Skip to content

GitLab

Commit d8f1458f authored by Jing Zhang's avatar Jing Zhang
Browse files

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

parents 6e983ba2 40b59a63
Hide whitespace changes
Inline Side-by-side
Showing with 331 additions and 143 deletions
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_partial_reduce_i8_i32_i8.hpp deleted 100644 → 0
View file @ 6e983ba2
#ifndef DEVICE_REDUCE_INSTANCE_MULTIBLOCK_PARTIAL_REDUCE_I8_I32_I8_HPP
#define DEVICE_REDUCE_INSTANCE_MULTIBLOCK_PARTIAL_REDUCE_I8_I32_I8_HPP
#include "reduction_enums.hpp"
#include "reduction_operator_mapping.hpp"
#include "device_reduce_instance_multiblock_partial_reduce.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_reduce_instance {
// clang-format off
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int32_t, int8_t, 0, 0, 0, 4, 3); // for ADD
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int32_t, int8_t, 0, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int32_t, int8_t, 0, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int32_t, int8_t, 0, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int32_t, int8_t, 5, 0, 0, 4, 3); // for AVG
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int32_t, int8_t, 5, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int32_t, int8_t, 5, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int32_t, int8_t, 5, 0, 0, 2, 1);
// clang-format on
} // namespace device_reduce_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
#endif
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_partial_reduce_i8_i8_i8.hpp deleted 100644 → 0
View file @ 6e983ba2
#ifndef DEVICE_REDUCE_INSTANCE_MULTIBLOCK_PARTIAL_REDUCE_I8_I8_I8_HPP
#define DEVICE_REDUCE_INSTANCE_MULTIBLOCK_PARTIAL_REDUCE_I8_I8_I8_HPP
#include "reduction_enums.hpp"
#include "reduction_operator_mapping.hpp"
#include "device_reduce_instance_multiblock_partial_reduce.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_reduce_instance {
// clang-format off
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 4, 3); // for MIN
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int8_t, int8_t, 2, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 4, 3); // for MAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int8_t, int8_t, 3, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 4, 3); // for AMAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int8_t, int8_t, 4, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 4, 3); // for MIN
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int8_t, int8_t, 2, 0, 1, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 4, 3); // for MAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int8_t, int8_t, 3, 0, 1, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 4, 3); // for AMAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(int8_t, int8_t, int8_t, 4, 0, 1, 2, 1);
// clang-format on
} // namespace device_reduce_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
#endif
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise.hpp
View file @ d8f1458f
...@@ -58,8 +58,8 @@ template <typename InDataType, ...@@ -58,8 +58,8 @@ template <typename InDataType,
int Rank, int Rank,
int NumReduceDim, int NumReduceDim,
ReduceTensorOp ReduceOpId, ReduceTensorOp ReduceOpId,
NanPropagation NanOpt, bool PropagateNan,
ReduceTensorIndices IndicesOpt> bool UseIndex>
