@ int main() ...@@ -100,6 +101,8 @@ int main()
gamma_dev.GetDeviceBuffer(), gamma_dev.GetDeviceBuffer(),
beta_dev.GetDeviceBuffer(), beta_dev.GetDeviceBuffer(),
y_dev.GetDeviceBuffer(), y_dev.GetDeviceBuffer(),
nullptr,
nullptr,
PassThrough{}); PassThrough{});
if(!device_instance.IsSupportedArgument(argument_ptr.get())) if(!device_instance.IsSupportedArgument(argument_ptr.get()))
...@@ -130,8 +133,7 @@ int main() ...@@ -130,8 +133,7 @@ int main()
ref_invoker.Run(ref_argument); ref_invoker.Run(ref_argument);
y_dev.FromDevice(y.mData.data()); y_dev.FromDevice(y.mData.data());
pass &= pass &= ck::utils::check_err(y, host_y, "Error: Incorrect results d1", 1e-3, 1e-3);
ck::utils::check_err(y.mData, host_y.mData, "Error: Incorrect results d1", 1e-3, 1e-3);
} }
return (pass ? 0 : 1); return (pass ? 0 : 1);
} }
example/28_grouped_gemm_bias_e_permute/grouped_gemm_bias_e_permute_xdl_fp16.cpp
View file @ 24af0144
...@@ -9,7 +9,7 @@ ...@@ -9,7 +9,7 @@
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/tensor_specialization.hpp" #include "ck/tensor_operation/gpu/device/tensor_specialization.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_grouped_contraction_multiple_d_xdl_cshuffle.hpp" #include "ck/tensor_operation/gpu/device/impl/device_grouped_contraction_multiple_d_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
...@@ -297,18 +297,10 @@ int main(int argc, char* argv[]) ...@@ -297,18 +297,10 @@ int main(int argc, char* argv[])
const auto e_ms_ns_lengths = contraction_descs[i].e_ms_ns_lengths; const auto e_ms_ns_lengths = contraction_descs[i].e_ms_ns_lengths;
const auto e_ms_ns_strides = contraction_descs[i].e_ms_ns_strides; const auto e_ms_ns_strides = contraction_descs[i].e_ms_ns_strides;
Tensor<ADataType> a_ms_ks( Tensor<ADataType> a_ms_ks(a_ms_ks_lengths, a_ms_ks_strides);
std::vector<std::size_t>(a_ms_ks_lengths.begin(), a_ms_ks_lengths.end()), Tensor<BDataType> b_ns_ks(b_ns_ks_lengths, b_ns_ks_strides);
std::vector<std::size_t>(a_ms_ks_strides.begin(), a_ms_ks_strides.end())); Tensor<DDataType> d_ms_ns(d_ms_ns_lengths, d_ms_ns_strides);
Tensor<BDataType> b_ns_ks( Tensor<EDataType> e_ms_ns_device_result(e_ms_ns_lengths, e_ms_ns_strides);
std::vector<std::size_t>(b_ns_ks_lengths.begin(), b_ns_ks_lengths.end()),
std::vector<std::size_t>(b_ns_ks_strides.begin(), b_ns_ks_strides.end()));
Tensor<DDataType> d_ms_ns(
std::vector<std::size_t>(d_ms_ns_lengths.begin(), d_ms_ns_lengths.end()),
std::vector<std::size_t>(d_ms_ns_strides.begin(), d_ms_ns_strides.end()));
Tensor<EDataType> e_ms_ns_device_result(
std::vector<std::size_t>(e_ms_ns_lengths.begin(), e_ms_ns_lengths.end()),
std::vector<std::size_t>(e_ms_ns_strides.begin(), e_ms_ns_strides.end()));
ck::index_t M_ = std::accumulate(e_ms_ns_lengths.begin(), ck::index_t M_ = std::accumulate(e_ms_ns_lengths.begin(),
e_ms_ns_lengths.begin() + NumDimM, e_ms_ns_lengths.begin() + NumDimM,
...@@ -423,13 +415,9 @@ int main(int argc, char* argv[]) ...@@ -423,13 +415,9 @@ int main(int argc, char* argv[])
const auto e_ms_ns_lengths = contraction_descs[i].e_ms_ns_lengths; const auto e_ms_ns_lengths = contraction_descs[i].e_ms_ns_lengths;
const auto e_ms_ns_strides = contraction_descs[i].e_ms_ns_strides; const auto e_ms_ns_strides = contraction_descs[i].e_ms_ns_strides;
Tensor<EDataType> c_ms_ns_host_result( Tensor<EDataType> c_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
std::vector<std::size_t>(e_ms_ns_lengths.begin(), e_ms_ns_lengths.end()),
std::vector<std::size_t>(e_ms_ns_strides.begin(), e_ms_ns_strides.end()));
Tensor<EDataType> e_ms_ns_host_result( Tensor<EDataType> e_ms_ns_host_result(e_ms_ns_lengths, e_ms_ns_strides);
std::vector<std::size_t>(e_ms_ns_lengths.begin(), e_ms_ns_lengths.end()),
std::vector<std::size_t>(e_ms_ns_strides.begin(), e_ms_ns_strides.end()));
e_tensors_device[i]->FromDevice(e_device_tensors[i].mData.data()); e_tensors_device[i]->FromDevice(e_device_tensors[i].mData.data());
...@@ -475,7 +463,7 @@ int main(int argc, char* argv[]) ...@@ -475,7 +463,7 @@ int main(int argc, char* argv[])
} }
} }
pass &= ck::utils::check_err(e_device_tensors[i].mData, e_ms_ns_host_result.mData); pass &= ck::utils::check_err(e_device_tensors[i], e_ms_ns_host_result);
} }
} }
......
