Skip to content

GitLab

Commit bb1f8082 authored by root's avatar root
Browse files

Merge remote-tracking branch 'origin/develop' into myamlak/cgemm

parents 97ac5007 82d7d993
Hide whitespace changes
Inline Side-by-side
Showing with 105 additions and 622 deletions
library/include/ck/library/host_tensor/host_reduce_util.hpp
View file @ bb1f8082
......@@ -28,9 +28,7 @@
#include <limits>
#include <cmath>
#include <cassert>
#include <stdexcept>
#include <string>
#include <functional>
#include "reduction_enums.hpp"
#include "data_type.hpp"
......@@ -214,13 +212,13 @@ binop_with_nan_check(std::function<void(AccDataType&, AccDataType)> opReduce,
};
};
template <typename AccDataType, bool PropagateNan>
template <typename AccDataType, typename IndexDataType, bool PropagateNan>
__host__ static inline void
binop_with_nan_check2(std::function<void(AccDataType&, AccDataType, bool&)> opReduce,
AccDataType& accuVal,
AccDataType currVal,
int& accuIndex,
int currIndex)
binop_with_index_and_nan_check(std::function<void(AccDataType&, AccDataType, bool&)> opReduce,
AccDataType& accuVal,
AccDataType currVal,
IndexDataType& accuIndex,
IndexDataType currIndex)
{
using ck::math::isnan;
......@@ -254,16 +252,6 @@ binop_with_nan_check2(std::function<void(AccDataType&, AccDataType, bool&)> opRe
}; // namespace host_reduce
static inline std::vector<int> to_int_vector(const std::vector<size_t>& inData)
{
std::vector<int> outData;
for(auto elem : inData)
outData.push_back(static_cast<int>(elem));
return (outData);
};
}; // namespace ck
#endif
library/include/ck/library/host_tensor/host_reduction.hpp
View file @ bb1f8082
......@@ -34,6 +34,7 @@
#include "reduction_enums.hpp"
#include "reduction_common.hpp"
#include "host_reduce_util.hpp"
#include "host_common_util.hpp"
#include "host_tensor.hpp"
#include "data_type.hpp"
......@@ -200,7 +201,7 @@ struct ReductionHost
using ck::float_equal_one;
using ck::float_equal_zero;
using ck::type_convert;
using ck::host_reduce::binop_with_nan_check2;
using ck::host_reduce::binop_with_index_and_nan_check;
using ck::host_reduce::ReduceOpFn2;
using ck::host_reduce::ReduceOpZeroVal;
......@@ -211,8 +212,7 @@ struct ReductionHost
AccDataType accuVal = ReduceOpZeroVal<AccDataType, ReduceOpId>();
IndexDataType accuIndex = 0;
for(IndexDataType i = 0; i < ck::type_convert<IndexDataType>(reduce_dim_indexes.size());
i++)
for(std::size_t i = 0; i < reduce_dim_indexes.size(); i++)
{
auto offset_reduce =
get_offset_from_index<NumReduceDim>(reduceStrides, reduce_dim_indexes[i]);
......@@ -221,9 +221,9 @@ struct ReductionHost
preUnaryOp(currVal);
auto currIndex = i;
auto currIndex = static_cast<IndexDataType>(i);
binop_with_nan_check2<AccDataType, PropagateNan>(
binop_with_index_and_nan_check<AccDataType, IndexDataType, PropagateNan>(
opReduce2, accuVal, currVal, accuIndex, currIndex);
};
......@@ -247,9 +247,7 @@ struct ReductionHost
auto offset_invariant =
get_offset_from_index<NumInvariantDim>(invariantStrides, invariant_index);
for(IndexDataType i = 0;
i < ck::type_convert<IndexDataType>(reduce_dim_indexes.size());
i++)
for(std::size_t i = 0; i < reduce_dim_indexes.size(); i++)
{
auto offset_reduce =
get_offset_from_index<NumReduceDim>(reduceStrides, reduce_dim_indexes[i]);
......@@ -259,9 +257,9 @@ struct ReductionHost
preUnaryOp(currVal);
auto currIndex = i;
auto currIndex = static_cast<IndexDataType>(i);
binop_with_nan_check2<AccDataType, PropagateNan>(
binop_with_index_and_nan_check<AccDataType, IndexDataType, PropagateNan>(
opReduce2, accuVal, currVal, accuIndex, currIndex);
};
......