void add_device_reduce_instance_threadwise( void add_device_reduce_instance_threadwise(
std::vector<deviceReduceThreadWisePtrType<AccDataType, ReduceOpId>>& device_op_instances) std::vector<deviceReduceThreadWisePtrType<AccDataType, ReduceOpId>>& device_op_instances)
{ {
...@@ -73,9 +73,7 @@ void add_device_reduce_instance_threadwise( ...@@ -73,9 +73,7 @@ void add_device_reduce_instance_threadwise(
constexpr bool Indexable = constexpr bool Indexable =
(ReduceOpId == ReduceTensorOp::MIN || ReduceOpId == ReduceTensorOp::MAX || (ReduceOpId == ReduceTensorOp::MIN || ReduceOpId == ReduceTensorOp::MAX ||
ReduceOpId == ReduceTensorOp::AMAX); ReduceOpId == ReduceTensorOp::AMAX);
constexpr bool NeedIndices = Indexable && (IndicesOpt != ReduceTensorIndices::NO_INDICES); constexpr bool OutputIndex = Indexable && UseIndex;
constexpr bool PropagateNan = (NanOpt == NanPropagation::NOT_PROPAGATE_NAN) ? false : true;
using cfg1 = ReductionConfiguration_1<256, 256, 1>; using cfg1 = ReductionConfiguration_1<256, 256, 1>;
...@@ -93,10 +91,9 @@ void add_device_reduce_instance_threadwise( ...@@ -93,10 +91,9 @@ void add_device_reduce_instance_threadwise(
InElementwiseOperation, InElementwiseOperation,
AccElementwiseOperation, AccElementwiseOperation,
PropagateNan, PropagateNan,
NeedIndices, OutputIndex,
false, // HaveIndexInputIfOutputIndex
cfg1::BlockSize_, cfg1::BlockSize_,
cfg1::MThreadClusterSize_,
cfg1::KThreadClusterSize_,
cfg2::MThreadSliceSize_, cfg2::MThreadSliceSize_,
cfg2::KThreadSliceSize_, cfg2::KThreadSliceSize_,
cfg2::InSrcVectorDim_, cfg2::InSrcVectorDim_,
...@@ -107,54 +104,54 @@ void add_device_reduce_instance_threadwise( ...@@ -107,54 +104,54 @@ void add_device_reduce_instance_threadwise(
}); });
}; };
#define ADD_THREADWISE_INST_BY_TYPE( \ #define ADD_THREADWISE_INST_BY_TYPE( \
inT, compT, outT, ReduceOpId, NanOpt, IndicesOpt, Rank, NumReduceDim) \ inT, compT, outT, ReduceOpId, PropagateNan, UseIndex, Rank, NumReduceDim) \
template void add_device_reduce_instance_threadwise<inT, \ template void add_device_reduce_instance_threadwise<inT, \
compT, \ compT, \
outT, \ outT, \
Rank, \ Rank, \
NumReduceDim, \ NumReduceDim, \
ReduceOpId, \ ReduceOpId, \
NanOpt, \ PropagateNan, \
IndicesOpt>( \ UseIndex>( \
std::vector<deviceReduceThreadWisePtrType<compT, ReduceOpId>> & device_op_instances) std::vector<deviceReduceThreadWisePtrType<compT, ReduceOpId>> & device_op_instances)
#define ADD_THREADWISE_INST_BY_ID( \ #define ADD_THREADWISE_INST_BY_ID( \
inT, compT, outT, ReduceOpId, NanOpt, IndicesOpt, Rank, NumReduceDim) \ inT, compT, outT, ReduceOpId, NanOpt, IndicesOpt, Rank, NumReduceDim) \
ADD_THREADWISE_INST_BY_TYPE(inT, \ ADD_THREADWISE_INST_BY_TYPE(inT, \
compT, \ compT, \
outT, \ outT, \
static_cast<ReduceTensorOp>(ReduceOpId), \ static_cast<ReduceTensorOp>(ReduceOpId), \
static_cast<NanPropagation>(NanOpt), \ static_cast<bool>(NanOpt), \
static_cast<ReduceTensorIndices>(IndicesOpt), \ static_cast<bool>(IndicesOpt), \
Rank, \ Rank, \
NumReduceDim) NumReduceDim)
#define ADD_THREADWISE_INST_REF_BY_TYPE( \ #define ADD_THREADWISE_INST_REF_BY_TYPE( \
inT, compT, outT, ReduceOpId, NanOpt, IndicesOpt, Rank, NumReduceDim) \ inT, compT, outT, ReduceOpId, PropagateNan, UseIndex, Rank, NumReduceDim) \
extern template void add_device_reduce_instance_threadwise<inT, \ extern template void add_device_reduce_instance_threadwise<inT, \
compT, \ compT, \
outT, \ outT, \
Rank, \ Rank, \
NumReduceDim, \ NumReduceDim, \