example/29_batched_gemm_bias_e_permute/batched_gemm_bias_e_permute_xdl_fp16.cpp
View file @ 24af0144
...@@ -8,7 +8,7 @@ ...@@ -8,7 +8,7 @@
#include "ck/ck.hpp" #include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/gemm_specialization.hpp" #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
#include "ck/tensor_operation/gpu/device/device_batched_contraction_multiple_d_xdl_cshuffle.hpp" #include "ck/tensor_operation/gpu/device/impl/device_batched_contraction_multiple_d_xdl_cshuffle.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp" #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
...@@ -246,21 +246,11 @@ int main(int argc, char* argv[]) ...@@ -246,21 +246,11 @@ int main(int argc, char* argv[])
exit(0); exit(0);
} }
Tensor<ADataType> a_gs_ms_ks( Tensor<ADataType> a_gs_ms_ks(a_gs_ms_ks_lengths, a_gs_ms_ks_strides);
std::vector<std::size_t>(a_gs_ms_ks_lengths.begin(), a_gs_ms_ks_lengths.end()), Tensor<BDataType> b_gs_ns_ks(b_gs_ns_ks_lengths, b_gs_ns_ks_strides);
std::vector<std::size_t>(a_gs_ms_ks_strides.begin(), a_gs_ms_ks_strides.end())); Tensor<DDataType> d_gs_ms_ns(d_gs_ms_ns_lengths, d_gs_ms_ns_strides);
Tensor<BDataType> b_gs_ns_ks( Tensor<EDataType> e_gs_ms_ns_host_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
std::vector<std::size_t>(b_gs_ns_ks_lengths.begin(), b_gs_ns_ks_lengths.end()), Tensor<EDataType> e_gs_ms_ns_device_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
std::vector<std::size_t>(b_gs_ns_ks_strides.begin(), b_gs_ns_ks_strides.end()));
Tensor<DDataType> d_gs_ms_ns(
std::vector<std::size_t>(d_gs_ms_ns_lengths.begin(), d_gs_ms_ns_lengths.end()),
std::vector<std::size_t>(d_gs_ms_ns_strides.begin(), d_gs_ms_ns_strides.end()));
Tensor<EDataType> e_gs_ms_ns_host_result(
std::vector<std::size_t>(e_gs_ms_ns_lengths.begin(), e_gs_ms_ns_lengths.end()),
std::vector<std::size_t>(e_gs_ms_ns_strides.begin(), e_gs_ms_ns_strides.end()));
Tensor<EDataType> e_gs_ms_ns_device_result(
std::vector<std::size_t>(e_gs_ms_ns_lengths.begin(), e_gs_ms_ns_lengths.end()),
std::vector<std::size_t>(e_gs_ms_ns_strides.begin(), e_gs_ms_ns_strides.end()));
std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl; std::cout << "a_gs_ms_ks: " << a_gs_ms_ks.mDesc << std::endl;
std::cout << "b_gs_ns_ks: " << b_gs_ns_ks.mDesc << std::endl; std::cout << "b_gs_ns_ks: " << b_gs_ns_ks.mDesc << std::endl;
...@@ -362,9 +352,7 @@ int main(int argc, char* argv[]) ...@@ -362,9 +352,7 @@ int main(int argc, char* argv[])
if(do_verification) if(do_verification)
{ {
Tensor<CShuffleDataType> c_ms_ns_host_result( Tensor<CShuffleDataType> c_ms_ns_host_result(e_gs_ms_ns_lengths, e_gs_ms_ns_strides);
std::vector<std::size_t>(e_gs_ms_ns_lengths.begin(), e_gs_ms_ns_lengths.end()),
std::vector<std::size_t>(e_gs_ms_ns_strides.begin(), e_gs_ms_ns_strides.end()));
using ReferenceOpInstance = ReferenceContraction_G2_M2_N2_K1<NumDimG, using ReferenceOpInstance = ReferenceContraction_G2_M2_N2_K1<NumDimG,
NumDimM, NumDimM,
...@@ -409,9 +397,7 @@ int main(int argc, char* argv[]) ...@@ -409,9 +397,7 @@ int main(int argc, char* argv[])
} }
} }
return ck::utils::check_err(e_gs_ms_ns_device_result.mData, e_gs_ms_ns_host_result.mData) return ck::utils::check_err(e_gs_ms_ns_device_result, e_gs_ms_ns_host_result) ? 0 : 1;
? 0
: 1;
} }
return 0; return 0;
......