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance.hpp
View file @ bb1f8082
......@@ -9,26 +9,11 @@
#include "device_reduce_instance_blockwise_i8_i8_i8.hpp"
#include "device_reduce_instance_blockwise_i8_i32_i8.hpp"
#include "device_reduce_instance_blockwise_b16_f32_b16.hpp"
#include "device_reduce_instance_blockwise_second_call_f16_f16_f16.hpp"
#include "device_reduce_instance_blockwise_second_call_f32_f32_f16.hpp"
#include "device_reduce_instance_blockwise_second_call_f32_f32_f32.hpp"
#include "device_reduce_instance_blockwise_second_call_f64_f64_f32.hpp"
#include "device_reduce_instance_blockwise_second_call_f64_f64_f64.hpp"
#include "device_reduce_instance_blockwise_second_call_i8_i8_i8.hpp"
#include "device_reduce_instance_blockwise_second_call_i32_i32_i8.hpp"
#include "device_reduce_instance_blockwise_second_call_f32_f32_b16.hpp"
#include "device_reduce_instance_multiblock_atomic_add_f16_f32_f32.hpp"
#include "device_reduce_instance_multiblock_atomic_add_f32_f32_f32.hpp"
#include "device_reduce_instance_multiblock_atomic_add_f32_f64_f32.hpp"
#include "device_reduce_instance_multiblock_atomic_add_f64_f64_f64.hpp"
#include "device_reduce_instance_multiblock_atomic_add_b16_f32_f32.hpp"
#include "device_reduce_instance_multiblock_partial_reduce_f16_f16_f16.hpp"
#include "device_reduce_instance_multiblock_partial_reduce_f16_f32_f16.hpp"
#include "device_reduce_instance_multiblock_partial_reduce_f32_f32_f32.hpp"
#include "device_reduce_instance_multiblock_partial_reduce_f32_f64_f32.hpp"
#include "device_reduce_instance_multiblock_partial_reduce_f64_f64_f64.hpp"
#include "device_reduce_instance_multiblock_partial_reduce_i8_i8_i8.hpp"
#include "device_reduce_instance_multiblock_partial_reduce_i8_i32_i8.hpp"
#include "device_reduce_instance_multiblock_partial_reduce_b16_f32_b16.hpp"
#include "device_reduce_instance_threadwise_f16_f16_f16.hpp"
#include "device_reduce_instance_threadwise_f16_f32_f16.hpp"
#include "device_reduce_instance_threadwise_f32_f32_f32.hpp"
......
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp
View file @ bb1f8082
......@@ -3,13 +3,27 @@
#include "reduction_operator_mapping.hpp"
#include "device_reduce_instance_impl_common.hpp"
#include "device_reduce_blockwise.hpp"
#include "device_reduce_multiblock.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_reduce_instance {
using reduce_configuration_1_instances_blockwise = std::tuple<
// clang-format off
// BlockSize | MThreadClusterSize | KThreadClusterSize
ReductionConfiguration_1<256, 128, 2>,
ReductionConfiguration_1<256, 64, 4>,
ReductionConfiguration_1<256, 32, 8>,
ReductionConfiguration_1<256, 16, 16>,
ReductionConfiguration_1<256, 8, 32>,
ReductionConfiguration_1<256, 4, 64>,
ReductionConfiguration_1<256, 2, 128>,
ReductionConfiguration_1<256, 1, 256>
// clang-format on
>;
#ifdef QUICK_REDUCE_TEST
using reduce_configuration_2_instances_blockwise = std::tuple<
// clang-format off
......@@ -58,8 +72,8 @@ template <typename InDataType,
int Rank,
int NumReduceDim,
ReduceTensorOp ReduceOpId,
NanPropagation NanOpt,
ReduceTensorIndices IndicesOpt>
bool PropagateNan,
bool UseIndex>
void add_device_reduce_instance_blockwise(
std::vector<deviceReduceBlockWisePtrType<AccDataType, ReduceOpId>>& device_op_instances)
{
......@@ -73,92 +87,94 @@ void add_device_reduce_instance_blockwise(
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;
static_for<0, std::tuple_size<reduce_configuration_1_instances>::value, 1>{}([&](auto i) {
using cfg1 =
remove_cvref_t<decltype(std::get<i.value>(reduce_configuration_1_instances{}))>;
static_for<0, std::tuple_size<reduce_configuration_2_instances_blockwise>::value, 1>{}(
[&](auto j) {
using cfg2 = remove_cvref_t<decltype(
