Skip to content

GitLab

Unverified Commit 80e05267 authored by Qianfeng's avatar Qianfeng Committed by GitHub
Browse files

Reduction external API and client examples (#493) 



* Change to the DeviceReduce base class template to include all problem description information

* Add external api for reduction

* Add client example to test the reduction external api

* Spelling correction

* Re-implement the host_reduction to follow the DeviceReduce base API format

* Change the reduce profiler to call the external API for collecting device instances

* Rename reduce client example directory from 08_reduce to 12_reduce

* Remove (void) before the functional call

* Tiny update in reduce client example

* Tiny update in profile_reduce_impl.hpp

* Rename the reduce client example directory
Co-authored-by: default avatarPo Yen Chen <PoYen.Chen@amd.com>
parent 7829d729
Hide whitespace changes
Inline Side-by-side
Showing with 99 additions and 102 deletions
library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.cpp
View file @ 80e05267
...@@ -11,10 +11,10 @@ namespace instance { ...@@ -11,10 +11,10 @@ namespace instance {
// clang-format off // clang-format off
// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex // InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex
template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&); template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&); template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&); template void add_device_reduce_instance_threadwise<I8, I32, I8, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
template void add_device_reduce_instance_threadwise<I8, I32, I8, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&); template void add_device_reduce_instance_threadwise<I8, I32, I8, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<I8, I32, I8, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>>&);
// clang-format on // clang-format on
} // namespace instance } // namespace instance
......
library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.cpp
View file @ 80e05267
...@@ -11,14 +11,14 @@ namespace instance { ...@@ -11,14 +11,14 @@ namespace instance {
// clang-format off // clang-format off
// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex // InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex
template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&); template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&); template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&); template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&); template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>>&);
template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&); template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&); template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&); template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&); template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>>&);
// clang-format on // clang-format on
} // namespace instance } // namespace instance
......
library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.cpp
View file @ 80e05267
...@@ -11,14 +11,14 @@ namespace instance { ...@@ -11,14 +11,14 @@ namespace instance {
// clang-format off // clang-format off
// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex // InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex
template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>>&);
template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>>&);
template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&); template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>>&);
template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>>&);
template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>>&);
template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&); template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>>&);
// clang-format on // clang-format on
} // namespace instance } // namespace instance
......
library/src/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.cpp
View file @ 80e05267
...@@ -11,14 +11,14 @@ namespace instance { ...@@ -11,14 +11,14 @@ namespace instance {
// clang-format off // clang-format off
// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex // InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex
template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>>&);
template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>>&);
template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&); template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<I8, I8, I8, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>>&);
template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&); template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>>&);
template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&); template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>>&);
template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&); template void add_device_reduce_instance_threadwise<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&); template void add_device_reduce_instance_threadwise<I8, I8, I8, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<I8, I8, I8, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>>&);
// clang-format on // clang-format on
} // namespace instance } // namespace instance
......
profiler/include/profiler/profile_reduce_impl.hpp
View file @ 80e05267
...@@ -6,11 +6,11 @@ ...@@ -6,11 +6,11 @@
#include "ck/utility/reduction_enums.hpp" #include "ck/utility/reduction_enums.hpp"
#include "ck/tensor_operation/gpu/device/device_reduce.hpp" #include "ck/tensor_operation/gpu/device/device_reduce.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance.hpp" #include "ck/library/tensor_operation_instance/gpu/reduce/reduce.hpp"
#include "ck/library/utility/algorithm.hpp" #include "ck/library/utility/algorithm.hpp"
#include "ck/library/utility/check_err.hpp" #include "ck/library/utility/check_err.hpp"
#include "ck/library/utility/device_memory.hpp" #include "ck/library/utility/device_memory.hpp"
#include "ck/library/utility/host_reduction.hpp" #include "ck/library/reference_tensor_operation/cpu/reference_reduce.hpp"
#include "ck/library/utility/host_common_util.hpp" #include "ck/library/utility/host_common_util.hpp"
#include "ck/library/utility/host_tensor_generator.hpp" #include "ck/library/utility/host_tensor_generator.hpp"
...@@ -158,11 +158,6 @@ bool profile_reduce_impl_impl(bool do_verification, ...@@ -158,11 +158,6 @@ bool profile_reduce_impl_impl(bool do_verification,
constexpr bool OutputIndex = (op_support_indices && UseIndex); constexpr bool OutputIndex = (op_support_indices && UseIndex);
constexpr bool out_support_atomic_add = std::is_same<OutDataType, float>::value;
constexpr bool op_support_atomic_add =
!op_support_indices && ReduceOpId != ReduceTensorOp::NORM2;
constexpr bool use_atomic_add = (out_support_atomic_add && op_support_atomic_add);
// 1) If InDataType is half_t, must use half_t as AccDataType for indexable reduction operations // 1) If InDataType is half_t, must use half_t as AccDataType for indexable reduction operations
// 2) If InDataType is half_t, must use float as AccDataType for non-indexable reduction // 2) If InDataType is half_t, must use float as AccDataType for non-indexable reduction
// operations // operations
...@@ -200,7 +195,8 @@ bool profile_reduce_impl_impl(bool do_verification, ...@@ -200,7 +195,8 @@ bool profile_reduce_impl_impl(bool do_verification,
constexpr bool invalid_reduce = (invalid_reduce_1 || invalid_reduce_2 || invalid_reduce_3 || constexpr bool invalid_reduce = (invalid_reduce_1 || invalid_reduce_2 || invalid_reduce_3 ||
invalid_reduce_4 || invalid_reduce_5 || invalid_reduce_6); invalid_reduce_4 || invalid_reduce_5 || invalid_reduce_6);
bool pass = true; int num_kernel = 0;
bool pass = true;
if constexpr(!invalid_reduce) if constexpr(!invalid_reduce)
{ {
...@@ -286,75 +282,25 @@ bool profile_reduce_impl_impl(bool do_verification, ...@@ -286,75 +282,25 @@ bool profile_reduce_impl_impl(bool do_verification,
reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator( reduce_unary_operator<ReduceOpId, true, true>::GetElementwiseOperator(
static_cast<int32_t>(reduce_total_length)); static_cast<int32_t>(reduce_total_length));
using DeviceReduceInstPtr = using ReduceOp = ck::tensor_operation::device::DeviceReduce<InDataType,
DeviceReducePtr<Rank, NumReduceDim, InElementwiseOperation, AccElementwiseOperation>; AccDataType,
OutDataType,
std::vector<DeviceReduceInstPtr> reduce_ptrs; Rank,
NumReduceDim,