ReduceOpId, \ ReduceOpId, \
NanOpt, \ PropagateNan, \
IndicesOpt>( \ UseIndex>( \
std::vector<DeviceReducePtr< \ std::vector<DeviceReducePtr< \
typename reduce_unary_operator<compT, ReduceOpId, true, true>::InElementwiseOperation, \ typename reduce_unary_operator<compT, ReduceOpId, true, true>::InElementwiseOperation, \
typename reduce_unary_operator<compT, ReduceOpId, true, true>:: \ typename reduce_unary_operator<compT, ReduceOpId, true, true>:: \
AccElementwiseOperation>> & \ AccElementwiseOperation>> & \
device_op_instances) device_op_instances)
#define ADD_THREADWISE_INST_REF_BY_ID( \ #define ADD_THREADWISE_INST_REF_BY_ID( \
inT, compT, outT, ReduceOpId, NanOpt, IndicesOpt, Rank, NumReduceDim) \ inT, compT, outT, ReduceOpId, NanOpt, IndicesOpt, Rank, NumReduceDim) \
ADD_THREADWISE_INST_REF_BY_TYPE(inT, \ ADD_THREADWISE_INST_REF_BY_TYPE(inT, \
compT, \ compT, \
outT, \ outT, \
static_cast<ReduceTensorOp>(ReduceOpId), \ static_cast<ReduceTensorOp>(ReduceOpId), \
static_cast<NanPropagation>(NanOpt), \ static_cast<bool>(NanOpt), \
static_cast<ReduceTensorIndices>(IndicesOpt), \ static_cast<bool>(IndicesOpt), \
Rank, \ Rank, \
NumReduceDim) NumReduceDim)
} // namespace device_reduce_instance } // namespace device_reduce_instance
......
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16.hpp
View file @ d8f1458f
#ifndef DEVICE_REDUCE_INSTANCE_THREADWISE_B16_F32_B16_HPP #ifndef DEVICE_REDUCE_INSTANCE_THREADWISE_B16_F32_B16_HPP
#define DEVICE_REDUCE_INSTANCE_THREADWISE_B16_F32_B16_HPP #define DEVICE_REDUCE_INSTANCE_THREADWISE_B16_F32_B16_HPP
#include "reduction_enums.hpp" #include "data_type.hpp"
#include "reduction_operator_mapping.hpp"
#include "device_reduce_instance_threadwise.hpp" #include "device_reduce_instance_threadwise.hpp"
namespace ck { namespace ck {
......
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16.hpp
View file @ d8f1458f
#ifndef DEVICE_REDUCE_INSTANCE_THREADWISE_F16_F16_F16_HPP #ifndef DEVICE_REDUCE_INSTANCE_THREADWISE_F16_F16_F16_HPP
#define DEVICE_REDUCE_INSTANCE_THREADWISE_F16_F16_F16_HPP #define DEVICE_REDUCE_INSTANCE_THREADWISE_F16_F16_F16_HPP
#include "reduction_enums.hpp" #include "data_type.hpp"
#include "reduction_operator_mapping.hpp"
#include "device_reduce_instance_threadwise.hpp" #include "device_reduce_instance_threadwise.hpp"
namespace ck { namespace ck {
......
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16.hpp
View file @ d8f1458f
#ifndef DEVICE_REDUCE_INSTANCE_THREADWISE_F16_F32_F16_HPP #ifndef DEVICE_REDUCE_INSTANCE_THREADWISE_F16_F32_F16_HPP
#define DEVICE_REDUCE_INSTANCE_THREADWISE_F16_F32_F16_HPP #define DEVICE_REDUCE_INSTANCE_THREADWISE_F16_F32_F16_HPP
#include "reduction_enums.hpp" #include "data_type.hpp"
#include "reduction_operator_mapping.hpp"
#include "device_reduce_instance_threadwise.hpp" #include "device_reduce_instance_threadwise.hpp"
namespace ck { namespace ck {
......
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32.hpp
View file @ d8f1458f
#ifndef DEVICE_REDUCE_INSTANCE_THREADWISE_F32_F32_F32_HPP #ifndef DEVICE_REDUCE_INSTANCE_THREADWISE_F32_F32_F32_HPP
#define DEVICE_REDUCE_INSTANCE_THREADWISE_F32_F32_F32_HPP #define DEVICE_REDUCE_INSTANCE_THREADWISE_F32_F32_F32_HPP
#include "reduction_enums.hpp"
#include "reduction_operator_mapping.hpp"
#include "device_reduce_instance_threadwise.hpp" #include "device_reduce_instance_threadwise.hpp"
namespace ck { namespace ck {
......