std::get<j.value>(reduce_configuration_2_instances_blockwise{}))>;
using ReduceOpInstance = DeviceReduceBlockWise<InDataType,
AccDataType,
OutDataType,
Rank,
NumReduceDim,
ReduceOperation,
InElementwiseOperation,
AccElementwiseOperation,
PropagateNan,
NeedIndices,
cfg1::BlockSize_,
cfg1::MThreadClusterSize_,
cfg1::KThreadClusterSize_,
cfg2::MThreadSliceSize_,
cfg2::KThreadSliceSize_,
cfg2::InSrcVectorDim_,
cfg2::InSrcVectorSize_,
cfg2::OutDstVectorSize_>;
device_op_instances.push_back(
std::make_unique<ReduceOpInstance>(ReduceOpInstance{}));
});
});
constexpr bool OutputIndex = Indexable && UseIndex;
static_for<0, std::tuple_size<reduce_configuration_1_instances_blockwise>::value, 1>{}(
[&](auto i) {
using cfg1 = remove_cvref_t<decltype(
std::get<i.value>(reduce_configuration_1_instances_blockwise{}))>;
static_for<0, std::tuple_size<reduce_configuration_2_instances_blockwise>::value, 1>{}(
[&](auto j) {
using cfg2 = remove_cvref_t<decltype(
std::get<j.value>(reduce_configuration_2_instances_blockwise{}))>;
using ReduceOpInstance =
DeviceReduceMultiBlock<InDataType,
AccDataType,
OutDataType,
Rank,
NumReduceDim,
ReduceOperation,
InElementwiseOperation,
AccElementwiseOperation,
InMemoryDataOperationEnum::Set,
PropagateNan,
OutputIndex,
false, // HaveIndexInputIfOutputIndex
cfg1::BlockSize_,
cfg1::MThreadClusterSize_,
cfg1::KThreadClusterSize_,
cfg2::MThreadSliceSize_,
cfg2::KThreadSliceSize_,
cfg2::InSrcVectorDim_,
cfg2::InSrcVectorSize_,
cfg2::OutDstVectorSize_>;
device_op_instances.push_back(
std::make_unique<ReduceOpInstance>(ReduceOpInstance{}));
});
});
};
#define ADD_BLOCKWISE_INST_BY_TYPE( \
inT, compT, outT, ReduceOpId, NanOpt, IndicesOpt, Rank, NumReduceDim) \
template void add_device_reduce_instance_blockwise<inT, \
compT, \
outT, \
Rank, \
NumReduceDim, \
ReduceOpId, \
NanOpt, \
IndicesOpt>( \
#define ADD_BLOCKWISE_INST_BY_TYPE( \
inT, compT, outT, ReduceOpId, PropagateNan, UseIndex, Rank, NumReduceDim) \
template void add_device_reduce_instance_blockwise<inT, \
compT, \
outT, \
Rank, \
NumReduceDim, \
ReduceOpId, \
PropagateNan, \
UseIndex>( \
std::vector<deviceReduceBlockWisePtrType<compT, ReduceOpId>> & device_op_instances)
#define ADD_BLOCKWISE_INST_BY_ID( \
inT, compT, outT, ReduceOpId, NanOpt, IndicesOpt, Rank, NumReduceDim) \
ADD_BLOCKWISE_INST_BY_TYPE(inT, \
compT, \
outT, \
static_cast<ReduceTensorOp>(ReduceOpId), \
static_cast<NanPropagation>(NanOpt), \
static_cast<ReduceTensorIndices>(IndicesOpt), \
Rank, \
#define ADD_BLOCKWISE_INST_BY_ID( \
inT, compT, outT, ReduceOpId, NanOpt, IndicesOpt, Rank, NumReduceDim) \
ADD_BLOCKWISE_INST_BY_TYPE(inT, \
compT, \
outT, \
static_cast<ReduceTensorOp>(ReduceOpId), \
static_cast<bool>(NanOpt), \
static_cast<bool>(IndicesOpt), \
Rank, \
NumReduceDim)
#define ADD_BLOCKWISE_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_blockwise<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_BLOCKWISE_INST_REF_BY_ID( \
inT, compT, outT, ReduceOpId, NanOpt, IndicesOpt, Rank, NumReduceDim) \
ADD_BLOCKWISE_INST_REF_BY_TYPE(inT, \
compT, \
outT, \
static_cast<ReduceTensorOp>(ReduceOpId), \
static_cast<NanPropagation>(NanOpt), \
static_cast<ReduceTensorIndices>(IndicesOpt), \
Rank, \
#define ADD_BLOCKWISE_INST_REF_BY_ID( \
inT, compT, outT, ReduceOpId, NanOpt, IndicesOpt, Rank, NumReduceDim) \
ADD_BLOCKWISE_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_blockwise_b16_f32_b16.hpp
View file @ bb1f8082
#ifndef DEVICE_REDUCE_INSTANCE_BLOCKWISE_B16_F32_B16_HPP
#define DEVICE_REDUCE_INSTANCE_BLOCKWISE_B16_F32_B16_HPP
#include "reduction_enums.hpp"
#include "reduction_operator_mapping.hpp"
#include "data_type.hpp"
#include "device_reduce_instance_blockwise.hpp"
namespace ck {
......
