mirror of
https://github.com/ROCm/composable_kernel.git
synced 2026-04-20 06:49:15 +00:00
Universal gemm splitk using reduce (with multi-d) (#1341)
* init for reduce_threadwise multi_d * add reduce_threadwise_multi_d * add reduce_multi_d * clean * start add an other splitk device op * add reduce template parameter to SplitKBatchOffset * add reduce c matrix * clean up code * change example data type to bf16 * add bf16Ai8B example * remove reduce template parameter * add splitk atomic status to v4 * example add multi d parameters * device op add multi-d parameters * add multi-d to reduce * fix kbach=1 bug * change B layout to col in bf16Ai8B example * remove float adding struct * change multi-d interface * change file and class name * remove multi-d of bf16Ai8B example * change IsReduce function to IsReduceAdd * change example layout to RRR from RCR * according layout to set ds stride * reset parameter layout * add gemm universal reduce instance * add reduce factory * add profile_gemm_universal_reduce * add reduce to profiler * fix reduce instance * fix profiler reduce compiling bug * format * format library instance code * add mem instance for reduce library * fix call instance names * add workspace for reduce in ckProfiler * format * add mnpading to reduce library instance * add fp16 instance to reduce of profiler * change copyright time * restore profiler cmake file * add reduce text to instances * add DsLayout and DsDataType to instances template parameter * fixed gemm_reduce_multi_d * add an example without multi_d * Update common.hpp * Update gtest.cmake * Update gemm_xdl_splitk_reduce_bf16.cpp * clean * Update gtest.cmake * format * fixe api * format * default parameter change to RRR * add vector_len for multi_d * format * Update gtest.cmake * fix bf16A iBB elementwiseop * add ReduceDataType * move ReduceDataType to end position * format * remove googletest git method address * fix copyright time * update init data --------- Co-authored-by: root <jizhan@amd.com> Co-authored-by: letaoqin <letaoqin@amd.com> Co-authored-by: Jing Zhang <jizhan@meta.com> Co-authored-by: zjing14 <zhangjing14@gmail.com>
This commit is contained in:
ltqin
@@ -1,3 +1,4 @@
|
||||
add_example_executable(example_reduce_blockwise reduce_blockwise.cpp)
|
||||
add_example_executable(example_reduce_threadwise_multi_d reduce_threadwise_multi_d.cpp)
|
||||
add_example_executable(example_reduce_multiblock_atomic_add reduce_multiblock_atomic_add.cpp)
|
||||
add_example_executable(example_reduce_blockwise_two_call reduce_blockwise_two_call.cpp)
|
||||
|
||||
229
example/12_reduce/reduce_threadwise_multi_d.cpp
Normal file
229
example/12_reduce/reduce_threadwise_multi_d.cpp
Normal file
@@ -0,0 +1,229 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#include <iostream>
|
||||
#include <initializer_list>
|
||||
#include <cstdlib>
|
||||
#include <getopt.h>
|
||||
|
||||
#include "ck/utility/reduction_enums.hpp"
|
||||
#include "reduce_threadwise_multi_d_impl.hpp"
|
||||
#include "reduce_example_common.hpp"
|
||||
|
||||
using namespace ck;
|
||||
using namespace ck::tensor_operation::device;
|
||||
|
||||
static struct option long_options[] = {{"inLengths", required_argument, nullptr, 'D'},
|
||||
{"verify", required_argument, nullptr, 'v'},
|
||||
{"help", no_argument, nullptr, '?'},
|
||||
{nullptr, 0, nullptr, 0}};
|
||||
|
||||
