mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-04-19 22:39:03 +00:00
Overhaul to Reducton and its dependants (#237)
* Tiny fix in dynamic_buffer.hpp to support vectorized AtomicAdd for double type * Update to host layer and host reduction * Merge and remove reduction kernels * Merge and remove reduction device interfaces and update pooling device interface * Merge and remove useless reduction device instances * Update to reduction profiler and reduction ctests * Update to reduction and pooling examples and add one reduction example * Change to reduction examples to let them testable by ctest * Add explicit pass checking for reduction and pooling examples * Explicit assignment of tensor shapes in example reduce_blockwise_two_call * Use atomic_add to repace atomicAdd and add atomic_add for double type * Add reduce ctest support for double data type * Replace to_int_vector() by using c++ std::vector::assign() * Keep DeviceReduceThreadWise separated from DeviceReduceBlockWise * Merge DeviceReduceBlockWise and DeviceReduceMultiBlockAtomicAdd into DeviceReduceMultiBlock * Add GetAtomicOperationZeroValue() support for AtomicMax * Tiny change to reduce example README.md * Fix some tiny issues due to branch merging * Revoke previous change in dynamic_buffer.hpp and add atomic_add for double2_t * Add reduce multiblock_atomic_add instances for fp64 to verify vectorized atomic_add on fp64 * Renaming * Clean the header includings in device_reduce instances header files
This commit is contained in:
Qianfeng
@@ -1,384 +1,10 @@
|
||||
#include "getopt.h"
|
||||
|
||||
#include "check_err.hpp"
|
||||
#include "device_reduce_instance.hpp"
|
||||
#include "reduction_enums.hpp"
|
||||
#include "host_tensor.hpp"
|
||||
#include "host_tensor_generator.hpp"
|
||||
#include "host_reduction.hpp"
|
||||
#include "reduce_util.hpp"
|
||||
#include "host_common_util.hpp"
|
||||
#include "profile_reduce_impl.hpp"
|
||||
|
||||
using namespace ck;
|
||||
|
||||
namespace {
|
||||
|
||||
template <index_t Rank, index_t NumReduceDim>
|
||||
static inline std::vector<int> get_invariant_dims(const std::vector<int>& reduceDims)
|
||||
{
|
||||
assert(NumReduceDim == reduceDims.size());
|
||||
|
||||
int reduceFlag = 0;
|
||||
|
||||
// flag the bits for the reduceDims
|
||||
for(int i = 0; i < NumReduceDim; i++)
|
||||
{
|
||||
reduceFlag |= 1 << reduceDims[i];
|
||||
};
|
||||
|
||||
std::vector<int> invariantDims;
|
||||
|
||||
// collect invariant dimensions
|
||||
for(int i = 0; i < Rank; i++)
|
||||
if((reduceFlag & (1 << i)) == 0)
|
||||
{
|
||||
invariantDims.push_back(i);
|
||||
};
|
||||
|
||||
return invariantDims;
|
||||
};
|
||||
|
||||
constexpr int Rank = 4;
|
||||
|
||||
constexpr ReduceTensorOp ReduceOpId = ReduceTensorOp::AVG;
|
||||
constexpr NanPropagation NanOpt = NanPropagation::PROPAGATE_NAN;
|
||||
constexpr bool PropagateNan = false;
|
||||
constexpr ReduceTensorIndices IndicesOpt = ReduceTensorIndices::NO_INDICES;
|
||||
constexpr bool NeedIndices = false;