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16.hpp
View file @ bb1f8082
#ifndef DEVICE_REDUCE_INSTANCE_BLOCKWISE_F16_F16_F16_HPP
#define DEVICE_REDUCE_INSTANCE_BLOCKWISE_F16_F16_F16_HPP
#include "reduction_enums.hpp"
#include "reduction_operator_mapping.hpp"
#include "data_type.hpp"
#include "device_reduce_instance_blockwise.hpp"
namespace ck {
......
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16.hpp
View file @ bb1f8082
#ifndef DEVICE_REDUCE_INSTANCE_BLOCKWISE_F16_F32_F16_HPP
#define DEVICE_REDUCE_INSTANCE_BLOCKWISE_F16_F32_F16_HPP
#include "reduction_enums.hpp"
#include "reduction_operator_mapping.hpp"
#include "data_type.hpp"
#include "device_reduce_instance_blockwise.hpp"
namespace ck {
......
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32.hpp
View file @ bb1f8082
#ifndef DEVICE_REDUCE_INSTANCE_BLOCKWISE_F32_F32_F32_HPP
#define DEVICE_REDUCE_INSTANCE_BLOCKWISE_F32_F32_F32_HPP
#include "reduction_enums.hpp"
#include "reduction_operator_mapping.hpp"
#include "device_reduce_instance_blockwise.hpp"
namespace ck {
......
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32.hpp
View file @ bb1f8082
#ifndef DEVICE_REDUCE_INSTANCE_BLOCKWISE_F32_F64_F32_HPP
#define DEVICE_REDUCE_INSTANCE_BLOCKWISE_F32_F64_F32_HPP
#include "reduction_enums.hpp"
#include "reduction_operator_mapping.hpp"
#include "device_reduce_instance_blockwise.hpp"
namespace ck {
......
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64.hpp
View file @ bb1f8082
#ifndef DEVICE_REDUCE_INSTANCE_BLOCKWISE_F64_F64_F64_HPP
#define DEVICE_REDUCE_INSTANCE_BLOCKWISE_F64_F64_F64_HPP
#include "reduction_enums.hpp"
#include "reduction_operator_mapping.hpp"
#include "device_reduce_instance_blockwise.hpp"
namespace ck {
......
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8.hpp
View file @ bb1f8082
#ifndef DEVICE_REDUCE_INSTANCE_BLOCKWISE_I8_I32_I8_HPP
#define DEVICE_REDUCE_INSTANCE_BLOCKWISE_I8_I32_I8_HPP
#include "reduction_enums.hpp"
#include "reduction_operator_mapping.hpp"
#include "device_reduce_instance_blockwise.hpp"
namespace ck {
......
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8.hpp
View file @ bb1f8082
#ifndef DEVICE_REDUCE_INSTANCE_BLOCKWISE_I8_I8_I8_HPP
#define DEVICE_REDUCE_INSTANCE_BLOCKWISE_I8_I8_I8_HPP
#include "reduction_enums.hpp"
#include "reduction_operator_mapping.hpp"
#include "device_reduce_instance_blockwise.hpp"
namespace ck {
......