class SimpleAppArgs
|
||||
{
|
||||
private:
|
||||
int option_index = 0;
|
||||
|
||||
public:
|
||||
std::vector<size_t> inLengths = {16, 64, 32, 16};
|
||||
std::vector<int> reduceDims = {0};
|
||||
std::vector<float> scales = {1.0f, 0.0f};
|
||||
|
||||
bool do_verification = true;
|
||||
int data_type = 1;
|
||||
int init_method = 2;
|
||||
bool time_kernel = true;
|
||||
|
||||
public:
|
||||
void show_usage(const char* cmd)
|
||||
{
|
||||
std::cout << "Usage of " << cmd << std::endl;
|
||||
std::cout << "--inLengths or -D, comma separated list of input tensor dimension lengths"
|
||||
<< std::endl;
|
||||
std::cout << "--reduceDims or -R, comma separated list of to-reduce dimensions"
|
||||
<< std::endl;
|
||||
std::cout << "--verify or -v, 1/0 to indicate whether to verify the reduction result by "
|
||||
"comparing with the host-based reduction"
|
||||
<< std::endl;
|
||||
std::cout << "Arg1: data type (0: fp16, 1: fp32, 3: int8, 5: bp16, 6: fp64, 7: int4)"
|
||||
<< std::endl;
|
||||
std::cout << "Arg2 -- init method (0=no init, 1=single integer value, 2=scope integer "
|
||||
"value, 3=decimal value)"
|
||||
<< std::endl;
|
||||
std::cout << "Arg3 -- time kernel (0=no, 1=yes)" << std::endl;
|
||||
};
|
||||
|
||||
int processArgs(int argc, char* argv[])
|
||||
{
|
||||
using ck::host_common::getTypeValuesFromString;
|
||||
|
||||
int ch;
|
||||
|
||||
while(1)
|
||||
{
|
||||
ch = getopt_long(argc, argv, "D:R:v:l:", long_options, &option_index);
|
||||
if(ch == -1)
|
||||
break;
|
||||
switch(ch)
|
||||
{
|
||||
case 'D':
|
||||
if(!optarg)
|
||||
throw std::runtime_error("Invalid option format!");
|
||||
|
||||
inLengths = getTypeValuesFromString<size_t>(optarg);
|
||||
break;
|
||||
case 'R':
|
||||
if(!optarg)
|
||||
throw std::runtime_error("Invalid option format!");
|
||||
|
||||
reduceDims = getTypeValuesFromString<int>(optarg);
|
||||
break;
|
||||
case 'v':
|
||||
if(!optarg)
|
||||
throw std::runtime_error("Invalid option format!");
|
||||
|
||||
do_verification = static_cast<bool>(std::atoi(optarg));
|
||||
break;
|
||||
case '?':
|
||||
if(std::string(long_options[option_index].name) == "help")
|
||||
{
|
||||
show_usage(argv[0]);
|
||||
return (-1);
|
||||
};
|
||||
break;
|
||||
default: show_usage(argv[0]); return (-1);
|
||||
};
|
||||
};
|
||||
|
||||
if(optind + 3 > argc)
|
||||
{
|
||||
throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
|
||||
};
|
||||
|
||||
data_type = std::atoi(argv[optind++]);
|
||||
init_method = std::atoi(argv[optind++]);
|
||||
time_kernel = static_cast<bool>(std::atoi(argv[optind]));
|
||||
|
||||
if(scales.empty())
|
||||
{
|
||||
scales.push_back(1.0f);
|
||||
scales.push_back(0.0f);
|
||||
};
|
||||
|
||||
return (0);
|
||||
};
|
||||
};
|
||||
|
||||
template <typename InOutDataType,
|
||||
typename AccDataType,
|
||||
ReduceTensorOp ReduceOpId,
|
||||
index_t PropagateNan,
|
||||
index_t OutputIndex>
|
||||
bool reduce_threadwise_multi_d_test(bool do_verification,
|
||||
int init_method,
|
||||
bool time_kernel,
|
||||
const std::vector<size_t>& inLengths,
|
||||
const std::vector<int>& reduceDims,
|
||||
float alpha,
|
||||
float beta)
|
||||
{
|
||||
bool matched = false;
|
||||
int result = 0;
|
||||
|
||||
const auto tuple_object = reduce_shape_instances{};
|
||||
|
||||
static_for<0, std::tuple_size<reduce_shape_instances>::value, 1>{}([&](auto i) {