example/30_grouped_conv_fwd_multiple_d/CMakeLists.txt 0 → 100644
View file @ 24af0144
add_custom_target(example_grouped_conv_fwd_multiple_d)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_fp16 grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_fp32 grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_bf16 grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_int8 grouped_conv_fwd_bias_relu_add_xdl_int8.cpp)
add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_fp16)
add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_fp32)
add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_bf16)
add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_int8)
if(USE_BITINT_EXTENSION_INT4)
add_example_executable(example_grouped_conv_fwd_bias_relu_add_xdl_int4 grouped_conv_fwd_bias_relu_add_xdl_int4.cpp)
add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_bias_relu_add_xdl_int4)
endif() # USE_BITINT_EXTENSION_INT4
add_example_executable(example_grouped_conv_fwd_xdl_fp16 grouped_conv_fwd_xdl_fp16.cpp)
add_dependencies(example_grouped_conv_fwd_multiple_d example_grouped_conv_fwd_xdl_fp16)
example/30_grouped_conv_fwd_multiple_d/README.md 0 → 100644
View file @ 24af0144
Command
```bash
arg1: verification (0=no, 1=yes)
arg2: initialization (0=no init, 1=integer value, 2=decimal value)
arg3: time kernel (0=no, 1=yes)
Following arguments (depending on number of spatial dims):
Number of spatial dimensions (1=Conv1d, 2=Conv2d, 3=Conv3d)
G, N, K, C,
<filter spatial dimensions>, (ie Y, X for 2D)
<input image spatial dimensions>, (ie Hi, Wi for 2D)
<strides>, (ie Sy, Sx for 2D)
<dilations>, (ie Dy, Dx for 2D)
<left padding>, (ie LeftPy, LeftPx for 2D)
<right padding>, (ie RightPy, RightPx for 2D)
./bin/example_grouped_conv_fwd_bias_relu_add_xdl_fp16 1 1 1
```
Result (MI100)
```
in: dim 5, lengths {1, 128, 192, 71, 71}, strides {192, 967872, 1, 13632, 192}
wei: dim 5, lengths {1, 256, 192, 3, 3}, strides {442368, 1728, 1, 576, 192}
bias: dim 5, lengths {1, 128, 256, 36, 36}, strides {256, 0, 1, 0, 0}
residual: dim 5, lengths {1, 128, 256, 36, 36}, strides {256, 0, 1, 0, 0}
out: dim 5, lengths {1, 128, 256, 36, 36}, strides {256, 331776, 1, 9216, 256}
launch_and_time_kernel: grid_dim {1296, 1, 1}, block_dim {256, 1, 1}
Warm up 1 time
Start running 10 times...
Perf: 1.55981 ms, 94.0927 TFlops, 213.868 GB/s, DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<256, 128, 256, 16, Default>
```
example/30_grouped_convnd_fwd_bias_relu_add/grouped_convnd_fwd_bias_relu_add_xdl_int4.cpp → example/30_grouped_conv_fwd_multiple_d/common.hpp
View file @ 24af0144
// SPDX-License-Identifier: MIT // SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved. // Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "grouped_convnd_fwd_bias_relu_add_common.hpp" #pragma once
#include "ck/tensor_operation/gpu/device/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp" #include <algorithm>
#include <array>
#include "ck/library/utility/convolution_host_tensor_descriptor_helper.hpp" #include <iostream>
#include <string>
// kernel data types #include <type_traits>
using InKernelDataType = int8_t;
using WeiKernelDataType = int8_t; #include "ck/ck.hpp"
using AccDataType = int32_t; #include "ck/tensor_operation/gpu/device/convolution_forward_specialization.hpp"
using CShuffleDataType = int8_t; #include "ck/tensor_operation/gpu/device/gemm_specialization.hpp"
using BiasKernelDataType = int8_t; #include "ck/tensor_operation/gpu/device/impl/device_grouped_conv_fwd_multiple_d_xdl_cshuffle.hpp"
using ResidualKernelDataType = int8_t; #include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
using OutKernelDataType = int8_t; #include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
// tensor data types #include "ck/library/utility/check_err.hpp"
using InUserDataType = ck::int4_t; #include "ck/library/utility/device_memory.hpp"
using WeiUserDataType = ck::int4_t; #include "ck/library/utility/host_tensor.hpp"
using OutUserDataType = ck::int4_t; #include "ck/library/utility/host_tensor_generator.hpp"
