Skip to content

GitLab

Commit a3b4c5cb authored by wangshaojie6's avatar wangshaojie6
Browse files

merge develop branch and add gridwise pipeline v3

parents 48918ab9 1677cf70
Hide whitespace changes
Inline Side-by-side
Showing with 123 additions and 373 deletions
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_partial_reduce_f16_f32_f16.hpp deleted 100644 → 0
View file @ 48918ab9
#ifndef DEVICE_REDUCE_INSTANCE_MULTIBLOCK_PARTIAL_REDUCE_F16_F32_F16_HPP
#define DEVICE_REDUCE_INSTANCE_MULTIBLOCK_PARTIAL_REDUCE_F16_F32_F16_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(half_t, float, half_t, 0, 0, 0, 4, 3); // for ADD
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(half_t, float, half_t, 0, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(half_t, float, half_t, 0, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(half_t, float, half_t, 0, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(half_t, float, half_t, 5, 0, 0, 4, 3); // for AVG
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(half_t, float, half_t, 5, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(half_t, float, half_t, 5, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(half_t, float, half_t, 5, 0, 0, 2, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(half_t, float, half_t, 7, 0, 0, 4, 3); // for NORM2
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(half_t, float, half_t, 7, 0, 0, 4, 4);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(half_t, float, half_t, 7, 0, 0, 4, 1);
ADD_MULTIBLOCK_PARTIAL_REDUCE_INST_REF_BY_ID(half_t, float, half_t, 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_f32_f32.hpp deleted 100644 → 0
View file @ 48918ab9
#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 @ 48918ab9
#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 @ 48918ab9
#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 @ 48918ab9
#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 @ 48918ab9
#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 @ a3b4c5cb
......@@ -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 @ a3b4c5cb
#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 @ a3b4c5cb
#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 @ a3b4c5cb
#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 @ a3b4c5cb
#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 @ a3b4c5cb
#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 @ a3b4c5cb
#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 @ a3b4c5cb
#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 @ a3b4c5cb
#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 @ a3b4c5cb
......@@ -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/include/ck/library/utility/conv_fwd_util.hpp → library/include/ck/library/utility/conv_util.hpp
View file @ a3b4c5cb
......@@ -146,19 +146,19 @@ struct ConvParams
const std::vector<ck::index_t>& left_pads,
const std::vector<ck::index_t>& right_pads);
ck::index_t num_dim_spatial;
ck::index_t N;
ck::index_t K;
ck::index_t C;
ck::index_t num_dim_spatial_;
ck::index_t N_;
ck::index_t K_;
ck::index_t C_;
std::vector<ck::index_t> filter_spatial_lengths;
std::vector<ck::index_t> input_spatial_lengths;
std::vector<ck::index_t> filter_spatial_lengths_;
std::vector<ck::index_t> input_spatial_lengths_;
std::vector<ck::index_t> conv_filter_strides;
std::vector<ck::index_t> conv_filter_dilations;
std::vector<ck::index_t> conv_filter_strides_;
std::vector<ck::index_t> conv_filter_dilations_;
std::vector<ck::index_t> input_left_pads;
std::vector<ck::index_t> input_right_pads;
std::vector<ck::index_t> input_left_pads_;
std::vector<ck::index_t> input_right_pads_;
std::vector<ck::index_t> GetOutputSpatialLengths() const;
};
......@@ -268,10 +268,10 @@ void run_reference_convolution_forward(const ConvParams& params,
auto ref_argument = ref_conv.MakeArgument(input,
weights,
output,
params.conv_filter_strides,
params.conv_filter_dilations,
params.input_left_pads,
params.input_right_pads,
params.conv_filter_strides_,
params.conv_filter_dilations_,