|
||||
if(matched)
|
||||
return;
|
||||
|
||||
using ShapeType = remove_cvref_t<decltype(std::get<i>(tuple_object))>;
|
||||
|
||||
if(ShapeType::Rank_ != inLengths.size() || ShapeType::NumReduceDim_ != reduceDims.size())
|
||||
return;
|
||||
|
||||
std::array<int, ShapeType::NumReduceDim_> arrReduceDims;
|
||||
|
||||
ck::ranges::copy(reduceDims, arrReduceDims.begin());
|
||||
|
||||
result = reduce_threadwise_multi_d_impl<InOutDataType,
|
||||
AccDataType,
|
||||
ReduceOpId,
|
||||
ShapeType::Rank_,
|
||||
ShapeType::NumReduceDim_,
|
||||
PropagateNan,
|
||||
OutputIndex>(
|
||||
do_verification, init_method, time_kernel, inLengths, arrReduceDims, alpha, beta);
|
||||
|
||||
matched = true;
|
||||
});
|
||||
|
||||
return (result == 0) ? true : false;
|
||||
};
|
||||
|
||||
constexpr ReduceTensorOp ReduceOpId = ReduceTensorOp::AVG;
|
||||
constexpr bool PropagateNan = true;
|
||||
constexpr bool OutputIndex = false;
|
||||
|
||||
int main(int argc, char* argv[])
|
||||
{
|
||||
bool pass = true;
|
||||
|
||||
if(argc > 1)
|
||||
{
|
||||
SimpleAppArgs arg;
|
||||
|
||||
if(arg.processArgs(argc, argv) < 0)
|
||||
return (-1);
|
||||
|
||||
if(arg.data_type == 0)
|
||||
{
|
||||
pass = reduce_threadwise_multi_d_test<ck::half_t,
|
||||
float,
|
||||
ReduceOpId,
|
||||
PropagateNan,
|
||||
OutputIndex>(arg.do_verification,
|
||||
arg.init_method,
|
||||
arg.time_kernel,
|
||||
arg.inLengths,
|
||||
arg.reduceDims,
|
||||
arg.scales[0],
|
||||
arg.scales[1]);
|
||||
}
|
||||
else if(arg.data_type == 1)
|
||||
{
|
||||
pass =
|
||||
reduce_threadwise_multi_d_test<float, float, ReduceOpId, PropagateNan, OutputIndex>(
|
||||
arg.do_verification,
|
||||
arg.init_method,
|
||||
arg.time_kernel,
|
||||
arg.inLengths,
|
||||
arg.reduceDims,
|
||||
arg.scales[0],
|
||||
arg.scales[1]);
|
||||
}
|
||||
}
|
||||
else
|
||||
{
|
||||
// for testing half_t
|
||||
pass = pass && reduce_threadwise_multi_d_test<ck::half_t,
|
||||
float,
|
||||
ReduceOpId,
|
||||
PropagateNan,
|
||||
OutputIndex>(
|
||||
true, 2, true, {16, 64, 32, 960}, {0}, 1.0f, 0.0f);
|
||||
|
||||
// for testing float
|
||||
pass = pass &&
|
||||
reduce_threadwise_multi_d_test<float, float, ReduceOpId, PropagateNan, OutputIndex>(
|
||||
true, 2, true, {16, 64, 32, 960}, {0}, 1.0f, 0.0f);
|
||||
|
||||
// for testing bhalf_t
|
||||
pass = pass && reduce_threadwise_multi_d_test<ck::bhalf_t,
|
||||
float,
|
||||
ReduceOpId,
|
||||
PropagateNan,
|
||||
OutputIndex>(
|
||||
true, 2, true, {16, 64, 32, 960}, {0}, 1.0f, 0.0f);
|
||||
}
|
||||
|
||||
return (pass ? 0 : 1);
|
||||
};
|
||||
307
example/12_reduce/reduce_threadwise_multi_d_impl.hpp
Normal file
307
example/12_reduce/reduce_threadwise_multi_d_impl.hpp
Normal file
@@ -0,0 +1,307 @@
|
||||
// SPDX-License-Identifier: MIT
|
||||
// Copyright (c) 2018-2024, Advanced Micro Devices, Inc. All rights reserved.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <iostream>
|
||||
|
||||
#include "ck/ck.hpp"
|
||||
#include "ck/utility/reduction_enums.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
|
||||
#include "ck/tensor_operation/gpu/device/impl/device_reduce_threadwise_multi_d.hpp"
|
||||
#include "ck/library/reference_tensor_operation/cpu/reference_reduce.hpp"
|
||||
|
||||
#include "ck/library/utility/algorithm.hpp"
|
||||
#include "ck/library/utility/check_err.hpp"
|
||||
