Skip to content

GitLab

Commit 6dfb4e78 authored by carlushuang's avatar carlushuang
Browse files

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

parents 397a68f2 1ced00a5
Hide whitespace changes
Inline Side-by-side
Showing with 158 additions and 356 deletions
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_partial_reduce_f32_f32_f32.hpp deleted 100644 → 0
View file @ 397a68f2
#ifndef DEVICE_REDUCE_INSTANCE_MULTIBLOCK_PARTIAL_REDUCE_F32_F32_F32_HPP
#define DEVICE_REDUCE_INSTANCE_MULTIBLOCK_PARTIAL_REDUCE_F32_F32_F32_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(float, float, float, 2, 0, 0, 4, 3); // for MIN
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(float, float, float, 2, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(float, float, float, 2, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(float, float, float, 2, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(float, float, float, 3, 0, 0, 4, 3); // for MAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(float, float, float, 3, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(float, float, float, 3, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(float, float, float, 3, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(float, float, float, 4, 0, 0, 4, 3); // for AMAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(float, float, float, 4, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(float, float, float, 4, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(float, float, float, 4, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(float, float, float, 2, 0, 1, 4, 3); // for MIN
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(float, float, float, 2, 0, 1, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(float, float, float, 2, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(float, float, float, 2, 0, 1, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(float, float, float, 3, 0, 1, 4, 3); // for MAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(float, float, float, 3, 0, 1, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(float, float, float, 3, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(float, float, float, 3, 0, 1, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(float, float, float, 4, 0, 1, 4, 3); // for AMAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(float, float, float, 4, 0, 1, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(float, float, float, 4, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(float, float, float, 4, 0, 1, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(float, float, float, 7, 0, 0, 4, 3); // for NORM2
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(float, float, float, 7, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(float, float, float, 7, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(float, float, float, 7, 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_f32_f64_f32.hpp deleted 100644 → 0
View file @ 397a68f2
#ifndef DEVICE_REDUCE_INSTANCE_MULTIBLOCK_PARTIAL_REDUCE_F32_F64_F32_HPP
#define DEVICE_REDUCE_INSTANCE_MULTIBLOCK_PARTIAL_REDUCE_F32_F64_F32_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(float, double, float, 7, 0, 0, 4, 3); // for NORM2
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(float, double, float, 7, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(float, double, float, 7, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(float, double, float, 7, 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_f64_f64_f64.hpp deleted 100644 → 0
View file @ 397a68f2
#ifndef DEVICE_REDUCE_INSTANCE_MULTIBLOCK_PARTIAL_REDUCE_F64_F64_F64_HPP
#define DEVICE_REDUCE_INSTANCE_MULTIBLOCK_PARTIAL_REDUCE_F64_F64_F64_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(double, double, double, 2, 0, 0, 4, 3); // for MIN
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 2, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 2, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 2, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 3, 0, 0, 4, 3); // for MAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 3, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 3, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 3, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 4, 0, 0, 4, 3); // for AMAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 4, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 4, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 4, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 2, 0, 1, 4, 3); // for MIN
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 2, 0, 1, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 2, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 2, 0, 1, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 3, 0, 1, 4, 3); // for MAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 3, 0, 1, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 3, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 3, 0, 1, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 4, 0, 1, 4, 3); // for AMAX
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 4, 0, 1, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 4, 0, 1, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 4, 0, 1, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 7, 0, 0, 4, 3); // for NORM2