|
||||
|
||||
template <typename InDataType,
|
||||
typename AccDataType,
|
||||
typename OutDataType,
|
||||
int Rank,
|
||||
int NumReduceDim>
|
||||
bool test_reduce_no_index_impl(int init_method,
|
||||
const std::vector<size_t>& inLengths,
|
||||
const std::vector<int>& reduceDims,
|
||||
float alpha,
|
||||
float beta)
|
||||
{
|
||||
using namespace ck::tensor_operation::device;
|
||||
using namespace ck::tensor_operation::device::device_reduce_instance;
|
||||
using namespace ck::host_reduce;
|
||||
|
||||
constexpr bool out_support_atomic_add = std::is_same<OutDataType, float>::value;
|
||||
constexpr bool op_support_atomic_add = true;
|
||||
constexpr bool use_atomic_add = (out_support_atomic_add && op_support_atomic_add);
|
||||
|
||||
Tensor<InDataType> in(inLengths);
|
||||
|
||||
std::vector<size_t> outLengths;
|
||||
|
||||
const auto invariantDims = get_invariant_dims<Rank, NumReduceDim>(reduceDims);
|
||||
|
||||
if(reduceDims.size() == Rank)
|
||||
outLengths.push_back(1);
|
||||
else
|
||||
for(auto dim : invariantDims)
|
||||
outLengths.push_back(inLengths[dim]);
|
||||
|
||||
Tensor<OutDataType> out_ref(outLengths);
|
||||
Tensor<OutDataType> out(outLengths);
|
||||
|
||||
// only used when the OutDataType is bhalf_t
|
||||
Tensor<float> out_ref_fp32(outLengths);
|
||||
Tensor<float> out_fp32(outLengths);
|
||||
|
||||
auto inStrides = in.mDesc.GetStrides();
|
||||
auto outStrides = out.mDesc.GetStrides();
|
||||
|
||||
size_t invariant_total_length = out.mDesc.GetElementSize();
|
||||
size_t reduce_total_length = in.mDesc.GetElementSize() / invariant_total_length;
|
||||
|
||||
std::size_t num_thread = 1;
|
||||
|
||||
switch(init_method)
|
||||
{
|
||||
case 0: break;
|
||||
case 1:
|
||||
in.GenerateTensorValue(GeneratorTensor_1<InDataType>{1}, num_thread);
|
||||
if(beta != 0.0f)
|
||||
out_ref.GenerateTensorValue(GeneratorTensor_1<InDataType>{1}, num_thread);
|
||||
break;
|
||||
case 2:
|
||||
in.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}, num_thread);
|
||||
if(beta != 0.0f)
|
||||
out_ref.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5}, num_thread);
|
||||
break;
|
||||
default:
|
||||
in.GenerateTensorValue(GeneratorTensor_3<InDataType>{-5.0, 5.0}, num_thread);
|
||||
if(beta != 0.0f)
|
||||
out_ref.GenerateTensorValue(GeneratorTensor_3<InDataType>{-5.0, 5.0}, num_thread);
|
||||
}
|
||||
|
||||
if(beta != 0.0f)
|
||||
for(size_t i = 0; i < out_ref.mDesc.GetElementSpace(); i++)
|
||||
out.mData[i] = out_ref.mData[i];
|
||||
|
||||
// these buffers are usually provided by the user application
|
||||
DeviceMem in_dev(sizeof(InDataType) * in.mDesc.GetElementSpace());
|
||||
DeviceMem out_dev(sizeof(OutDataType) * out.mDesc.GetElementSpace());
|
||||
|
||||
in_dev.ToDevice(in.mData.data());
|
||||
|
||||
if(beta != 0.0f)
|
||||
out_dev.ToDevice(out.mData.data());
|
||||
|
||||
using InElementwiseOperation_0 =
|
||||
typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::InElementwiseOperation;
|
||||