params.input_left_pads_,
params.input_right_pads_,
PassThrough{},
PassThrough{},
PassThrough{});
......@@ -437,17 +437,17 @@ class ConvFwdOpInstance : public ck::utils::OpInstance<OutDataType, InDataType,
virtual InTensorsTuple GetInputTensors() const override
{
std::vector<std::size_t> input_dims{static_cast<std::size_t>(params_.N),
static_cast<std::size_t>(params_.C)};
std::vector<std::size_t> input_dims{static_cast<std::size_t>(params_.N_),
static_cast<std::size_t>(params_.C_)};
input_dims.insert(std::end(input_dims),
std::begin(params_.input_spatial_lengths),
std::end(params_.input_spatial_lengths));
std::begin(params_.input_spatial_lengths_),
std::end(params_.input_spatial_lengths_));
std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params_.K),
static_cast<std::size_t>(params_.C)};
std::vector<std::size_t> filter_dims{static_cast<std::size_t>(params_.K_),
static_cast<std::size_t>(params_.C_)};
filter_dims.insert(std::end(filter_dims),
std::begin(params_.filter_spatial_lengths),
std::end(params_.filter_spatial_lengths));
std::begin(params_.filter_spatial_lengths_),
std::end(params_.filter_spatial_lengths_));
auto input = std::make_unique<Tensor<InDataType>>(
get_host_tensor_descriptor(input_dims, InLayout{}));
......@@ -465,8 +465,8 @@ class ConvFwdOpInstance : public ck::utils::OpInstance<OutDataType, InDataType,
virtual TensorPtr<OutDataType> GetOutputTensor() const override
{
std::vector<std::size_t> output_dims{static_cast<std::size_t>(params_.N),
static_cast<std::size_t>(params_.K)};
std::vector<std::size_t> output_dims{static_cast<std::size_t>(params_.N_),
static_cast<std::size_t>(params_.K_)};
output_dims.insert(std::end(output_dims),
std::begin(output_spatial_lengths_),
std::end(output_spatial_lengths_));
......@@ -522,16 +522,16 @@ class ConvFwdOpInstance : public ck::utils::OpInstance<OutDataType, InDataType,
static_cast<InDataType*>(in_device_buffers[0]->GetDeviceBuffer()),
static_cast<WeiDataType*>(in_device_buffers[1]->GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buffer->GetDeviceBuffer()),
params_.N,
params_.K,
params_.C,
params_.input_spatial_lengths,
params_.filter_spatial_lengths,
params_.N_,
params_.K_,
params_.C_,
params_.input_spatial_lengths_,
params_.filter_spatial_lengths_,
output_spatial_lengths_,
params_.conv_filter_strides,
params_.conv_filter_dilations,
params_.input_left_pads,
params_.input_right_pads,
params_.conv_filter_strides_,
params_.conv_filter_dilations_,
params_.input_left_pads_,
params_.input_right_pads_,
InElementwiseOp{},
WeiElementwiseOp{},
OutElementwiseOp{});
......@@ -539,20 +539,20 @@ class ConvFwdOpInstance : public ck::utils::OpInstance<OutDataType, InDataType,
virtual std::size_t GetFlops() const override
{
return get_flops(params_.N,
params_.C,
params_.K,
params_.filter_spatial_lengths,
return get_flops(params_.N_,
params_.C_,
params_.K_,
params_.filter_spatial_lengths_,
output_spatial_lengths_);
}
virtual std::size_t GetBtype() const override
{
return get_btype<InDataType, WeiDataType, OutDataType>(params_.N,
params_.C,
params_.K,
params_.input_spatial_lengths,
params_.filter_spatial_lengths,
return get_btype<InDataType, WeiDataType, OutDataType>(params_.N_,
params_.C_,
params_.K_,
params_.input_spatial_lengths_,
params_.filter_spatial_lengths_,
output_spatial_lengths_);
}
......
library/include/ck/library/utility/op_instance_engine.hpp
View file @ a3b4c5cb
......@@ -128,7 +128,7 @@ class OpInstanceRunEngine
template <typename OpInstancePtr>
ProfileBestConfig Profile(const std::vector<OpInstancePtr>& op_ptrs,
int nrepeat = 100,
bool time_kernel = false,
bool do_verification = false,
bool do_log = false)
{
......@@ -143,7 +143,7 @@ class OpInstanceRunEngine
if(op_ptr->IsSupportedArgument(argument.get()))
{
std::string op_name = op_ptr->GetTypeString();
float avg_time = invoker->Run(argument.get(), nrepeat);
float avg_time = invoker->Run(argument.get(), StreamConfig{nullptr, time_kernel});
std::size_t flops = op_instance_.GetFlops();
std::size_t num_btype = op_instance_.GetBtype();
......