#include "ck/library/utility/convolution_parameter.hpp"
#include "ck/library/reference_tensor_operation/cpu/reference_conv_fwd.hpp"
using BF16 = ck::bhalf_t;
using FP16 = ck::half_t;
using FP32 = float;
#ifdef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
using I4 = ck::int4_t;
#endif
using I8 = std::int8_t;
using I32 = std::int32_t;
template <ck::index_t... Is> template <ck::index_t... Is>
using S = ck::Sequence<Is...>; using S = ck::Sequence<Is...>;
using InElementOp = ck::tensor_operation::element_wise::PassThrough; using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::AddReluAdd;
static constexpr auto ConvSpec = static constexpr auto ConvSpec =
ck::tensor_operation::device::ConvolutionForwardSpecialization::Default; ck::tensor_operation::device::ConvolutionForwardSpecialization::Default;
static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding; static constexpr auto GemmSpec = ck::tensor_operation::device::GemmSpecialization::MNKPadding;
template <ck::index_t NDimSpatial, template <typename InputLay, typename WeightLay, typename OutputLay>
typename InLayout, struct CommonLayoutSetting
typename WeiLayout, {
typename BiasLayout, using InputLayout = InputLay;
typename ResidualLayout, using WeightLayout = WeightLay;
typename OutLayout> using OutputLayout = OutputLay;
using DeviceGroupedConvNDFwdInstance = };
ck::tensor_operation::device::DeviceGroupedConvFwdMultipleD_Xdl_CShuffle<
NDimSpatial, template <ck::index_t NDimSpatial>
InLayout, struct CommonLayoutSettingSelector;
WeiLayout,
ck::Tuple<BiasLayout, ResidualLayout>, namespace ctl = ck::tensor_layout::convolution;
OutLayout,
InKernelDataType, template <>
WeiKernelDataType, struct CommonLayoutSettingSelector<1> final
AccDataType, : CommonLayoutSetting<ctl::G_NW_C, ctl::G_K_X_C, ctl::G_NW_K>
CShuffleDataType,
ck::Tuple<BiasKernelDataType, ResidualKernelDataType>,
OutKernelDataType,
InElementOp,
WeiElementOp,
OutElementOp,
ConvSpec, // ConvForwardSpecialization
GemmSpec, // GemmSpecialization
1, //
256, // BlockSize
128, // MPerBlock
256, // NPerBlock
64, // KPerBlock
16, // AK1
16, // BK1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
4, // NXdlPerWave
S<4, 64, 1>, // ABlockTransferThreadClusterLengths_AK0_M_AK1
S<1, 0, 2>, // ABlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
16, // ABlockTransferSrcScalarPerVector
16, // ABlockTransferDstScalarPerVector_AK1
1, // ABlockLdsExtraM
S<4, 64, 1>, // BBlockTransferThreadClusterLengths_BK0_N_BK1
S<1, 0, 2>, // BBlockTransferThreadClusterArrangeOrder
S<1, 0, 2>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
16, // BBlockTransferSrcScalarPerVector
16, // BBlockTransferDstScalarPerVector_BK1
1, // BBlockLdsExtraN
1,
1,
S<1, 64, 1, 4>,
16>;
int main(int argc, char* argv[])
{ {
namespace ctc = ck::tensor_layout::convolution; };
print_helper_msg(); template <>
struct CommonLayoutSettingSelector<2> final
: CommonLayoutSetting<ctl::G_NHW_C, ctl::G_K_YX_C, ctl::G_NHW_K>
{
};
template <>
struct CommonLayoutSettingSelector<3> final
: CommonLayoutSetting<ctl::G_NDHW_C, ctl::G_K_ZYX_C, ctl::G_NDHW_K>
{
};
template <ck::index_t NDimSpatial>
using InputLayout = typename CommonLayoutSettingSelector<NDimSpatial>::InputLayout;
template <ck::index_t NDimSpatial>
using WeightLayout = typename CommonLayoutSettingSelector<NDimSpatial>::WeightLayout;
template <ck::index_t NDimSpatial>
using OutputLayout = typename CommonLayoutSettingSelector<NDimSpatial>::OutputLayout;
struct ExecutionConfig final
{
bool do_verification = true; bool do_verification = true;
int init_method = 1; int init_method = 1;
bool time_kernel = false; bool time_kernel = true;
};
// conventional group conv definition
// G = 2 #define DefaultConvParam \
// [N, C, Hi, Wi] = [128, 384, 71, 71] ck::utils::conv::ConvParam \
// [K, C, Y, X] = [512, 192, 3, 3] { \
// [N, K, Ho, Wo] = [128, 512, 36, 36] 2, 32, 2, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, { 1, 1 } \
// CK group conv definition }