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 7, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 7, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 7, 0, 0, 2, 1);
// Will be moved to use MultiBlockAtomicAdd
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 0, 0, 0, 4, 3); // for ADD
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 0, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 0, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 0, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 5, 0, 0, 4, 3); // for AVG
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 5, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 5, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(double, double, double, 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_i32_i8.hpp deleted 100644 → 0
View file @ 397a68f2
#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 @ 397a68f2
#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 @ 6dfb4e78
......@@ -58,8 +58,8 @@ template <typename InDataType,
int Rank,
int NumReduceDim,
ReduceTensorOp ReduceOpId,
NanPropagation NanOpt,
ReduceTensorIndices IndicesOpt>
bool PropagateNan,
bool UseIndex>
void add_device_reduce_instance_threadwise(
std::vector<deviceReduceThreadWisePtrType<AccDataType, ReduceOpId>>& device_op_instances)
{
......@@ -73,9 +73,7 @@ void add_device_reduce_instance_threadwise(
constexpr bool Indexable =
(ReduceOpId == ReduceTensorOp::MIN || ReduceOpId == ReduceTensorOp::MAX ||
ReduceOpId == ReduceTensorOp::AMAX);
constexpr bool NeedIndices = Indexable && (IndicesOpt != ReduceTensorIndices::NO_INDICES);
constexpr bool PropagateNan = (NanOpt == NanPropagation::NOT_PROPAGATE_NAN) ? false : true;
constexpr bool OutputIndex = Indexable && UseIndex;
using cfg1 = ReductionConfiguration_1<256, 256, 1>;
......@@ -93,10 +91,9 @@ void add_device_reduce_instance_threadwise(
InElementwiseOperation,
AccElementwiseOperation,
PropagateNan,
NeedIndices,
OutputIndex,
false, // HaveIndexInputIfOutputIndex
cfg1::BlockSize_,
cfg1::MThreadClusterSize_,
cfg1::KThreadClusterSize_,
cfg2::MThreadSliceSize_,
cfg2::KThreadSliceSize_,
cfg2::InSrcVectorDim_,
......@@ -107,54 +104,54 @@ void add_device_reduce_instance_threadwise(
});
};
#define ADD_THREADWISE_INST_BY_TYPE( \
inT, compT, outT, ReduceOpId, NanOpt, IndicesOpt, Rank, NumReduceDim) \
template void add_device_reduce_instance_threadwise<inT, \
compT, \
outT, \
Rank, \
NumReduceDim, \
ReduceOpId, \
NanOpt, \
IndicesOpt>( \
#define ADD_THREADWISE_INST_BY_TYPE( \
inT, compT, outT, ReduceOpId, PropagateNan, UseIndex, Rank, NumReduceDim) \
template void add_device_reduce_instance_threadwise<inT, \
compT, \
outT, \
Rank, \
NumReduceDim, \
ReduceOpId, \
PropagateNan, \
UseIndex>( \
std::vector<deviceReduceThreadWisePtrType<compT, ReduceOpId>> & device_op_instances)
#define ADD_THREADWISE_INST_BY_ID( \
inT, compT, outT, ReduceOpId, NanOpt, IndicesOpt, Rank, NumReduceDim) \
ADD_THREADWISE_INST_BY_TYPE(inT, \
compT, \
outT, \
static_cast<ReduceTensorOp>(ReduceOpId), \
static_cast<NanPropagation>(NanOpt), \
static_cast<ReduceTensorIndices>(IndicesOpt), \
Rank, \
#define ADD_THREADWISE_INST_BY_ID( \
inT, compT, outT, ReduceOpId, NanOpt, IndicesOpt, Rank, NumReduceDim) \
ADD_THREADWISE_INST_BY_TYPE(inT, \
compT, \
outT, \
static_cast<ReduceTensorOp>(ReduceOpId), \
static_cast<bool>(NanOpt), \
static_cast<bool>(IndicesOpt), \
Rank, \
NumReduceDim)
#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, \
compT, \
outT, \
Rank, \
NumReduceDim, \
ReduceOpId, \
NanOpt, \
IndicesOpt>( \
PropagateNan, \
UseIndex>( \
std::vector<DeviceReducePtr< \
typename reduce_unary_operator<compT, ReduceOpId, true, true>::InElementwiseOperation, \
typename reduce_unary_operator<compT, ReduceOpId, true, true>:: \
AccElementwiseOperation>> & \
device_op_instances)
#define ADD_THREADWISE_INST_REF_BY_ID( \
inT, compT, outT, ReduceOpId, NanOpt, IndicesOpt, Rank, NumReduceDim) \
ADD_THREADWISE_INST_REF_BY_TYPE(inT, \
compT, \
outT, \
static_cast<ReduceTensorOp>(ReduceOpId), \
static_cast<NanPropagation>(NanOpt), \
static_cast<ReduceTensorIndices>(IndicesOpt), \
Rank, \
#define ADD_THREADWISE_INST_REF_BY_ID( \
inT, compT, outT, ReduceOpId, NanOpt, IndicesOpt, Rank, NumReduceDim) \
ADD_THREADWISE_INST_REF_BY_TYPE(inT, \
compT, \
outT, \
static_cast<ReduceTensorOp>(ReduceOpId), \
static_cast<bool>(NanOpt), \
static_cast<bool>(IndicesOpt), \
Rank, \
NumReduceDim)
} // namespace device_reduce_instance
......
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16.hpp
View file @ 6dfb4e78
#ifndef DEVICE_REDUCE_INSTANCE_THREADWISE_B16_F32_B16_HPP
#define DEVICE_REDUCE_INSTANCE_THREADWISE_B16_F32_B16_HPP
#include "reduction_enums.hpp"
#include "reduction_operator_mapping.hpp"
#include "data_type.hpp"
#include "device_reduce_instance_threadwise.hpp"
namespace ck {
......
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16.hpp
View file @ 6dfb4e78
#ifndef DEVICE_REDUCE_INSTANCE_THREADWISE_F16_F16_F16_HPP
#define DEVICE_REDUCE_INSTANCE_THREADWISE_F16_F16_F16_HPP
#include "reduction_enums.hpp"
#include "reduction_operator_mapping.hpp"
#include "data_type.hpp"
#include "device_reduce_instance_threadwise.hpp"
namespace ck {
......