#include "ck/library/utility/device_memory.hpp"
|
||||
#include "ck/library/utility/host_tensor.hpp"
|
||||
#include "ck/library/utility/host_tensor_generator.hpp"
|
||||
#include "ck/library/utility/host_common_util.hpp"
|
||||
|
||||
#include "reduce_example_common.hpp"
|
||||
|
||||
template <typename InOutDataType,
|
||||
typename AccDataType,
|
||||
ck::ReduceTensorOp ReduceOpId,
|
||||
ck::index_t Rank,
|
||||
ck::index_t NumReduceDim,
|
||||
bool PropagateNan,
|
||||
bool OutputIndex>
|
||||
int reduce_threadwise_multi_d_impl(bool do_verification,
|
||||
int init_method,
|
||||
bool time_kernel,
|
||||
const std::vector<size_t>& inLengths,
|
||||
const std::array<int, NumReduceDim>& reduceDims,
|
||||
float alpha,
|
||||
float beta)
|
||||
|
||||
{
|
||||
using namespace ck;
|
||||
using namespace ck::tensor_operation::device;
|
||||
|
||||
constexpr index_t NumOutDim = (Rank - NumReduceDim == 0) ? 1 : Rank - NumReduceDim;
|
||||
|
||||
constexpr bool op_support_indices =
|
||||
(ReduceOpId == ReduceTensorOp::MIN || ReduceOpId == ReduceTensorOp::MAX ||
|
||||
ReduceOpId == ReduceTensorOp::AMAX);
|
||||
|
||||
constexpr bool invalid_reduce_1 = OutputIndex && !op_support_indices;
|
||||
|
||||
// 1) If InOutDataType is half_t, must use half_t as AccDataType for indexable reduction
|
||||
// operations 2) If InOutDataType is half_t, must use float as AccDataType for non-indexable
|
||||
// reduction operations
|
||||
constexpr bool invalid_reduce_2 =
|
||||
std::is_same<InOutDataType, half_t>::value &&
|
||||
((!op_support_indices && !std::is_same<AccDataType, float>::value) ||
|
||||
(op_support_indices && !std::is_same<AccDataType, half_t>::value));
|
||||
|
||||
// 1) If InOutDataType is float, must use float as AccDataType for indexable reduction
|
||||
// operations
|
||||
constexpr bool invalid_reduce_3 =
|
||||
std::is_same<InOutDataType, float>::value &&
|
||||
(op_support_indices && !std::is_same<AccDataType, float>::value);
|
||||
|
||||
// 1) If InOutDataType is int8_t or int4_t, must use int8_t as AccDataType for indexable
|
||||
// reduction operations 2) If InOutDataType is int8_t or int4_t, must use int32_t as AccDataType
|
||||
// for non-indexable reduction operations
|
||||
constexpr bool invalid_reduce_4 =
|
||||
std::is_same<InOutDataType, int8_t>::value &&
|
||||
((!op_support_indices && !std::is_same<AccDataType, int32_t>::value) ||
|
||||
(op_support_indices && !std::is_same<AccDataType, int8_t>::value));
|
||||
|
||||
// 1) If InOutDataType is int8_t or int4_t, the supported operation must be either indexable
|
||||
// operations or ADD/AVG
|
||||
constexpr bool invalid_reduce_5 = std::is_same<InOutDataType, int8_t>::value &&
|
||||
(!op_support_indices && ReduceOpId != ReduceTensorOp::ADD &&
|
||||
ReduceOpId != ReduceTensorOp::AVG);
|
||||
|
||||
// 1) If InOutDataType is bhalf_t, must use float as AccDataType for all reduction operations
|
||||
constexpr bool invalid_reduce_6 =
|
||||
std::is_same<InOutDataType, bhalf_t>::value && !std::is_same<AccDataType, float>::value;
|
||||
|
||||
constexpr bool invalid_reduce = (invalid_reduce_1 || invalid_reduce_2 || invalid_reduce_3 ||
|
||||
invalid_reduce_4 || invalid_reduce_5 || invalid_reduce_6);
|
||||
|
||||
if constexpr(invalid_reduce)
|
||||
{
|
||||
std::cerr << "The reduction setting is invalid, exiting!" << std::endl;
|
||||
return (-1);