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_second_call.hpp deleted 100644 → 0
View file @ 97ac5007
#ifndef DEVICE_REDUCE_INSTANCE_BLOCKWISE_SECOND_CALL_HPP
#define DEVICE_REDUCE_INSTANCE_BLOCKWISE_SECOND_CALL_HPP
#include "reduction_operator_mapping.hpp"
#include "device_reduce_instance_impl_common.hpp"
#include "device_reduce_blockwise_second_call.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_reduce_instance {
#ifdef QUICK_REDUCE_TEST
using reduce_configuration_2_instances_blockwise_second_call = std::tuple<
// clang-format off
// InSrcVectorDim | InSrcVectorSize | OutDstVectorSize | MThreadSliceSize | KThreadSliceSize
ReductionConfiguration_2<1, 2, 1, 1, 2>,
ReductionConfiguration_2<1, 1, 1, 1, 3>
// clang-format on
>;
#else
using reduce_configuration_2_instances_blockwise_second_call = std::tuple<
// clang-format off
// InSrcVectorDim | InSrcVectorSize | OutDstVectorSize | MThreadSliceSize | KThreadSliceSize
ReductionConfiguration_2<1, 4, 1, 1, 8>,
ReductionConfiguration_2<1, 4, 1, 1, 4>,
ReductionConfiguration_2<1, 2, 1, 1, 2>,
ReductionConfiguration_2<1, 1, 1, 1, 3>,
ReductionConfiguration_2<1, 1, 1, 1, 5>,
ReductionConfiguration_2<1, 1, 1, 1, 7>,
ReductionConfiguration_2<1, 1, 1, 1, 11>
// clang-format on
>;
#endif
template <typename AccDataType, ReduceTensorOp ReduceOpId>
using deviceReduceBlockWiseSecondCallPtrType = DeviceReducePtr<
typename reduce_unary_operator<AccDataType, ReduceOpId, false, true>::InElementwiseOperation,
typename reduce_unary_operator<AccDataType, ReduceOpId, false, true>::AccElementwiseOperation>;
template <typename InDataType,
typename AccDataType,
typename OutDataType,
int Rank,
int NumReduceDim,
ReduceTensorOp ReduceOpId,
NanPropagation NanOpt,
ReduceTensorIndices IndicesOpt>
void add_device_reduce_instance_blockwise_second_call(
std::vector<deviceReduceBlockWiseSecondCallPtrType<AccDataType, ReduceOpId>>&
device_op_instances)
{
using ReduceOperation = typename reduce_binary_operator<AccDataType, ReduceOpId>::opType;
using InElementwiseOperation =
typename reduce_unary_operator<AccDataType, ReduceOpId, false, true>::
InElementwiseOperation;
using AccElementwiseOperation =
typename reduce_unary_operator<AccDataType, ReduceOpId, false, true>::
AccElementwiseOperation;
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;
static_assert(std::is_same<InDataType, AccDataType>::value,
"InDataType and AccDataType should be the same to use "
"add_device_reduce_instance_blockwise_second_call!");
static_for<0, std::tuple_size<reduce_configuration_1_instances>::value, 1>{}([&](auto i) {
using cfg1 =
remove_cvref_t<decltype(std::get<i.value>(reduce_configuration_1_instances{}))>;
static_for<0,
std::tuple_size<reduce_configuration_2_instances_blockwise_second_call>::value,
1>{}([&](auto j) {
using cfg2 = remove_cvref_t<decltype(
std::get<j.value>(reduce_configuration_2_instances_blockwise_second_call{}))>;
using ReduceOpInstance = DeviceReduceBlockWiseSecondCall<InDataType,
AccDataType,
OutDataType,
Rank,
NumReduceDim,
ReduceOperation,
InElementwiseOperation,
AccElementwiseOperation,
PropagateNan,
NeedIndices,
cfg1::BlockSize_,
cfg1::MThreadClusterSize_,
cfg1::KThreadClusterSize_,
cfg2::MThreadSliceSize_,
cfg2::KThreadSliceSize_,
cfg2::InSrcVectorDim_,
cfg2::InSrcVectorSize_,
cfg2::OutDstVectorSize_>;
device_op_instances.push_back(std::make_unique<ReduceOpInstance>(ReduceOpInstance{}));
});
});
};
#define ADD_BLOCKWISE_SECOND_CALL_INST_BY_TYPE( \
inT, compT, outT, ReduceOpId, NanOpt, IndicesOpt, Rank, NumReduceDim) \
template void add_device_reduce_instance_blockwise_second_call<inT, \
compT, \
outT, \
Rank, \
NumReduceDim, \
ReduceOpId, \
NanOpt, \
IndicesOpt>( \
std::vector<deviceReduceBlockWiseSecondCallPtrType<compT, ReduceOpId>> & \
device_op_instances)
#define ADD_BLOCKWISE_SECOND_CALL_INST_BY_ID( \
inT, compT, outT, ReduceOpId, NanOpt, IndicesOpt, Rank, NumReduceDim) \
ADD_BLOCKWISE_SECOND_CALL_INST_BY_TYPE(inT, \
compT, \
outT, \
static_cast<ReduceTensorOp>(ReduceOpId), \
static_cast<NanPropagation>(NanOpt), \
static_cast<ReduceTensorIndices>(IndicesOpt), \
Rank, \
NumReduceDim)
#define ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_TYPE( \
inT, compT, outT, ReduceOpId, NanOpt, IndicesOpt, Rank, NumReduceDim) \
extern template void add_device_reduce_instance_blockwise_second_call<inT, \
compT, \
outT, \
Rank, \
NumReduceDim, \
ReduceOpId, \
NanOpt, \
IndicesOpt>( \
std::vector< \
DeviceReducePtr<typename reduce_unary_operator<compT, ReduceOpId, false, true>:: \
InElementwiseOperation, \
typename reduce_unary_operator<compT, ReduceOpId, false, true>:: \
AccElementwiseOperation>> & \
device_op_instances)
#define ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID( \
inT, compT, outT, ReduceOpId, NanOpt, IndicesOpt, Rank, NumReduceDim) \
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_TYPE(inT, \
compT, \
outT, \
static_cast<ReduceTensorOp>(ReduceOpId), \
static_cast<NanPropagation>(NanOpt), \
static_cast<ReduceTensorIndices>(IndicesOpt), \
Rank, \
NumReduceDim)
} // namespace device_reduce_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
#endif