// [G, N, C, Hi, Wi] = [2, 128, 192, 71, 71]
// [G, K, C, Y, X] = [2, 256, 192, 3, 3] inline void print_help_msg()
// [G, N, K, Ho, Wo] = [2, 128, 256, 36, 36] {
ck::utils::conv::ConvParam conv_param{ std::cerr << "arg1: verification (0=no, 1=yes)\n"
2, 2, 128, 256, 192, {3, 3}, {71, 71}, {2, 2}, {1, 1}, {1, 1}, {1, 1}}; << "arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n"
<< "arg3: time kernel (0=no, 1=yes)\n"
<< ck::utils::conv::get_conv_param_parser_helper_msg() << std::endl;
}
inline bool parse_cmd_args(int argc,
char* argv[],
ExecutionConfig& config,
ck::utils::conv::ConvParam& conv_param)
{
constexpr int num_execution_config_args =
3; // arguments for do_verification, init_method, time_kernel
constexpr int num_conv_param_leading_args = 5; // arguments for num_dim_spatial_, G_, N_, K_, C_
constexpr int threshold_to_catch_partial_args = 1 + num_execution_config_args;
constexpr int threshold_to_catch_all_args =
threshold_to_catch_partial_args + num_conv_param_leading_args;
if(argc == 1) if(argc == 1)
{ {
// use default // use default
} }
else if(argc == 4) // catch only ExecutionConfig arguments
else if(argc == threshold_to_catch_partial_args)
{ {
do_verification = std::stoi(argv[1]); config.do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]); config.init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]); config.time_kernel = std::stoi(argv[3]);
} }
else // catch both ExecutionConfig & ConvParam arguments
else if(threshold_to_catch_all_args < argc && ((argc - threshold_to_catch_all_args) % 3 == 0))
{ {
do_verification = std::stoi(argv[1]); config.do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]); config.init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]); config.time_kernel = std::stoi(argv[3]);
const ck::index_t num_dim_spatial = std::stoi(argv[4]);
conv_param = ck::utils::conv::parse_conv_param(num_dim_spatial, 5, argv); const ck::index_t num_dim_spatial = std::stoi(argv[4]);
conv_param = ck::utils::conv::parse_conv_param(
num_dim_spatial, threshold_to_catch_partial_args, argv);
}
else
{
print_help_msg();
return false;
} }
const auto in_element_op = InElementOp{}; return true;
const auto wei_element_op = WeiElementOp{}; }
const auto out_element_op = OutElementOp{};
if(conv_param.num_dim_spatial_ == 1) inline HostTensorDescriptor make_input_descriptor(const ck::utils::conv::ConvParam& conv_param)
{
switch(conv_param.num_dim_spatial_)
{ {
using InLayout = ctc::G_NW_C; case 1:
using WeiLayout = ctc::G_K_X_C; return HostTensorDescriptor(
using BiasLayout = ctc::G_K;
using ResidualLayout = ctc::G_NW_K;
using OutLayout = ctc::G_NW_K;
const auto in_g_n_c_wis_desc = HostTensorDescriptor(
{conv_param.G_, conv_param.N_, conv_param.C_, conv_param.input_spatial_lengths_[0]}, {conv_param.G_, conv_param.N_, conv_param.C_, conv_param.input_spatial_lengths_[0]},
{ {
conv_param.C_, // g conv_param.C_, // g
...@@ -150,81 +161,8 @@ int main(int argc, char* argv[]) ...@@ -150,81 +161,8 @@ int main(int argc, char* argv[])
conv_param.G_ * conv_param.C_ // wi conv_param.G_ * conv_param.C_ // wi
}); });
const auto wei_g_k_c_xs_desc = HostTensorDescriptor( case 2:
{conv_param.G_, conv_param.K_, conv_param.C_, conv_param.filter_spatial_lengths_[0]}, return HostTensorDescriptor(
{
conv_param.K_ * conv_param.filter_spatial_lengths_[0] * conv_param.C_, // g
conv_param.filter_spatial_lengths_[0] * conv_param.C_, // k
1, // c
conv_param.C_ // x
});
const auto bias_g_n_k_wos_desc = HostTensorDescriptor(
{conv_param.G_, conv_param.N_, conv_param.K_, conv_param.output_spatial_lengths_[0]},
{
conv_param.K_, // g
0, // k
1, // c
0 // x
});
const auto residual_g_n_k_wos_desc = HostTensorDescriptor(
{conv_param.G_, conv_param.N_, conv_param.K_, conv_param.output_spatial_lengths_[0]},
{
conv_param.K_, // g
0, // k
1, // c
0 // x
});
const auto out_g_n_k_wos_desc = HostTensorDescriptor(
{conv_param.G_, conv_param.N_, conv_param.K_, conv_param.output_spatial_lengths_[0]},
{
conv_param.K_, // g
conv_param.output_spatial_lengths_[0] * conv_param.G_ * conv_param.K_, // n