|
||||
};
|
||||
|
||||
using PassThrough = tensor_operation::element_wise::PassThrough;
|
||||
using Add = tensor_operation::element_wise::Add;
|
||||
|
||||
using ReduceOperation = typename reduce_binary_operator<ReduceOpId>::opType;
|
||||
using InElementwiseOperation = PassThrough;
|
||||
using OutElementwiseOperation = Add;
|
||||
|
||||
using InOutDataTypeInDevice = InOutDataType;
|
||||
|
||||
using DeviceReduceInstance =
|
||||
ck::tensor_operation::device::DeviceReduceThreadWiseMultiD<InOutDataTypeInDevice,
|
||||
ck::Tuple<InOutDataTypeInDevice>,
|
||||
AccDataType,
|
||||
InOutDataTypeInDevice,
|
||||
Rank,
|
||||
NumReduceDim,
|
||||
ReduceOperation,
|
||||
InElementwiseOperation,
|
||||
OutElementwiseOperation,
|
||||
256, // BlockSize
|
||||
4, // MThreadSliceSize
|
||||
1, // KThreadSliceSize
|
||||
0, // InSrcVectorDim
|
||||
1, // InSrceVectorSize
|
||||
1,
|
||||
Sequence<1>>; // OutDstVectorSize
|
||||
|
||||
Tensor<InOutDataType> in(inLengths);
|
||||
|
||||
std::vector<size_t> outLengths;
|
||||
|
||||
auto invariantDims = get_invariant_dims<Rank, NumReduceDim>(reduceDims);
|
||||
|
||||
if(invariantDims.empty())
|
||||
outLengths.push_back(1);
|
||||
else
|
||||
for(auto dim : invariantDims)
|
||||
outLengths.push_back(inLengths[dim]);
|
||||
|
||||
Tensor<InOutDataType> out_ref(outLengths);
|
||||
Tensor<InOutDataType> out(outLengths);
|
||||
|
||||
Tensor<InOutDataType> d0(outLengths);
|
||||
|
||||
Tensor<int> out_indices_ref(outLengths);
|
||||
Tensor<int> out_indices(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;
|
||||
|
||||
if(do_verification)
|
||||
{
|
||||
switch(init_method)
|
||||
{
|
||||
case 0: break;
|
||||
case 1:
|
||||
in.GenerateTensorValue(GeneratorTensor_1<InOutDataType>{1}, num_thread);
|
||||
d0.GenerateTensorValue(GeneratorTensor_1<InOutDataType>{1}, num_thread);
|
||||
if(beta != 0.0f)
|
||||
out_ref.GenerateTensorValue(GeneratorTensor_1<InOutDataType>{1}, num_thread);
|
||||
break;
|
||||
case 2:
|
||||
in.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
|
||||
d0.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
|
||||
if(beta != 0.0f)
|
||||
out_ref.GenerateTensorValue(GeneratorTensor_2<InOutDataType>{-5, 5}, num_thread);
|
||||
break;
|
||||
default:
|
||||
in.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0, 5.0}, num_thread);
|
||||
d0.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0, 5.0}, num_thread);
|
||||
if(beta != 0.0f)
|
||||
out_ref.GenerateTensorValue(GeneratorTensor_3<InOutDataType>{-5.0, 5.0},
|
||||
num_thread);
|
||||
}
|
||||
|
||||
if(beta != 0.0f)
|
||||
for(size_t i = 0; i < out_ref.mDesc.GetElementSpaceSize(); i++)
|
||||
out.mData[i] = out_ref.mData[i];
|
||||
};
|
||||
|
||||
// these buffers are usually provided by the user application
|
||||
DeviceMem in_dev(sizeof(InOutDataTypeInDevice) * in.mDesc.GetElementSpaceSize());
|
||||
DeviceMem d0_dev(sizeof(InOutDataTypeInDevice) * d0.mDesc.GetElementSpaceSize());
|
||||
DeviceMem out_dev(sizeof(InOutDataTypeInDevice) * out.mDesc.GetElementSpaceSize());
|
||||
|
||||
in_dev.ToDevice(in.mData.data());
|
||||
d0_dev.ToDevice(d0.mData.data());
|
||||
|
||||
if(beta != 0.0f)
|
||||
{
|
||||
out_dev.ToDevice(out.mData.data());
|
||||
};
|
||||
|
||||
size_t indicesSizeInBytes = OutputIndex ? out.mDesc.GetElementSize() * sizeof(int32_t) : 0;
|
||||
|
||||