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_second_call_f16_f16_f16.hpp deleted 100644 → 0
View file @ 97ac5007
#ifndef DEVICE_REDUCE_INSTANCE_BLOCKWISE_SECOND_CALL_F16_F16_F16_HPP
#define DEVICE_REDUCE_INSTANCE_BLOCKWISE_SECOND_CALL_F16_F16_F16_HPP
#include "reduction_enums.hpp"
#include "reduction_operator_mapping.hpp"
#include "device_reduce_instance_blockwise_second_call.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_reduce_instance {
// clang-format off
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(half_t, half_t, half_t, 2, 0, 0, 4, 3); // for MIN
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(half_t, half_t, half_t, 2, 0, 0, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(half_t, half_t, half_t, 2, 0, 0, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(half_t, half_t, half_t, 2, 0, 0, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(half_t, half_t, half_t, 3, 0, 0, 4, 3); // for MAX
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(half_t, half_t, half_t, 3, 0, 0, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(half_t, half_t, half_t, 3, 0, 0, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(half_t, half_t, half_t, 3, 0, 0, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(half_t, half_t, half_t, 4, 0, 0, 4, 3); // for AMAX
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(half_t, half_t, half_t, 4, 0, 0, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(half_t, half_t, half_t, 4, 0, 0, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(half_t, half_t, half_t, 4, 0, 0, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(half_t, half_t, half_t, 2, 0, 1, 4, 3); // for MIN
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(half_t, half_t, half_t, 2, 0, 1, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(half_t, half_t, half_t, 2, 0, 1, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(half_t, half_t, half_t, 2, 0, 1, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(half_t, half_t, half_t, 3, 0, 1, 4, 3); // for MAX
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(half_t, half_t, half_t, 3, 0, 1, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(half_t, half_t, half_t, 3, 0, 1, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(half_t, half_t, half_t, 3, 0, 1, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(half_t, half_t, half_t, 4, 0, 1, 4, 3); // for AMAX
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(half_t, half_t, half_t, 4, 0, 1, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(half_t, half_t, half_t, 4, 0, 1, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(half_t, half_t, half_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_blockwise_second_call_f32_f32_b16.hpp deleted 100644 → 0
View file @ 97ac5007
#ifndef DEVICE_REDUCE_INSTANCE_BLOCKWISE_SECOND_CALL_F32_F32_B16_HPP
#define DEVICE_REDUCE_INSTANCE_BLOCKWISE_SECOND_CALL_F32_F32_B16_HPP
#include "reduction_enums.hpp"
#include "reduction_operator_mapping.hpp"
#include "device_reduce_instance_blockwise_second_call.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_reduce_instance {
// clang-format off
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 0, 0, 0, 4, 3); // for ADD
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 0, 0, 0, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 0, 0, 0, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 0, 0, 0, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 5, 0, 0, 4, 3); // for AVG
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 5, 0, 0, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 5, 0, 0, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 5, 0, 0, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 7, 0, 0, 4, 3); // for NORM2
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 7, 0, 0, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 7, 0, 0, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 7, 0, 0, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 2, 0, 0, 4, 3); // for MIN
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 2, 0, 0, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 2, 0, 0, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 2, 0, 0, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 3, 0, 0, 4, 3); // for MAX
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 3, 0, 0, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 3, 0, 0, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 3, 0, 0, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 4, 0, 0, 4, 3); // for AMAX
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 4, 0, 0, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 4, 0, 0, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 4, 0, 0, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 2, 0, 1, 4, 3); // for MIN
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 2, 0, 1, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 2, 0, 1, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 2, 0, 1, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 3, 0, 1, 4, 3); // for MAX