1, // k
conv_param.G_ * conv_param.K_ // wo
});
return run_grouped_conv_fwd_bias_relu_add<1,
InKernelDataType,
WeiKernelDataType,
CShuffleDataType,
OutKernelDataType,
InElementOp,
WeiElementOp,
OutElementOp,
InUserDataType,
WeiUserDataType,
OutUserDataType,
DeviceGroupedConvNDFwdInstance<1,
InLayout,
WeiLayout,
BiasLayout,
ResidualLayout,
OutLayout>>(
do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
bias_g_n_k_wos_desc,
residual_g_n_k_wos_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op);
}
else if(conv_param.num_dim_spatial_ == 2)
{
using InLayout = ctc::G_NHW_C;
using WeiLayout = ctc::G_K_YX_C;
using BiasLayout = ctc::G_K;
using ResidualLayout = ctc::G_NHW_K;
using OutLayout = ctc::G_NHW_K;
const auto in_g_n_c_wis_desc = HostTensorDescriptor(
{conv_param.G_, {conv_param.G_,
conv_param.N_, conv_param.N_,
conv_param.C_, conv_param.C_,
...@@ -239,104 +177,8 @@ int main(int argc, char* argv[]) ...@@ -239,104 +177,8 @@ int main(int argc, char* argv[])
conv_param.G_ * conv_param.C_ // wi conv_param.G_ * conv_param.C_ // wi
}); });
const auto wei_g_k_c_xs_desc = case 3:
HostTensorDescriptor({conv_param.G_, return HostTensorDescriptor(
conv_param.K_,
conv_param.C_,
conv_param.filter_spatial_lengths_[0],
conv_param.filter_spatial_lengths_[1]},
{
conv_param.K_ * conv_param.filter_spatial_lengths_[0] *
conv_param.filter_spatial_lengths_[1] * conv_param.C_, // g
conv_param.filter_spatial_lengths_[0] *
conv_param.filter_spatial_lengths_[1] * conv_param.C_, // k
1, // c
conv_param.filter_spatial_lengths_[1] * conv_param.C_, // y
conv_param.C_ // x
});
const auto bias_g_n_k_wos_desc =
HostTensorDescriptor({conv_param.G_,
conv_param.N_,
conv_param.K_,
conv_param.output_spatial_lengths_[0],
conv_param.output_spatial_lengths_[1]},
{
conv_param.K_, // g
0, // n
1, // k
0, // ho
0 // wo
});
const auto residual_g_n_k_wos_desc =
HostTensorDescriptor({conv_param.G_,
conv_param.N_,
conv_param.K_,
conv_param.output_spatial_lengths_[0],
conv_param.output_spatial_lengths_[1]},
{
conv_param.K_, // g
0, // n
1, // k
0, // ho
0 // wo
});
const auto out_g_n_k_wos_desc = HostTensorDescriptor(
{conv_param.G_,
conv_param.N_,
conv_param.K_,
conv_param.output_spatial_lengths_[0],
conv_param.output_spatial_lengths_[1]},
{
conv_param.K_, // g
conv_param.output_spatial_lengths_[0] * conv_param.output_spatial_lengths_[1] *
conv_param.G_ * conv_param.K_, // n
1, // k
conv_param.output_spatial_lengths_[1] * conv_param.G_ * conv_param.K_, // ho
conv_param.G_ * conv_param.K_ // wo
});
return run_grouped_conv_fwd_bias_relu_add<2,
InKernelDataType,
WeiKernelDataType,
CShuffleDataType,
OutKernelDataType,
InElementOp,
WeiElementOp,
OutElementOp,
InUserDataType,
WeiUserDataType,
OutUserDataType,
DeviceGroupedConvNDFwdInstance<2,
InLayout,
WeiLayout,
BiasLayout,
ResidualLayout,
OutLayout>>(
do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
bias_g_n_k_wos_desc,
residual_g_n_k_wos_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op);
}
else if(conv_param.num_dim_spatial_ == 3)
{
using InLayout = ctc::G_NDHW_C;
using WeiLayout = ctc::G_K_ZYX_C;
using BiasLayout = ctc::G_K;
using ResidualLayout = ctc::G_NDHW_K;
using OutLayout = ctc::G_NDHW_K;
const auto in_g_n_c_wis_desc = HostTensorDescriptor(
{conv_param.G_, {conv_param.G_,
conv_param.N_, conv_param.N_,
conv_param.C_, conv_param.C_,
...@@ -353,8 +195,42 @@ int main(int argc, char* argv[]) ...@@ -353,8 +195,42 @@ int main(int argc, char* argv[])
conv_param.input_spatial_lengths_[2] * conv_param.G_ * conv_param.C_, // hi conv_param.input_spatial_lengths_[2] * conv_param.G_ * conv_param.C_, // hi
conv_param.G_ * conv_param.C_ // wi conv_param.G_ * conv_param.C_ // wi
}); });
}
const auto wei_g_k_c_xs_desc = HostTensorDescriptor( throw std::runtime_error("unsuppored # dim spatial");
}
inline HostTensorDescriptor make_weight_descriptor(const ck::utils::conv::ConvParam& conv_param)
{
switch(conv_param.num_dim_spatial_)
{
case 1:
return HostTensorDescriptor(