using AccElementwiseOperation_0 =
|
||||
typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::
|
||||
AccElementwiseOperation;
|
||||
using InElementwiseOperation_1 =
|
||||
typename reduce_unary_operator<AccDataType, ReduceOpId, true, false>::
|
||||
InElementwiseOperation;
|
||||
using AccElementwiseOperation_1 =
|
||||
typename reduce_unary_operator<AccDataType, ReduceOpId, true, false>::
|
||||
AccElementwiseOperation;
|
||||
using InElementwiseOperation_2 =
|
||||
typename reduce_unary_operator<AccDataType, ReduceOpId, false, true>::
|
||||
InElementwiseOperation;
|
||||
using AccElementwiseOperation_2 =
|
||||
typename reduce_unary_operator<AccDataType, ReduceOpId, false, true>::
|
||||
AccElementwiseOperation;
|
||||
|
||||
using DeviceReduceInstPtr0 =
|
||||
DeviceReducePtr<InElementwiseOperation_0, AccElementwiseOperation_0>;
|
||||
using DeviceReduceInstPtr1 =
|
||||
DeviceReducePtr<InElementwiseOperation_1, AccElementwiseOperation_1>;
|
||||
using DeviceReduceInstPtr2 =
|
||||
DeviceReducePtr<InElementwiseOperation_2, AccElementwiseOperation_2>;
|
||||
|
||||
std::vector<DeviceReduceInstPtr0> reduce0_ptrs;
|
||||
std::vector<DeviceReduceInstPtr1> reduce1_ptrs;
|
||||
std::vector<DeviceReduceInstPtr2> reduce2_ptrs;
|
||||
|
||||
add_device_reduce_instance_threadwise<InDataType,
|
||||
AccDataType,
|
||||
OutDataType,
|
||||
Rank,
|
||||
NumReduceDim,
|
||||
ReduceOpId,
|
||||
NanOpt,
|
||||
IndicesOpt>(reduce0_ptrs);
|
||||
|
||||
add_device_reduce_instance_blockwise<InDataType,
|
||||
AccDataType,
|
||||
OutDataType,
|
||||
Rank,
|
||||
NumReduceDim,
|
||||
ReduceOpId,
|
||||
NanOpt,
|
||||
IndicesOpt>(reduce0_ptrs);
|
||||
|
||||
if constexpr(use_atomic_add)
|
||||
{
|
||||
add_device_reduce_instance_multiblock_atomic_add<InDataType,
|
||||
AccDataType,
|
||||
OutDataType,
|
||||
Rank,
|
||||
NumReduceDim,
|
||||
ReduceOpId,
|
||||
NanOpt,
|
||||
IndicesOpt>(reduce0_ptrs);
|
||||
}
|
||||
else
|
||||
{
|
||||
add_device_reduce_instance_multiblock_partial_reduce<InDataType,
|
||||
AccDataType,
|
||||
OutDataType,
|
||||
Rank,
|
||||
NumReduceDim,
|
||||
ReduceOpId,
|
||||
NanOpt,
|
||||
IndicesOpt>(reduce1_ptrs);
|
||||
};
|
||||
|
||||
// used for secondary reduction
|
||||
if constexpr(!use_atomic_add)
|
||||
{
|
||||
add_device_reduce_instance_blockwise_second_call<AccDataType,
|
||||
AccDataType,
|
||||
OutDataType,
|
||||
Rank,
|
||||
NumReduceDim,
|
||||
ReduceOpId,
|
||||
NanOpt,
|
||||
IndicesOpt>(reduce2_ptrs);
|
||||
};
|
||||
|
||||
if(reduce0_ptrs.empty() && reduce1_ptrs.empty())
|
||||
{
|
||||
throw std::runtime_error("Wrong! No device REDUCE instance found");
|
||||
};
|
||||
|
||||
bool result = true;
|
||||
|
||||
ReductionHost<InDataType,
|
||||
AccDataType,
|
||||
OutDataType,
|
||||
ReduceOpId,
|
||||
Rank,
|
||||
NumReduceDim,
|
||||
PropagateNan,
|
||||
NeedIndices>
|
||||
hostReduce(in.mDesc, out_ref.mDesc, invariantDims, reduceDims);
|
||||
|
||||
hostReduce.Run(alpha, in.mData.data(), beta, out_ref.mData.data(), nullptr);