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 3, 0, 1, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 3, 0, 1, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 3, 0, 1, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 4, 0, 1, 4, 3); // for AMAX
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 4, 0, 1, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_t, 4, 0, 1, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, bhalf_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_blockwise_second_call_f32_f32_f16.hpp deleted 100644 → 0
View file @ 97ac5007
#ifndef DEVICE_REDUCE_INSTANCE_BLOCKWISE_SECOND_CALL_F32_F32_F16_HPP
#define DEVICE_REDUCE_INSTANCE_BLOCKWISE_SECOND_CALL_F32_F32_F16_HPP
#include "reduction_enums.hpp"
#include "reduction_operator_mapping.hpp"
#include "device_reduce_instance_blockwise_second_call.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_reduce_instance {
// clang-format off
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, half_t, 0, 0, 0, 4, 3); // for ADD
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, half_t, 0, 0, 0, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, half_t, 0, 0, 0, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, half_t, 0, 0, 0, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, half_t, 5, 0, 0, 4, 3); // for AVG
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, half_t, 5, 0, 0, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, half_t, 5, 0, 0, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, half_t, 5, 0, 0, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, half_t, 7, 0, 0, 4, 3); // for NORM2
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, half_t, 7, 0, 0, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, half_t, 7, 0, 0, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, 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_blockwise_second_call_f32_f32_f32.hpp deleted 100644 → 0
View file @ 97ac5007
#ifndef DEVICE_REDUCE_INSTANCE_BLOCKWISE_SECOND_CALL_F32_F32_F32_HPP
#define DEVICE_REDUCE_INSTANCE_BLOCKWISE_SECOND_CALL_F32_F32_F32_HPP
#include "reduction_enums.hpp"
#include "reduction_operator_mapping.hpp"
#include "device_reduce_instance_blockwise_second_call.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_reduce_instance {
// clang-format off
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 0, 0, 0, 4, 3); // for ADD
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 0, 0, 0, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 0, 0, 0, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 0, 0, 0, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 5, 0, 0, 4, 3); // for AVG
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 5, 0, 0, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 5, 0, 0, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 5, 0, 0, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 7, 0, 0, 4, 3); // for NORM2
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 7, 0, 0, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 7, 0, 0, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 7, 0, 0, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 2, 0, 0, 4, 3); // for MIN
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 2, 0, 0, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 2, 0, 0, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 2, 0, 0, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 3, 0, 0, 4, 3); // for MAX
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 3, 0, 0, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 3, 0, 0, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 3, 0, 0, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 4, 0, 0, 4, 3); // for AMAX
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 4, 0, 0, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 4, 0, 0, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 4, 0, 0, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 2, 0, 1, 4, 3); // for MIN
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 2, 0, 1, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 2, 0, 1, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 2, 0, 1, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 3, 0, 1, 4, 3); // for MAX
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 3, 0, 1, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 3, 0, 1, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 3, 0, 1, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 4, 0, 1, 4, 3); // for AMAX
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 4, 0, 1, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 4, 0, 1, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(float, float, float, 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_blockwise_second_call_f64_f64_f32.hpp deleted 100644 → 0
View file @ 97ac5007
#ifndef DEVICE_REDUCE_INSTANCE_BLOCKWISE_SECOND_CALL_F64_F64_F32_HPP
#define DEVICE_REDUCE_INSTANCE_BLOCKWISE_SECOND_CALL_F64_F64_F32_HPP
#include "reduction_enums.hpp"
#include "reduction_operator_mapping.hpp"
#include "device_reduce_instance_blockwise_second_call.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_reduce_instance {