{conv_param.G_, conv_param.K_, conv_param.C_, conv_param.filter_spatial_lengths_[0]},
{
conv_param.K_ * conv_param.filter_spatial_lengths_[0] * conv_param.C_, // g
conv_param.filter_spatial_lengths_[0] * conv_param.C_, // k
1, // c
conv_param.C_ // x
});
case 2:
return HostTensorDescriptor(
{conv_param.G_,
conv_param.K_,
conv_param.C_,
conv_param.filter_spatial_lengths_[0],
conv_param.filter_spatial_lengths_[1]},
{
conv_param.K_ * conv_param.filter_spatial_lengths_[0] *
conv_param.filter_spatial_lengths_[1] * conv_param.C_, // g
conv_param.filter_spatial_lengths_[0] * conv_param.filter_spatial_lengths_[1] *
conv_param.C_, // k
1, // c
conv_param.filter_spatial_lengths_[1] * conv_param.C_, // y
conv_param.C_ // x
});
case 3:
return HostTensorDescriptor(
{conv_param.G_, {conv_param.G_,
conv_param.K_, conv_param.K_,
conv_param.C_, conv_param.C_,
...@@ -373,40 +249,89 @@ int main(int argc, char* argv[]) ...@@ -373,40 +249,89 @@ int main(int argc, char* argv[])
conv_param.filter_spatial_lengths_[2] * conv_param.C_, // y conv_param.filter_spatial_lengths_[2] * conv_param.C_, // y
conv_param.C_ // x conv_param.C_ // x
}); });
}
const auto bias_g_n_k_wos_desc = throw std::runtime_error("unsuppored # dim spatial");
HostTensorDescriptor({conv_param.G_, }
conv_param.N_,
conv_param.K_, inline HostTensorDescriptor make_bias_descriptor(const ck::utils::conv::ConvParam& conv_param)
conv_param.output_spatial_lengths_[0], {
conv_param.output_spatial_lengths_[1], switch(conv_param.num_dim_spatial_)
conv_param.output_spatial_lengths_[2]}, {
{ case 1:
conv_param.K_, // g return HostTensorDescriptor(
0, // n {conv_param.G_, conv_param.N_, conv_param.K_, conv_param.output_spatial_lengths_[0]},
1, // k {
0, // z conv_param.K_, // g
0, // y 0, // k
0 // x 1, // c
}); 0 // x
});
const auto residual_g_n_k_wos_desc = case 2:
HostTensorDescriptor({conv_param.G_, return HostTensorDescriptor({conv_param.G_,
conv_param.N_, conv_param.N_,
conv_param.K_, conv_param.K_,
conv_param.output_spatial_lengths_[0], conv_param.output_spatial_lengths_[0],
conv_param.output_spatial_lengths_[1], conv_param.output_spatial_lengths_[1]},
conv_param.output_spatial_lengths_[2]}, {
{ conv_param.K_, // g
conv_param.K_, // g 0, // n
0, // n 1, // k
1, // k 0, // ho
0, // z 0 // wo
0, // y });
0 // x case 3:
}); return HostTensorDescriptor({conv_param.G_,
conv_param.N_,
const auto out_g_n_k_wos_desc = HostTensorDescriptor( conv_param.K_,
conv_param.output_spatial_lengths_[0],
conv_param.output_spatial_lengths_[1],
conv_param.output_spatial_lengths_[2]},
{
conv_param.K_, // g
0, // n
1, // k
0, // z
0, // y
0 // x
});
}
throw std::runtime_error("unsuppored # dim spatial");
}
inline HostTensorDescriptor make_output_descriptor(const ck::utils::conv::ConvParam& conv_param)
{
switch(conv_param.num_dim_spatial_)
{
case 1:
return HostTensorDescriptor(
{conv_param.G_, conv_param.N_, conv_param.K_, conv_param.output_spatial_lengths_[0]},
{
conv_param.K_, // g
conv_param.output_spatial_lengths_[0] * conv_param.G_ * conv_param.K_, // n
1, // k
conv_param.G_ * conv_param.K_ // wo
});
case 2:
return HostTensorDescriptor(
{conv_param.G_,
conv_param.N_,
conv_param.K_,
conv_param.output_spatial_lengths_[0],
conv_param.output_spatial_lengths_[1]},
{
conv_param.K_, // g
conv_param.output_spatial_lengths_[0] * conv_param.output_spatial_lengths_[1] *
conv_param.G_ * conv_param.K_, // n
1, // k
conv_param.output_spatial_lengths_[1] * conv_param.G_ * conv_param.K_, // ho
conv_param.G_ * conv_param.K_ // wo
});
case 3:
return HostTensorDescriptor(
{conv_param.G_, {conv_param.G_,
conv_param.N_, conv_param.N_,
conv_param.K_, conv_param.K_,
...@@ -423,37 +348,7 @@ int main(int argc, char* argv[]) ...@@ -423,37 +348,7 @@ int main(int argc, char* argv[])
conv_param.output_spatial_lengths_[2] * conv_param.G_ * conv_param.K_, // ho conv_param.output_spatial_lengths_[2] * conv_param.G_ * conv_param.K_, // ho
conv_param.G_ * conv_param.K_ // wo conv_param.G_ * conv_param.K_ // wo
}); });
return run_grouped_conv_fwd_bias_relu_add<3,
InKernelDataType,
WeiKernelDataType,