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32.hpp
View file @ d8f1458f
#ifndef DEVICE_REDUCE_INSTANCE_THREADWISE_F32_F64_F32_HPP #ifndef DEVICE_REDUCE_INSTANCE_THREADWISE_F32_F64_F32_HPP
#define DEVICE_REDUCE_INSTANCE_THREADWISE_F32_F64_F32_HPP #define DEVICE_REDUCE_INSTANCE_THREADWISE_F32_F64_F32_HPP
#include "reduction_enums.hpp"
#include "reduction_operator_mapping.hpp"
#include "device_reduce_instance_threadwise.hpp" #include "device_reduce_instance_threadwise.hpp"
namespace ck { namespace ck {
......
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64.hpp
View file @ d8f1458f
#ifndef DEVICE_REDUCE_INSTANCE_THREADWISE_F64_F64_F64_HPP #ifndef DEVICE_REDUCE_INSTANCE_THREADWISE_F64_F64_F64_HPP
#define DEVICE_REDUCE_INSTANCE_THREADWISE_F64_F64_F64_HPP #define DEVICE_REDUCE_INSTANCE_THREADWISE_F64_F64_F64_HPP
#include "reduction_enums.hpp"
#include "reduction_operator_mapping.hpp"
#include "device_reduce_instance_threadwise.hpp" #include "device_reduce_instance_threadwise.hpp"
namespace ck { namespace ck {
......
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8.hpp
View file @ d8f1458f
#ifndef DEVICE_REDUCE_INSTANCE_THREADWISE_I8_I32_I8_HPP #ifndef DEVICE_REDUCE_INSTANCE_THREADWISE_I8_I32_I8_HPP
#define DEVICE_REDUCE_INSTANCE_THREADWISE_I8_I32_I8_HPP #define DEVICE_REDUCE_INSTANCE_THREADWISE_I8_I32_I8_HPP
#include "reduction_enums.hpp"
#include "reduction_operator_mapping.hpp"
#include "device_reduce_instance_threadwise.hpp" #include "device_reduce_instance_threadwise.hpp"
namespace ck { namespace ck {
......
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8.hpp
View file @ d8f1458f
#ifndef DEVICE_REDUCE_INSTANCE_THREADWISE_I8_I8_I8_HPP #ifndef DEVICE_REDUCE_INSTANCE_THREADWISE_I8_I8_I8_HPP
#define DEVICE_REDUCE_INSTANCE_THREADWISE_I8_I8_I8_HPP #define DEVICE_REDUCE_INSTANCE_THREADWISE_I8_I8_I8_HPP
#include "reduction_enums.hpp"
#include "reduction_operator_mapping.hpp"
#include "device_reduce_instance_threadwise.hpp" #include "device_reduce_instance_threadwise.hpp"
namespace ck { namespace ck {
......
library/include/ck/library/utility/check_err.hpp
View file @ d8f1458f
...@@ -24,7 +24,7 @@ check_err(const std::vector<T>& out, ...@@ -24,7 +24,7 @@ check_err(const std::vector<T>& out,
const std::vector<T>& ref, const std::vector<T>& ref,
const std::string& msg = "Error: Incorrect results!", const std::string& msg = "Error: Incorrect results!",
double rtol = 1e-5, double rtol = 1e-5,
double atol = 1e-8) double atol = 3e-6)
{ {
if(out.size() != ref.size()) if(out.size() != ref.size())
{ {
...@@ -173,8 +173,8 @@ check_err(const std::vector<T>& out, ...@@ -173,8 +173,8 @@ check_err(const std::vector<T>& out,
{ {
if(out[i] != ref[i]) if(out[i] != ref[i])
{ {
std::cout << "out[" << i << "] != ref[" << i << "]: " << out[i] << " != " << ref[i] std::cout << "out[" << i << "] != ref[" << i << "]: " << static_cast<int>(out[i])
<< std::endl << " != " << static_cast<int>(ref[i]) << std::endl
<< msg << std::endl; << msg << std::endl;
return false; return false;
} }
......