|
||||
|
||||
const auto i_inLengths = to_int_vector(inLengths);
|
||||
const auto i_inStrides = to_int_vector(inStrides);
|
||||
const auto i_outLengths = to_int_vector(outLengths);
|
||||
const auto i_outStrides = to_int_vector(outStrides);
|
||||
|
||||
for(auto& reduce_ptr : reduce0_ptrs)
|
||||
{
|
||||
auto wsSizeInBytes = reduce_ptr->GetWorkspaceSizeInBytes(i_inLengths, reduceDims);
|
||||
|
||||
DeviceMem ws_dev(wsSizeInBytes);
|
||||
|
||||
InElementwiseOperation_0 in_elementwise_op_0(static_cast<int32_t>(reduce_total_length));
|
||||
AccElementwiseOperation_0 acc_elementwise_op_0(static_cast<int32_t>(reduce_total_length));
|
||||
|
||||
auto argument_ptr = reduce_ptr->MakeArgumentPointer(i_inLengths,
|
||||
i_inStrides,
|
||||
i_outLengths,
|
||||
i_outStrides,
|
||||
reduceDims,
|
||||
alpha,
|
||||
beta,
|
||||
in_dev.GetDeviceBuffer(),
|
||||
out_dev.GetDeviceBuffer(),
|
||||
nullptr,
|
||||
ws_dev.GetDeviceBuffer(),
|
||||
in_elementwise_op_0,
|
||||
acc_elementwise_op_0);
|
||||
|
||||
if(!reduce_ptr->IsSupportedArgument(argument_ptr.get()))
|
||||
continue;
|
||||
|
||||
auto invoker_ptr = reduce_ptr->MakeInvokerPointer();
|
||||
|
||||
(void)invoker_ptr->Run(argument_ptr.get());
|
||||
|
||||
out_dev.FromDevice(out.mData.data());
|
||||
|
||||
bool single_result = true;
|
||||
|
||||
if constexpr(std::is_same<OutDataType, ck::half_t>::value ||
|
||||
std::is_same<OutDataType, ck::bhalf_t>::value)
|
||||
{
|
||||
reduce_util::to_f32_vector(out, out_fp32);
|
||||
reduce_util::to_f32_vector(out_ref, out_ref_fp32);
|
||||
single_result = ck::utils::check_err(
|
||||
out_fp32.mData, out_ref_fp32.mData, "Error: incorrect data result!");
|
||||
}
|
||||
else
|
||||
{
|
||||
single_result =
|
||||
ck::utils::check_err(out.mData, out_ref.mData, "Error: incorrect data result!");
|
||||
};
|
||||
|
||||
if(!single_result)
|
||||
{
|
||||
std::cout << "Fail Info: " << reduce_ptr->GetTypeString() << std::endl;
|
||||
result = false;
|
||||
}
|
||||
};
|
||||
|
||||
for(auto& reduce_ptr : reduce1_ptrs)
|
||||
{
|
||||
auto wsSizeInBytes = reduce_ptr->GetWorkspaceSizeInBytes(i_inLengths, reduceDims);
|
||||
|
||||
DeviceMem ws_dev(wsSizeInBytes);
|
||||
|
||||
InElementwiseOperation_1 in_elementwise_op_1(static_cast<int32_t>(reduce_total_length));
|
||||
AccElementwiseOperation_1 acc_elementwise_op_1(static_cast<int32_t>(reduce_total_length));
|
||||
|
||||
auto argument_ptr = reduce_ptr->MakeArgumentPointer(i_inLengths,
|
||||
i_inStrides,
|
||||
i_outLengths,
|
||||
i_outStrides,
|
||||
reduceDims,
|
||||
alpha,
|
||||
beta,
|
||||
in_dev.GetDeviceBuffer(),
|
||||
out_dev.GetDeviceBuffer(),
|
||||
nullptr,
|
||||
ws_dev.GetDeviceBuffer(),
|
||||
in_elementwise_op_1,
|
||||
acc_elementwise_op_1);
|
||||
|
||||
if(!reduce_ptr->IsSupportedArgument(argument_ptr.get()))
|
||||
continue;
|
||||
|
||||
auto invoker_ptr = reduce_ptr->MakeInvokerPointer();
|
||||
|
||||
(void)invoker_ptr->Run(argument_ptr.get());
|
||||
|
||||