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16.hpp
View file @ 6dfb4e78
#ifndef DEVICE_REDUCE_INSTANCE_THREADWISE_F16_F32_F16_HPP
#define DEVICE_REDUCE_INSTANCE_THREADWISE_F16_F32_F16_HPP
#include "reduction_enums.hpp"
#include "reduction_operator_mapping.hpp"
#include "data_type.hpp"
#include "device_reduce_instance_threadwise.hpp"
namespace ck {
......
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32.hpp
View file @ 6dfb4e78
#ifndef 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"
namespace ck {
......
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32.hpp
View file @ 6dfb4e78
#ifndef 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"
namespace ck {
......
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64.hpp
View file @ 6dfb4e78
#ifndef 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"
namespace ck {
......
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8.hpp
View file @ 6dfb4e78
#ifndef 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"
namespace ck {
......
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8.hpp
View file @ 6dfb4e78
#ifndef 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"
namespace ck {
......
library/include/ck/library/utility/check_err.hpp
View file @ 6dfb4e78
......@@ -24,7 +24,7 @@ check_err(const std::vector<T>& out,
const std::vector<T>& ref,
const std::string& msg = "Error: Incorrect results!",
double rtol = 1e-5,
double atol = 1e-8)
double atol = 3e-6)
{
if(out.size() != ref.size())
{
......@@ -173,8 +173,8 @@ check_err(const std::vector<T>& out,
{
if(out[i] != ref[i])
{
std::cout << "out[" << i << "] != ref[" << i << "]: " << out[i] << " != " << ref[i]
<< std::endl
std::cout << "out[" << i << "] != ref[" << i << "]: " << static_cast<int>(out[i])
<< " != " << static_cast<int>(ref[i]) << std::endl
<< msg << std::endl;
return false;
}
......
library/src/tensor_operation_instance/gpu/CMakeLists.txt
View file @ 6dfb4e78
include_directories(BEFORE
${PROJECT_SOURCE_DIR}/include/ck
${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
${PROJECT_SOURCE_DIR}/include/ck/problem_transform
......
library/src/tensor_operation_instance/gpu/batched_gemm_reduce/CMakeLists.txt
View file @ 6dfb4e78
set(DEVICE_BATCHED_GEMM_REDUCE_INSTANCE_SOURCE
set(DEVICE_BATCHED_GEMM_REDUCE_INSTANCE_SOURCE
device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp
device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gnk_gmn_instance.cpp
device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp
......
library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instance.cpp
View file @ 6dfb4e78
......@@ -10,8 +10,9 @@ namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using F16 = ck::half_t;
using F32 = float;
using DPtrsGlobal = ck::Tuple<F32*, F32*>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
......@@ -21,41 +22,52 @@ using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::reduce::Add<F32>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using ReduceOps = ck::Tuple<ReduceSum, ReduceSum>;
using Identity = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Identity, Identity>;
using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
ck::InMemoryDataOperationEnum::AtomicAdd>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// c[g, m, n] = a[g, m, k] * b[g, n, k]
// d0[g, m] = reduce0(c[g, m, n])
// d1[g, m] = reduce1(c[g, m, n])
using device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instances =
std::tuple<
// clang-format off
//##################################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| D0| D1| D1EleOp| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//##################################| | | | Type| Type| Type| DataType| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//##################################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 4, 4, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 2, 2, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 2, 2, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 2, 2, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>
//##################################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsAccEleOp| D| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//##################################| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData|Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//##################################| | | | | | | | | | | Operation| Operation| Operation| Operation| | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 4, 4, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 2, 2, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 2, 2, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 2, 2, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>
// clang-format on
>;
void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gkn_gmn_instances(
std::vector<DeviceGemmReducePtr<PassThrough, PassThrough, PassThrough, Square>>& instances)
std::vector<DeviceGemmReducePtr<DPtrsGlobal,
PassThrough,
PassThrough,
PassThrough,
DInElementOps,
DOutElementOps>>& instances)
{
add_device_operation_instances(
instances,
......