DeviceMem out_index_dev(indicesSizeInBytes);
|
||||
|
||||
InElementwiseOperation in_elementwise_op;
|
||||
OutElementwiseOperation out_elementwise_op;
|
||||
|
||||
std::array<index_t, Rank> arrInLengths;
|
||||
std::array<index_t, Rank> arrInStrides;
|
||||
|
||||
std::array<index_t, NumOutDim> arrOutLengths;
|
||||
std::array<index_t, NumOutDim> arrOutStrides;
|
||||
|
||||
ck::ranges::copy(inLengths, arrInLengths.begin());
|
||||
ck::ranges::copy(inStrides, arrInStrides.begin());
|
||||
|
||||
ck::ranges::copy(outLengths, arrOutLengths.begin());
|
||||
ck::ranges::copy(outStrides, arrOutStrides.begin());
|
||||
|
||||
if(do_verification)
|
||||
{
|
||||
using ReferenceReduceInstance =
|
||||
ck::tensor_operation::host::ReferenceReduce<InOutDataType,
|
||||
AccDataType,
|
||||
InOutDataType,
|
||||
Rank,
|
||||
NumReduceDim,
|
||||
ReduceOperation,
|
||||
InElementwiseOperation,
|
||||
PassThrough,
|
||||
PropagateNan,
|
||||
OutputIndex>;
|
||||
|
||||
auto reduce_ref = ReferenceReduceInstance{};
|
||||
|
||||
auto argument_ptr_ref = reduce_ref.MakeArgumentPointer(arrInLengths,
|
||||
arrInStrides,
|
||||
arrOutLengths,
|
||||
arrOutStrides,
|
||||
reduceDims,
|
||||
static_cast<double>(alpha),
|
||||
static_cast<double>(beta),
|
||||
in.mData.data(),
|
||||
nullptr,
|
||||
out_ref.mData.data(),
|
||||
out_indices_ref.mData.data(),
|
||||
in_elementwise_op,
|
||||
PassThrough{});
|
||||
|
||||
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();
|
||||
|
||||
invoker_ptr_ref->Run(argument_ptr_ref.get());
|
||||
|
||||
for(std::size_t i = 0; i < out_ref.GetElementSize(); i++)
|
||||
out_elementwise_op(out_ref.mData[i], out_ref.mData[i], d0.mData[i]);
|
||||
};
|
||||
|
||||
auto reduce = DeviceReduceInstance{};
|
||||
|
||||
auto argument_ptr = reduce.MakeArgumentPointer(arrInLengths,
|
||||
arrInStrides,
|
||||
{arrOutLengths},
|
||||
{arrOutStrides},
|
||||
arrOutLengths,
|
||||
arrOutStrides,
|
||||
reduceDims,
|
||||
in_dev.GetDeviceBuffer(),
|
||||
{d0_dev.GetDeviceBuffer()},
|
||||
out_dev.GetDeviceBuffer(),
|
||||
in_elementwise_op,
|
||||
out_elementwise_op);
|
||||
|
||||
if(!reduce.IsSupportedArgument(argument_ptr.get()))
|
||||
{
|
||||
std::cerr << "The runtime parameters not supported by the DeviceReduce instance, exiting!"
|
||||
<< std::endl;
|
||||
|
||||
return (-2);
|
||||
};
|
||||
|
||||
std::string reduce_name = reduce.GetTypeString();
|
||||
|
||||
auto invoker_ptr = reduce.MakeInvokerPointer();
|
||||
|
||||
float avg_time = invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
|
||||
|
||||
std::size_t num_bytes = invariant_total_length * reduce_total_length * sizeof(InOutDataType) +
|
||||
invariant_total_length * sizeof(InOutDataType);
|
||||
|
||||
float gb_per_sec = num_bytes / 1.E6 / avg_time;
|
||||
|
||||
std::cout << "Perf: " << avg_time << " ms, " << gb_per_sec << " GB/s, " << reduce_name
|
||||
<< std::endl;
|
||||
|
||||
bool pass = true;
|
||||
|
||||
if(do_verification)
|
||||
{
|
||||
|
||||
out_dev.FromDevice(out.mData.data());
|
||||
|
||||
pass = pass && ck::utils::check_err(out, out_ref);
|
||||
|
||||
if(OutputIndex)
|
||||
{
|
||||
out_index_dev.FromDevice(out_indices.mData.data());
|
||||
pass = pass && ck::utils::check_err(out_indices, out_indices_ref);
|
||||
};
|
||||
};
|
||||
|
||||
return (pass ? 0 : 1);
|
||||
}
|
||||
Reference in New Issue
Block a user