std::vector<int> inLengths2 = reduce_ptr->GetWorkspace2dLengths(argument_ptr.get());
|
||||
std::vector<int> inStrides2{inLengths2[1], 1};
|
||||
|
||||
for(auto& reduce2_ptr : reduce2_ptrs)
|
||||
{
|
||||
InElementwiseOperation_2 in_elementwise_op_2(static_cast<int32_t>(reduce_total_length));
|
||||
AccElementwiseOperation_2 acc_elementwise_op_2(
|
||||
static_cast<int32_t>(reduce_total_length));
|
||||
|
||||
auto argument2_ptr = reduce2_ptr->MakeArgumentPointer(inLengths2,
|
||||
inStrides2,
|
||||
i_outLengths,
|
||||
i_outStrides,
|
||||
reduceDims,
|
||||
alpha,
|
||||
beta,
|
||||
ws_dev.GetDeviceBuffer(),
|
||||
out_dev.GetDeviceBuffer(),
|
||||
nullptr,
|
||||
ws_dev.GetDeviceBuffer(),
|
||||
in_elementwise_op_2,
|
||||
acc_elementwise_op_2);
|
||||
|
||||
if(!reduce2_ptr->IsSupportedArgument(argument2_ptr.get()))
|
||||
continue;
|
||||
|
||||
std::string reduce2_name = reduce2_ptr->GetTypeString();
|
||||
|
||||
auto invoker2_ptr = reduce2_ptr->MakeInvokerPointer();
|
||||
|
||||
(void)invoker2_ptr->Run(argument2_ptr.get());
|
||||
|
||||
out_dev.FromDevice(out.mData.data());
|
||||
|
||||
bool single_result = true;
|
||||
|
||||
if constexpr(std::is_same<OutDataType, ck::half_t>::value ||
|
||||
std::is_same<OutDataType, ck::bhalf_t>::value)
|
||||
{
|
||||
reduce_util::to_f32_vector(out, out_fp32);
|
||||
reduce_util::to_f32_vector(out_ref, out_ref_fp32);
|
||||
single_result = ck::utils::check_err(
|
||||
out_fp32.mData, out_ref_fp32.mData, "Error: incorrect data result!");
|
||||
}
|
||||
else
|
||||
{
|
||||
single_result =
|
||||
ck::utils::check_err(out.mData, out_ref.mData, "Error: incorrect data result!");
|
||||
};
|
||||
|
||||
if(!single_result)
|
||||
{
|
||||
std::cout << "Fail Info: " << reduce_ptr->GetTypeString() << " => "
|
||||
<< reduce2_ptr->GetTypeString() << std::endl;
|
||||
result = false;
|
||||
}
|
||||
};
|
||||
};
|
||||
|
||||
return (result);
|
||||
};
|
||||
|
||||
} // anonymous namespace
|
||||
|
||||
static struct option long_options[] = {{"inLengths", required_argument, nullptr, 'D'},
|
||||
{"reduceDimensions", required_argument, nullptr, 'R'},
|
||||
{"scales", required_argument, nullptr, 'S'},
|
||||
@@ -387,48 +13,6 @@ static struct option long_options[] = {{"inLengths", required_argument, nullptr,
|
||||
|
||||
class SimpleAppArgs
|
||||
{
|
||||
template <typename T>
|
||||
static T getSingleValueFromString(const std::string& valueStr)
|
||||
{
|
||||
std::istringstream iss(valueStr);
|
||||
|
||||
T ret;
|
||||
|
||||
iss >> ret;
|
||||
|
||||
return (ret);
|
||||
};
|
||||
|
||||
template <typename T>
|
||||
static std::vector<T> getTypeValuesFromString(const char* cstr_values)
|
||||
{
|
||||
std::string valuesStr(cstr_values);
|
||||
|
||||
std::vector<T> values;
|
||||
std::size_t pos = 0;
|
||||
std::size_t new_pos;
|
||||
|
||||
new_pos = valuesStr.find(',', pos);
|
||||
while(new_pos != std::string::npos)
|
||||
{
|
||||
const std::string sliceStr = valuesStr.substr(pos, new_pos - pos);
|
||||
|
||||