library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instance.cpp
View file @ 6dfb4e78
......@@ -10,8 +10,9 @@ namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using F16 = ck::half_t;
using F32 = float;
using DPtrsGlobal = ck::Tuple<F32*, F32*>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
......@@ -21,41 +22,52 @@ using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::reduce::Add<F32>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using ReduceOps = ck::Tuple<ReduceSum, ReduceSum>;
using Identity = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Identity, Identity>;
using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
ck::InMemoryDataOperationEnum::AtomicAdd>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// c[g, m, n] = a[g, m, k] * b[g, n, k]
// d0[g, m] = reduce0(c[g, m, n])
// d1[g, m] = reduce1(c[g, m, n])
using device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instances =
std::tuple<
// clang-format off
//##################################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| D0| D1| D1EleOp| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//##################################| | | | Type| Type| Type| DataType| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//##################################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 2, 8, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 2, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 2, 8, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 2, 8, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>
//##################################| ALayout| BLayout| CLayout|AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsAccEleOp| D| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//##################################| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData|Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//##################################| | | | | | | | | | | Operation| Operation| Operation| Operation| | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 2, 8, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 2, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 2, 8, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 2, 8, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Col, Col, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>
// clang-format on
>;
void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gkm_gnk_gmn_instances(
std::vector<DeviceGemmReducePtr<PassThrough, PassThrough, PassThrough, Square>>& instances)
std::vector<DeviceGemmReducePtr<DPtrsGlobal,
PassThrough,
PassThrough,
PassThrough,
DInElementOps,
DOutElementOps>>& instances)
{
add_device_operation_instances(
instances,
......
library/src/tensor_operation_instance/gpu/batched_gemm_reduce/device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instance.cpp
View file @ 6dfb4e78
......@@ -10,8 +10,9 @@ namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using F16 = ck::half_t;
using F32 = float;
using DPtrsGlobal = ck::Tuple<F32*, F32*>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
......@@ -21,41 +22,52 @@ using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using ReduceSum = ck::reduce::Add<F32>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using ReduceOps = ck::Tuple<ReduceSum, ReduceSum>;
using Identity = ck::tensor_operation::element_wise::UnaryIdentic<F32, F32, false>;
using Square = ck::tensor_operation::element_wise::UnarySquare<F32, F32, false>;
using DInElementOps = ck::Tuple<Identity, Square>;
using DOutElementOps = ck::Tuple<Identity, Identity>;
using ReduceMemOp = ck::InMemoryDataOperationEnumSequence<ck::InMemoryDataOperationEnum::AtomicAdd,
ck::InMemoryDataOperationEnum::AtomicAdd>;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// c[g, m, n] = a[g, m, k] * b[g, n, k]
// d0[g, m] = reduce0(c[g, m, n])
// d1[g, m] = reduce1(c[g, m, n])
using device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances =
std::tuple<
// clang-format off
//##################################| ALayout| BLayout| CLayout| AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| D0| D1| D1EleOp| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//##################################| | | | Type| Type| Type| DataType| DataType| DataType| Type| Elementwise| Elementwise| Elementwise| Reduce| Reduce| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//##################################| | | | | | | | | | | Operation| Operation| Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 8, 2, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 8, 2, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 8, 2, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, F32, PassThrough, PassThrough, PassThrough, ReduceSum, ReduceSum, Square, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>
//##################################| ALayout| BLayout| CLayout| AData| BData| CData| GemmAcc| CShuffle| ReduceAcc| DData| A| B| C| Dxs| DxsInEleOp| DxsAccEleOp| D| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer| CReduce| CReduceThreadLds2VGprCopy| CReduceThreadVgpr2GlobalCopy|
//##################################| | | | Type| Type| Type| DataType| DataType| DataType| Type Tuple| Elementwise| Elementwise| Elementwise| Reduce| | | MemoryData|Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MPerBlock| ScalarPerVector| ThreadClusterLengths| SrcDstScalarPerVector| SrcDstScalarPerVector|
//##################################| | | | | | | | | | | Operation| Operation| Operation| Operation| | | Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NPerBlock| _NPerBlock| _MPerBlock_NPerBlock| _NPerBlock| _MPerBlock|
//##################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 8, 2, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 4>, 8, S<64, 2>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, true, 1, 1, S<1, 16, 1, 8>, 8, S<32, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 8, 2, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, true, 1, 1, S<1, 16, 1, 4>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 8, 2, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, false, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>,
DeviceBatchedGemmReduce_Xdl_CShuffle< Row, Row, Row, F16, F16, F16, F32, F32, F32, DPtrsGlobal, PassThrough, PassThrough, PassThrough, ReduceOps, DInElementOps, DOutElementOps, ReduceMemOp, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, true, 1, 1, S<1, 32, 1, 8>, 8, S<64, 4>, 4, 1>
// clang-format on
>;
void add_device_batched_gemm_reduce_xdl_cshuffle_f16_f16_f16_f32_f32_gmk_gkn_gmn_instances(
std::vector<DeviceGemmReducePtr<PassThrough, PassThrough, PassThrough, Square>>& instances)
std::vector<DeviceGemmReducePtr<DPtrsGlobal,
PassThrough,
PassThrough,
PassThrough,
DInElementOps,
DOutElementOps>>& instances)
{
add_device_operation_instances(
instances,
......
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