library/src/tensor_operation_instance/gpu/CMakeLists.txt
View file @ d8f1458f
include_directories(BEFORE include_directories(BEFORE
${PROJECT_SOURCE_DIR}/include/ck ${PROJECT_SOURCE_DIR}/include/ck
${PROJECT_SOURCE_DIR}/include/ck/utility ${PROJECT_SOURCE_DIR}/include/ck/utility
${PROJECT_SOURCE_DIR}/include/ck/host_utility
${PROJECT_SOURCE_DIR}/include/ck/tensor_description ${PROJECT_SOURCE_DIR}/include/ck/tensor_description
${PROJECT_SOURCE_DIR}/include/ck/tensor ${PROJECT_SOURCE_DIR}/include/ck/tensor
${PROJECT_SOURCE_DIR}/include/ck/problem_transform ${PROJECT_SOURCE_DIR}/include/ck/problem_transform
......
library/src/tensor_operation_instance/gpu/gemm/CMakeLists.txt
View file @ d8f1458f
# device_gemm_instance
set(DEVICE_GEMM_INSTANCE_SOURCE set(DEVICE_GEMM_INSTANCE_SOURCE
device_gemm_xdl_f32_f32_f32_mk_kn_mn_instance.cpp; device_gemm_xdl_f32_f32_f32_mk_kn_mn_instance.cpp;
device_gemm_xdl_f32_f32_f32_mk_nk_mn_instance.cpp; device_gemm_xdl_f32_f32_f32_mk_nk_mn_instance.cpp;
...@@ -8,10 +7,10 @@ set(DEVICE_GEMM_INSTANCE_SOURCE ...@@ -8,10 +7,10 @@ set(DEVICE_GEMM_INSTANCE_SOURCE
device_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp; device_gemm_xdl_f16_f16_f16_mk_nk_mn_instance.cpp;
device_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp; device_gemm_xdl_f16_f16_f16_km_kn_mn_instance.cpp;
device_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp; device_gemm_xdl_f16_f16_f16_km_nk_mn_instance.cpp;
device_gemm_xdl_c_shuffle_int8_int8_int8_mk_kn_mn_instance.cpp; device_gemm_xdl_c_shuffle_i8_i8_i8_mk_kn_mn_instance.cpp;
device_gemm_xdl_c_shuffle_int8_int8_int8_mk_nk_mn_instance.cpp; device_gemm_xdl_c_shuffle_i8_i8_i8_mk_nk_mn_instance.cpp;
device_gemm_xdl_c_shuffle_int8_int8_int8_km_kn_mn_instance.cpp; device_gemm_xdl_c_shuffle_i8_i8_i8_km_kn_mn_instance.cpp;
device_gemm_xdl_c_shuffle_int8_int8_int8_km_nk_mn_instance.cpp; device_gemm_xdl_c_shuffle_i8_i8_i8_km_nk_mn_instance.cpp;
device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instance.cpp; device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_kn_mn_instance.cpp;
device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp; device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp;
device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instance.cpp; device_gemm_xdl_c_shuffle_bf16_bf16_bf16_km_kn_mn_instance.cpp;
...@@ -33,11 +32,21 @@ set(DEVICE_GEMM_INSTANCE_SOURCE ...@@ -33,11 +32,21 @@ set(DEVICE_GEMM_INSTANCE_SOURCE
device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp; device_gemm_xdl_splitk_f16_f16_f16_mk_nk_mn_instance.cpp;
device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instance.cpp; device_gemm_xdl_splitk_f16_f16_f16_km_kn_mn_instance.cpp;
device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instance.cpp; device_gemm_xdl_splitk_f16_f16_f16_km_nk_mn_instance.cpp;
device_gemm_dl_f32_f32_f32_mk_kn_mn_instance.cpp;
device_gemm_dl_f32_f32_f32_mk_nk_mn_instance.cpp;
device_gemm_dl_f32_f32_f32_km_kn_mn_instance.cpp;
device_gemm_dl_f32_f32_f32_km_nk_mn_instance.cpp;
device_gemm_dl_f16_f16_f16_mk_kn_mn_instance.cpp;
device_gemm_dl_f16_f16_f16_mk_nk_mn_instance.cpp;
device_gemm_dl_f16_f16_f16_km_kn_mn_instance.cpp;
device_gemm_dl_f16_f16_f16_km_nk_mn_instance.cpp;
device_gemm_dl_i8_i8_i8_mk_kn_mn_instance.cpp;
device_gemm_dl_i8_i8_i8_mk_nk_mn_instance.cpp;
device_gemm_dl_i8_i8_i8_km_kn_mn_instance.cpp;
device_gemm_dl_i8_i8_i8_km_nk_mn_instance.cpp;
) )
add_library(device_gemm_instance OBJECT ${DEVICE_GEMM_INSTANCE_SOURCE}) add_library(device_gemm_instance OBJECT ${DEVICE_GEMM_INSTANCE_SOURCE})
target_compile_features(device_gemm_instance PUBLIC) target_compile_features(device_gemm_instance PUBLIC)
set_target_properties(device_gemm_instance PROPERTIES POSITION_INDEPENDENT_CODE ON) set_target_properties(device_gemm_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