T val = getSingleValueFromString<T>(sliceStr);
|
||||
|
||||
values.push_back(val);
|
||||
|
||||
pos = new_pos + 1;
|
||||
new_pos = valuesStr.find(',', pos);
|
||||
};
|
||||
|
||||
std::string sliceStr = valuesStr.substr(pos);
|
||||
T val = getSingleValueFromString<T>(sliceStr);
|
||||
|
||||
values.push_back(val);
|
||||
|
||||
return (values);
|
||||
};
|
||||
|
||||
private:
|
||||
int option_index = 0;
|
||||
|
||||
@@ -460,6 +44,8 @@ class SimpleAppArgs
|
||||
|
||||
int processArgs(int argc, char* argv[])
|
||||
{
|
||||
using ck::host_common::getTypeValuesFromString;
|
||||
|
||||
int ch;
|
||||
|
||||
while(1)
|
||||
@@ -514,7 +100,7 @@ class SimpleAppArgs
|
||||
(reduceDims.size() != 1 && reduceDims.size() != 3 && reduceDims.size() != 4))
|
||||
return (-1);
|
||||
|
||||
if(data_type != 0 && data_type != 1 && data_type != 3 && data_type != 5)
|
||||
if(data_type != 0 && data_type != 1 && data_type != 3 && data_type != 5 && data_type != 6)
|
||||
return (-1);
|
||||
|
||||
return (0);
|
||||
@@ -525,87 +111,92 @@ bool test_reduce_no_index(int data_type,
|
||||
int init_method,
|
||||
std::vector<int> reduceDims,
|
||||
std::vector<size_t> inLengths,
|
||||
ReduceTensorOp reduceOpId,
|
||||
bool propagateNan,
|
||||
float alpha,
|
||||
float beta)
|
||||
{
|
||||
using ck::profiler::profile_reduce_impl;
|
||||
|
||||
bool result = true;
|
||||
|
||||
if(data_type == 0)
|
||||
{
|
||||
switch(reduceDims.size())
|
||||
{
|
||||
case 1:
|
||||
result = test_reduce_no_index_impl<float, float, float, Rank, 1>(
|
||||
init_method, inLengths, reduceDims, alpha, beta);
|
||||
break;
|
||||
case 3:
|
||||
result = test_reduce_no_index_impl<float, float, float, Rank, 3>(
|
||||
init_method, inLengths, reduceDims, alpha, beta);
|
||||
break;
|
||||
case 4:
|
||||
result = test_reduce_no_index_impl<float, float, float, Rank, 4>(
|
||||
init_method, inLengths, reduceDims, alpha, beta);
|
||||
break;
|
||||
};
|
||||
result = profile_reduce_impl<float, float, float>(true,
|
||||
init_method,
|
||||
false,
|
||||
false,
|
||||
inLengths,
|
||||
reduceDims,
|
||||
reduceOpId,
|
||||
propagateNan,
|
||||
false,
|
||||
alpha,
|
||||
beta);
|
||||
}
|
||||
else if(data_type == 1)
|
||||
{
|
||||
switch(reduceDims.size())
|
||||
{
|
||||
case 1:
|
||||
result = test_reduce_no_index_impl<ck::half_t, float, ck::half_t, Rank, 1>(
|
||||
init_method, inLengths, reduceDims, alpha, beta);
|
||||
break;
|
||||
case 3:
|
||||
result = test_reduce_no_index_impl<ck::half_t, float, ck::half_t, Rank, 3>(
|
||||
init_method, inLengths, reduceDims, alpha, beta);
|
||||
break;
|
||||
case 4:
|
||||
result = test_reduce_no_index_impl<ck::half_t, float, ck::half_t, Rank, 4>(
|
||||
init_method, inLengths, reduceDims, alpha, beta);
|
||||
break;
|
||||
};
|
||||
result = profile_reduce_impl<ck::half_t, float, ck::half_t>(true,
|
||||
init_method,
|
||||
false,
|
||||
false,
|
||||
inLengths,
|
||||
reduceDims,
|
||||
reduceOpId,
|
||||
propagateNan,
|
||||
false,
|
||||
alpha,
|
||||
beta);