CShuffleDataType,
OutKernelDataType,
InElementOp,
WeiElementOp,
OutElementOp,
InUserDataType,
WeiUserDataType,
OutUserDataType,
DeviceGroupedConvNDFwdInstance<3,
InLayout,
WeiLayout,
BiasLayout,
ResidualLayout,
OutLayout>>(
do_verification,
init_method,
time_kernel,
conv_param,
in_g_n_c_wis_desc,
wei_g_k_c_xs_desc,
bias_g_n_k_wos_desc,
residual_g_n_k_wos_desc,
out_g_n_k_wos_desc,
in_element_op,
wei_element_op,
out_element_op);
} }
return 0; throw std::runtime_error("unsuppored # dim spatial");
} }
example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_bf16.cpp 0 → 100644
View file @ 24af0144
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
// kernel data types
using InKernelDataType = BF16;
using WeiKernelDataType = BF16;
using AccDataType = FP32;
using CShuffleDataType = FP32;
using BiasKernelDataType = BF16;
using ResidualKernelDataType = BF16;
using OutKernelDataType = BF16;
// tensor data types
using InUserDataType = InKernelDataType;
using WeiUserDataType = WeiKernelDataType;
using OutUserDataType = OutKernelDataType;
using InElementOp = PassThrough;
using WeiElementOp = PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::AddReluAdd;
#include "run_grouped_conv_fwd_bias_relu_add_example.inc"
int main(int argc, char* argv[]) { return !run_grouped_conv_fwd_bias_relu_add_example(argc, argv); }
example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp16.cpp 0 → 100644
View file @ 24af0144
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
// kernel data types
using InKernelDataType = FP16;
using WeiKernelDataType = FP16;
using AccDataType = FP32;
using CShuffleDataType = FP16;
using BiasKernelDataType = FP16;
using ResidualKernelDataType = FP16;
using OutKernelDataType = FP16;
// tensor data types
using InUserDataType = InKernelDataType;
using WeiUserDataType = WeiKernelDataType;
using OutUserDataType = OutKernelDataType;
using InElementOp = PassThrough;
using WeiElementOp = PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::AddReluAdd;
#include "run_grouped_conv_fwd_bias_relu_add_example.inc"
int main(int argc, char* argv[]) { return !run_grouped_conv_fwd_bias_relu_add_example(argc, argv); }
example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_fp32.cpp 0 → 100644
View file @ 24af0144
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#include "common.hpp"
// kernel data types
using InKernelDataType = FP32;
using WeiKernelDataType = FP32;
using AccDataType = FP32;
using CShuffleDataType = FP32;
using BiasKernelDataType = FP32;
using ResidualKernelDataType = FP32;
using OutKernelDataType = FP32;
// tensor data types
using InUserDataType = InKernelDataType;
using WeiUserDataType = WeiKernelDataType;
using OutUserDataType = OutKernelDataType;
using InElementOp = PassThrough;
using WeiElementOp = PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::AddReluAdd;
#include "run_grouped_conv_fwd_bias_relu_add_example.inc"
int main(int argc, char* argv[]) { return !run_grouped_conv_fwd_bias_relu_add_example(argc, argv); }
example/30_grouped_conv_fwd_multiple_d/grouped_conv_fwd_bias_relu_add_xdl_int4.cpp 0 → 100644
View file @ 24af0144
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#ifndef CK_EXPERIMENTAL_BIT_INT_EXTENSION_INT4
#error Should compile this file with ck::int4_t support
#endif
#include "common.hpp"
// kernel data types
using InKernelDataType = I8;
using WeiKernelDataType = I8;
using AccDataType = I32;
using CShuffleDataType = I8;
using BiasKernelDataType = I8;
using ResidualKernelDataType = I8;
using OutKernelDataType = I8;
// tensor data types
using InUserDataType = I4;
using WeiUserDataType = I4;
using OutUserDataType = I4;
using InElementOp = PassThrough;
using WeiElementOp = PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::AddReluAdd;
#define BUILD_INT4_EXAMPLE
#include "run_grouped_conv_fwd_bias_relu_add_example.inc"
int main(int argc, char* argv[]) { return !run_grouped_conv_fwd_bias_relu_add_example(argc, argv); }
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