clang_tidy_check(device_gemm_instance)
library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_kn_mn_instance.cpp 0 → 100644
View file @ d8f1458f
#include <stdlib.h>
#include "config.hpp"
#include "device_gemm_dl.hpp"
#include "element_wise_operation.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
using device_gemm_dl_f16_f16_f16_km_kn_mn_instances = std::tuple<
// clang-format off
// #########| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
// #########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
// #########| | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| Order| | |
// #########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmDl< F16, F16, F16, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>
// clang-format on
>;
void add_device_gemm_dl_f16_f16_f16_km_kn_mn_instances(
std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
{
add_device_operation_instances(instances, device_gemm_dl_f16_f16_f16_km_kn_mn_instances{});
}
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_km_nk_mn_instance.cpp 0 → 100644
View file @ d8f1458f
#include <stdlib.h>
#include "config.hpp"
#include "device_gemm_dl.hpp"
#include "element_wise_operation.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
using device_gemm_dl_f16_f16_f16_km_nk_mn_instances = std::tuple<
// clang-format off
// #########| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
// #########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
// #########| | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
// #########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmDl< F16, F16, F16, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>
// clang-format on
>;
void add_device_gemm_dl_f16_f16_f16_km_nk_mn_instances(
std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
{
add_device_operation_instances(instances, device_gemm_dl_f16_f16_f16_km_nk_mn_instances{});
}
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_kn_mn_instance.cpp 0 → 100644
View file @ d8f1458f
#include <stdlib.h>
#include "config.hpp"
#include "device_gemm_dl.hpp"
#include "element_wise_operation.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// Compilation parameters for a[m, k] * b[k, n] = c[m, n]
using device_gemm_dl_f16_f16_f16_mk_kn_mn_instances = std::tuple<
// clang-format off
// #########| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
// #########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
// #########| | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
// #########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmDl< F16, F16, F16, F32, Row, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<2, 1, 4, 2>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>
// clang-format on
>;
void add_device_gemm_dl_f16_f16_f16_mk_kn_mn_instances(
std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
{
add_device_operation_instances(instances, device_gemm_dl_f16_f16_f16_mk_kn_mn_instances{});
}
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f16_f16_f16_mk_nk_mn_instance.cpp 0 → 100644
View file @ d8f1458f
#include <stdlib.h>
#include "config.hpp"
#include "device_gemm_dl.hpp"
#include "element_wise_operation.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
using device_gemm_dl_f16_f16_f16_mk_nk_mn_instances =
std::tuple<
// clang-format off
// ########| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
// ########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
// ########| | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
// ########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmDl< F16, F16, F16, F32, Row, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 2, 4, 4, 1, S<8, 2>, S<8, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<8, 1, 1, 2>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 2>, S<1, 2, 0, 3>, S<1, 1, 1, 2>, S<0, 1, 2, 3, 4, 5>, 5, 4>