|
||||
}
|
||||
else if(data_type == 3)
|
||||
{
|
||||
switch(reduceDims.size())
|
||||
{
|
||||
case 1:
|
||||
result = test_reduce_no_index_impl<int8_t, int32_t, int8_t, Rank, 1>(
|
||||
init_method, inLengths, reduceDims, alpha, beta);
|
||||
break;
|
||||
case 3:
|
||||
result = test_reduce_no_index_impl<int8_t, int32_t, int8_t, Rank, 3>(
|
||||
init_method, inLengths, reduceDims, alpha, beta);
|
||||
break;
|
||||
case 4:
|
||||
result = test_reduce_no_index_impl<int8_t, int32_t, int8_t, Rank, 4>(
|
||||
init_method, inLengths, reduceDims, alpha, beta);
|
||||
break;
|
||||
};
|
||||
result = profile_reduce_impl<int8_t, int32_t, int8_t>(true,
|
||||
init_method,
|
||||
false,
|
||||
false,
|
||||
inLengths,
|
||||
reduceDims,
|
||||
reduceOpId,
|
||||
propagateNan,
|
||||
false,
|
||||
alpha,
|
||||
beta);
|
||||
}
|
||||
else if(data_type == 5)
|
||||
{
|
||||
switch(reduceDims.size())
|
||||
{
|
||||
case 1:
|
||||
result = test_reduce_no_index_impl<ck::bhalf_t, float, ck::bhalf_t, Rank, 1>(
|
||||
init_method, inLengths, reduceDims, alpha, beta);
|
||||
break;
|
||||
case 3:
|
||||
result = test_reduce_no_index_impl<ck::bhalf_t, float, ck::bhalf_t, Rank, 3>(
|
||||
init_method, inLengths, reduceDims, alpha, beta);
|
||||
break;
|
||||
case 4:
|
||||
result = test_reduce_no_index_impl<ck::bhalf_t, float, ck::bhalf_t, Rank, 4>(
|
||||
init_method, inLengths, reduceDims, alpha, beta);
|
||||
break;
|
||||
};
|
||||
result = profile_reduce_impl<ck::bhalf_t, float, ck::bhalf_t>(true,
|
||||
init_method,
|
||||
false,
|
||||
false,
|
||||
inLengths,
|
||||
reduceDims,
|
||||
reduceOpId,
|
||||
propagateNan,
|
||||
false,
|
||||
alpha,
|
||||
beta);
|
||||
}
|
||||
else if(data_type == 6)
|
||||
{
|
||||
result = profile_reduce_impl<double, double, double>(true,
|
||||
init_method,
|
||||
false,
|
||||
false,
|
||||
inLengths,
|
||||
reduceDims,
|
||||
reduceOpId,
|
||||
propagateNan,
|
||||
false,
|
||||
alpha,
|
||||
beta);
|
||||
}
|
||||
|
||||
return (result);
|
||||
};
|
||||
|
||||
constexpr ReduceTensorOp reduceOpId = ReduceTensorOp::AVG;
|
||||
constexpr bool propagateNan = false;
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
SimpleAppArgs args;
|
||||
@@ -621,8 +212,14 @@ int main(int argc, char* argv[])
|
||||
{0, 1, 2, 3}, {0, 1, 2}, {1, 2, 3}, {0, 1, 3}, {0, 2, 3}, {0}, {1}, {2}, {3}};
|
||||
|
||||
for(auto& reduceDims : v_reduceDims)
|
||||
result = result && test_reduce_no_index(
|
||||
data_type, init_method, reduceDims, inLengths, 1.0f, 0.0f);
|
||||
result = result && test_reduce_no_index(data_type,
|
||||
init_method,
|
||||
reduceDims,
|
||||
inLengths,
|
||||
reduceOpId,
|
||||
propagateNan,
|
||||
1.0f,
|
||||
0.0f);
|
||||
}
|
||||
else
|
||||
{
|
||||
@@ -636,6 +233,8 @@ int main(int argc, char* argv[])
|
||||
args.init_method,
|
||||
args.reduceDims,
|
||||
args.inLengths,
|
||||
reduceOpId,
|
||||
propagateNan,
|
||||
args.scales[0],
|
||||
args.scales[1]);
|
||||
}
|
||||
|
||||
Reference in New Issue
Block a user