add_device_reduce_instance_threadwise<InDataType, ReduceOperation,
AccDataType, InElementwiseOperation,
OutDataType, AccElementwiseOperation,
Rank, PropagateNan,
NumReduceDim, OutputIndex>;
ReduceOperation, const auto reduce_ptrs =
InElementwiseOperation, ck::tensor_operation::device::instance::DeviceOperationInstanceFactory<
AccElementwiseOperation, ReduceOp>::GetInstances();
PropagateNan,
UseIndex>(reduce_ptrs);
add_device_reduce_instance_blockwise<InDataType,
AccDataType,
OutDataType,
Rank,
NumReduceDim,
ReduceOperation,
InElementwiseOperation,
AccElementwiseOperation,
PropagateNan,
UseIndex>(reduce_ptrs);
if constexpr(use_atomic_add)
{
add_device_reduce_instance_multiblock_atomic_add<InDataType,
AccDataType,
OutDataType,
Rank,
NumReduceDim,
ReduceOperation,
InElementwiseOperation,
AccElementwiseOperation,
PropagateNan,
UseIndex>(reduce_ptrs);
}
if(reduce_ptrs.empty()) if(reduce_ptrs.empty())
{ {
throw std::runtime_error("Wrong! No device REDUCE instance found"); throw std::runtime_error("Wrong! No device REDUCE instance found");
}; };
if(do_verification)
{
ReductionHost<InDataType,
AccDataType,
OutDataType,
ReduceOperation,
InElementwiseOperation,
AccElementwiseOperation,
Rank,
NumReduceDim,
PropagateNan,
OutputIndex>
hostReduce(in.mDesc, out_ref.mDesc, invariantDims, reduceDims);
hostReduce.Run(alpha,
in.mData.data(),
beta,
out_ref.mData.data(),
out_indices_ref.mData.data(),
in_elementwise_op,
acc_elementwise_op);
};
std::array<index_t, Rank> arrInLengths; std::array<index_t, Rank> arrInLengths;
std::array<index_t, Rank> arrInStrides; std::array<index_t, Rank> arrInStrides;
std::array<index_t, NumOutDim> arrOutLengths; std::array<index_t, NumOutDim> arrOutLengths;
...@@ -365,6 +311,49 @@ bool profile_reduce_impl_impl(bool do_verification, ...@@ -365,6 +311,49 @@ bool profile_reduce_impl_impl(bool do_verification,
ck::ranges::copy(outLengths, arrOutLengths.begin()); ck::ranges::copy(outLengths, arrOutLengths.begin());
ck::ranges::copy(outStrides, arrOutStrides.begin()); ck::ranges::copy(outStrides, arrOutStrides.begin());
if(do_verification)
{
using ReferenceReduceInstance =
ck::tensor_operation::host::ReferenceReduce<InDataType,
AccDataType,
OutDataType,
Rank,
NumReduceDim,
ReduceOperation,
InElementwiseOperation,
AccElementwiseOperation,
PropagateNan,
OutputIndex>;
auto reduce_ref = ReferenceReduceInstance{};
auto argument_ptr_ref = reduce_ref.MakeArgumentPointer(arrInLengths,
arrInStrides,
arrOutLengths,
arrOutStrides,
reduceDims,
alpha,
beta,
in.mData.data(),
nullptr,
out_ref.mData.data(),
out_indices_ref.mData.data(),
in_elementwise_op,
acc_elementwise_op);
if(!reduce_ref.IsSupportedArgument(argument_ptr_ref.get()))
{
std::cout
<< "The runtime parameters not supported by the reduce reference, exiting!"
<< std::endl;
return (false);
};
auto invoker_ptr_ref = reduce_ref.MakeInvokerPointer();
(void)invoker_ptr_ref->Run(argument_ptr_ref.get());
};
for(auto& reduce_ptr : reduce_ptrs) for(auto& reduce_ptr : reduce_ptrs)
{ {
auto argument_ptr = reduce_ptr->MakeArgumentPointer(arrInLengths, auto argument_ptr = reduce_ptr->MakeArgumentPointer(arrInLengths,
...@@ -383,6 +372,8 @@ bool profile_reduce_impl_impl(bool do_verification, ...@@ -383,6 +372,8 @@ bool profile_reduce_impl_impl(bool do_verification,
if(!reduce_ptr->IsSupportedArgument(argument_ptr.get())) if(!reduce_ptr->IsSupportedArgument(argument_ptr.get()))
continue; continue;
else
num_kernel++;
std::string reduce_name = reduce_ptr->GetTypeString(); std::string reduce_name = reduce_ptr->GetTypeString();
...@@ -446,14 +437,20 @@ bool profile_reduce_impl_impl(bool do_verification, ...@@ -446,14 +437,20 @@ bool profile_reduce_impl_impl(bool do_verification,
}; };
}; };
if(time_kernel) if(time_kernel && num_kernel > 0)
std::cout << "Best Perf: " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s" std::cout << "Best Perf: " << best_avg_time << " ms, " << best_gb_per_sec << " GB/s"
<< std::endl; << std::endl;
} }
else else
{ {
std::cout << "The requested reduction operation is not supported, please check !!!" throw std::runtime_error(
<< std::endl; "The requested reduction operation is not supported, please check!");
};
if(num_kernel == 0)
{
std::cout << "Error: No kernel is applicable" << std::endl;
return false;
}; };
return pass; return pass;
......
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