library/src/host_tensor/CMakeLists.txt
View file @ a3b4c5cb
......@@ -10,10 +10,31 @@ set(HOST_TENSOR_SOURCE
host_tensor.cpp
)
add_library(host_tensor SHARED ${HOST_TENSOR_SOURCE})
add_library(host_tensor STATIC ${HOST_TENSOR_SOURCE})
add_library(composable_kernel::host_tensor ALIAS host_tensor)
target_compile_features(host_tensor PUBLIC)
set_target_properties(host_tensor PROPERTIES POSITION_INDEPENDENT_CODE ON)
target_include_directories(host_tensor SYSTEM PUBLIC $<BUILD_INTERFACE:${HALF_INCLUDE_DIR}>)
install(TARGETS host_tensor LIBRARY DESTINATION lib)
target_include_directories(host_tensor PUBLIC
"$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck>"
"$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/utility>"
"$<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/ck/library/host_tensor>"
)
install(TARGETS host_tensor
EXPORT host_tensorTargets
LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}
ARCHIVE DESTINATION ${CMAKE_INSTALL_LIBDIR}
RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR}
INCLUDES DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}
)
install(EXPORT host_tensorTargets
FILE composable_kernelhost_tensorTargets.cmake
NAMESPACE composable_kernel::
DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/composable_kernel
)
clang_tidy_check(host_tensor)
library/src/host_tensor/device.cpp
View file @ a3b4c5cb
......@@ -2,7 +2,7 @@
DeviceMem::DeviceMem(std::size_t mem_size) : mMemSize(mem_size)
{
hipGetErrorString(hipMalloc(static_cast<void**>(&mpDeviceBuf), mMemSize));
hip_check_error(hipMalloc(static_cast<void**>(&mpDeviceBuf), mMemSize));
}
void* DeviceMem::GetDeviceBuffer() { return mpDeviceBuf; }
......@@ -11,49 +11,48 @@ std::size_t DeviceMem::GetBufferSize() { return mMemSize; }
void DeviceMem::ToDevice(const void* p)
{
hipGetErrorString(
hipMemcpy(mpDeviceBuf, const_cast<void*>(p), mMemSize, hipMemcpyHostToDevice));
hip_check_error(hipMemcpy(mpDeviceBuf, const_cast<void*>(p), mMemSize, hipMemcpyHostToDevice));
}
void DeviceMem::FromDevice(void* p)
{
hipGetErrorString(hipMemcpy(p, mpDeviceBuf, mMemSize, hipMemcpyDeviceToHost));
hip_check_error(hipMemcpy(p, mpDeviceBuf, mMemSize, hipMemcpyDeviceToHost));
}
void DeviceMem::SetZero() { hipGetErrorString(hipMemset(mpDeviceBuf, 0, mMemSize)); }
void DeviceMem::SetZero() { hip_check_error(hipMemset(mpDeviceBuf, 0, mMemSize)); }
DeviceMem::~DeviceMem() { hipGetErrorString(hipFree(mpDeviceBuf)); }
DeviceMem::~DeviceMem() { hip_check_error(hipFree(mpDeviceBuf)); }
struct KernelTimerImpl
{
KernelTimerImpl()
{
hipGetErrorString(hipEventCreate(&mStart));
hipGetErrorString(hipEventCreate(&mEnd));
hip_check_error(hipEventCreate(&mStart));
hip_check_error(hipEventCreate(&mEnd));
}
~KernelTimerImpl()
{
hipGetErrorString(hipEventDestroy(mStart));
hipGetErrorString(hipEventDestroy(mEnd));
hip_check_error(hipEventDestroy(mStart));
hip_check_error(hipEventDestroy(mEnd));
}
void Start()
{
hipGetErrorString(hipDeviceSynchronize());
hipGetErrorString(hipEventRecord(mStart, nullptr));
hip_check_error(hipDeviceSynchronize());
hip_check_error(hipEventRecord(mStart, nullptr));
}
void End()
{
hipGetErrorString(hipEventRecord(mEnd, nullptr));
hipGetErrorString(hipEventSynchronize(mEnd));
hip_check_error(hipEventRecord(mEnd, nullptr));
hip_check_error(hipEventSynchronize(mEnd));
}
float GetElapsedTime() const
{
float time;
hipGetErrorString(hipEventElapsedTime(&time, mStart, mEnd));
hip_check_error(hipEventElapsedTime(&time, mStart, mEnd));
return time;
}
......
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