// clang-format off
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, float, 0, 0, 0, 4, 3); // for ADD
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, float, 0, 0, 0, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, float, 0, 0, 0, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, float, 0, 0, 0, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, float, 5, 0, 0, 4, 3); // for AVG
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, float, 5, 0, 0, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, float, 5, 0, 0, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, float, 5, 0, 0, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, float, 7, 0, 0, 4, 3); // for NORM2
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, float, 7, 0, 0, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, float, 7, 0, 0, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, 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_blockwise_second_call_f64_f64_f64.hpp deleted 100644 → 0
View file @ 97ac5007
#ifndef DEVICE_REDUCE_INSTANCE_BLOCKWISE_SECOND_CALL_F64_F64_F64_HPP
#define DEVICE_REDUCE_INSTANCE_BLOCKWISE_SECOND_CALL_F64_F64_F64_HPP
#include "reduction_enums.hpp"
#include "reduction_operator_mapping.hpp"
#include "device_reduce_instance_blockwise_second_call.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_reduce_instance {
// clang-format off
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 0, 0, 0, 4, 3); // for ADD
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 0, 0, 0, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 0, 0, 0, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 0, 0, 0, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 5, 0, 0, 4, 3); // for AVG
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 5, 0, 0, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 5, 0, 0, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 5, 0, 0, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 7, 0, 0, 4, 3); // for NORM2
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 7, 0, 0, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 7, 0, 0, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 7, 0, 0, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 2, 0, 0, 4, 3); // for MIN
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 2, 0, 0, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 2, 0, 0, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 2, 0, 0, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 3, 0, 0, 4, 3); // for MAX
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 3, 0, 0, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 3, 0, 0, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 3, 0, 0, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 4, 0, 0, 4, 3); // for AMAX
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 4, 0, 0, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 4, 0, 0, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 4, 0, 0, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 2, 0, 1, 4, 3); // for MIN
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 2, 0, 1, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 2, 0, 1, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 2, 0, 1, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 3, 0, 1, 4, 3); // for MAX
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 3, 0, 1, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 3, 0, 1, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 3, 0, 1, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 4, 0, 1, 4, 3); // for AMAX
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 4, 0, 1, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 4, 0, 1, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(double, double, double, 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_blockwise_second_call_i32_i32_i8.hpp deleted 100644 → 0
View file @ 97ac5007
#ifndef DEVICE_REDUCE_INSTANCE_BLOCKWISE_SECOND_CALL_I32_I32_I8_HPP
#define DEVICE_REDUCE_INSTANCE_BLOCKWISE_SECOND_CALL_I32_I32_I8_HPP
#include "reduction_enums.hpp"
#include "reduction_operator_mapping.hpp"
#include "device_reduce_instance_blockwise_second_call.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_reduce_instance {
// clang-format off
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(int32_t, int32_t, int8_t, 0, 0, 0, 4, 3); // for ADD
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(int32_t, int32_t, int8_t, 0, 0, 0, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(int32_t, int32_t, int8_t, 0, 0, 0, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(int32_t, int32_t, int8_t, 0, 0, 0, 2, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(int32_t, int32_t, int8_t, 5, 0, 0, 4, 3); // for AVG
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(int32_t, int32_t, int8_t, 5, 0, 0, 4, 4);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(int32_t, int32_t, int8_t, 5, 0, 0, 4, 1);
ADD_BLOCKWISE_SECOND_CALL_INST_REF_BY_ID(int32_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
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