// clang-format on
>;
void add_device_gemm_dl_f16_f16_f16_mk_nk_mn_instances(
std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
{
add_device_operation_instances(instances, device_gemm_dl_f16_f16_f16_mk_nk_mn_instances{});
}
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_kn_mn_instance.cpp 0 → 100644
View file @ d8f1458f
#include <stdlib.h>
#include "config.hpp"
#include "device_gemm_dl.hpp"
#include "element_wise_operation.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// Compilation parameters for a[k, m] * b[k, n] = c[m, n]
using device_gemm_dl_f32_f32_f32_km_kn_mn_instances = std::tuple<
// clang-format off
// ########| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
// ########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
// ########| | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
// ########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmDl< F32, F32, F32, F32, Col, Row, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<2, 1, 4, 1>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<2, 1, 4, 1>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>
// clang-format on
>;
void add_device_gemm_dl_f32_f32_f32_km_kn_mn_instances(
std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
{
add_device_operation_instances(instances, device_gemm_dl_f32_f32_f32_km_kn_mn_instances{});
}
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm/device_gemm_dl_f32_f32_f32_km_nk_mn_instance.cpp 0 → 100644
View file @ d8f1458f
#include <stdlib.h>
#include "config.hpp"
#include "device_gemm_dl.hpp"
#include "element_wise_operation.hpp"
#include "device_operation_instance.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// Compilation parameters for a[k, m] * b[n, k] = c[m, n]
using device_gemm_dl_f32_f32_f32_km_nk_mn_instances =
std::tuple<
// clang-format off
// ########| AData| BData| CData| AccData| ALayout| BLayout| CLayout| A| B| C| GEMM| Block| MPer| NPer| K0Per| K1| M1Per| N1Per| KPer| M11N11Thread| M11N11Thread| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| BBlockTransfer| CThreadTransfer| CThreadTransfer| CThreadTransfer|
// ########| Type| Type| Type| Type| | | | Elementwise| Elementwise| Elementwise| Spacialization| Size| Block| Block| Block| | ThreadM111| ThreadN111| Thread| ClusterM110Xs| ClusterN110Xs| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| ThreadSliceLengths| ThreadClusterLengths| ThreadCluster| SrcAccess| SrcVectorTensor| SrcVectorTensor| DstVectorTensor| SrcDstAccess| SrcDstVectorDim| DstScalarPerVector|
// ########| | | | | | | | Operation| Operation| Operation| | | | | | | | | | | | K0_M0_M1_K1| K0_M0_M1_K1| ArrangeOrder| Order| Lengths_K0_M0_M1_K1| ContiguousDimOrder| Lengths_K0_M0_M1_K1| K0_N0_N1_K1| K0_N0_N1_K1| ArrangeOrder| Order| Lengths_K0_N0_N1_K1| ContiguousDimOrder| Lengths_K0_N0_N1_K1| Order| | |
// ########| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmDl< F32, F32, F32, F32, Col, Col, Row, PassThrough, PassThrough, PassThrough, GemmDefault, 256, 128, 128, 16, 1, 4, 4, 1, S<8, 2>, S<8, 2>, S<2, 1, 4, 1>, S<8, 1, 32, 1>, S<0, 3, 1, 2>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<0, 3, 1, 2>, S<1, 1, 4, 1>, S<8, 1, 1, 1>, S<2, 1, 128, 1>, S<1, 2, 0, 3>, S<1, 2, 0, 3>, S<4, 1, 1, 1>, S<1, 2, 0, 3>, S<1, 1, 1, 1>, S<0, 1, 2, 3, 4, 5>, 5, 4>
// clang-format on
>;
void add_device_gemm_dl_f32_f32_f32_km_nk_mn_instances(
std::vector<DeviceGemmPtr<PassThrough, PassThrough, PassThrough>>& instances)
{
add_device_operation_instances(instances, device_gemm_dl_f32_f32_f32_km_nk_mn_instances{});
}
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment