ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-14 10:09:41 +00:00
Code Issues Packages Projects Releases Wiki Activity

Overhaul to Reducton and its dependants (#237)

Browse Source 
* 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

[ROCm/composable_kernel commit: 63eee2d999]
This commit is contained in:
Qianfeng
2022-05-25 01:19:12 +08:00
committed by GitHub
parent eb85e53cb6
commit 5bcaa25bb7
94 changed files with 2429 additions and 6785 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
example/12_reduce/CMakeLists.txt
View File

@@ -1 +1,2 @@
add_example_executable(example_reduce_blockwise reduce_blockwise.cpp -D 16,64,32,960 -v 1 1 10)
add_example_executable(example_reduce_blockwise reduce_blockwise.cpp)
add_example_executable(example_reduce_blockwise_two_call reduce_blockwise_two_call.cpp)

41
example/12_reduce/README.md
View File

@@ -5,23 +5,38 @@
# -D <xxx> : input 4-d tensor lengths
# -v <x> : verification (0=no, 1=yes)
#arg1: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
#arg2: run kernel # of times (>1)
./bin/example_reduce_blockwise -D 16,64,32,960 -v 1 1 10
#arg2: time kernel (0=no, 1=yes)
./bin/example_reduce_blockwise -D 16,64,32,960 -v 1 1 1
```
Result
```
./bin/example_reduce_blockwise -D 16,64,32,960 -v 1 1 1
launch_and_time_kernel: grid_dim {240, 1, 1}, block_dim {256, 1, 1}
Warm up
Start running 3 times...
Perf: 0.23536 ms, 267.32 GB/s, DeviceReduceBlockWise<256,M_C4_S1,K_C64_S1,InSrcVectorDim_0_InSrcVectorSize_1_OutDstVectorSize_1>
error: 0
max_diff: 0, 529, 529
root@dc-smc-18:/data/composable_kernel/Build3# bin/example_reduce_blockwise -D 16,64,32,960 -v 1 1 10
launch_and_time_kernel: grid_dim {240, 1, 1}, block_dim {256, 1, 1}
Warm up
Warm up 1 time
Start running 10 times...
Perf: 0.23392 ms, 268.966 GB/s, DeviceReduceBlockWise<256,M_C4_S1,K_C64_S1,InSrcVectorDim_0_InSrcVectorSize_1_OutDstVectorSize_1>
error: 0
max_diff: 0, 528, 528
Perf: 0.282592 ms, 222.641 GB/s, DeviceReduceBlockWise<256,M_C4_S1,K_C64_S1,InSrcVectorDim_0_InSrcVectorSize_1_OutDstVectorSize_1>
```
# Instructions for ```example_reduce_blockwise_two_call```
## Run ```example_reduce_blockwise_two_call```
```bash
#arg1: verification (0=no, 1=yes(
#arg2: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
#arg3: time kernel (0=no, 1=yes)
./bin/example_reduce_blockwise_two_call 1 2 1
Result
```
./bin/example_reduce_blockwise_two_call 1 2 1
launch_and_time_kernel: grid_dim {204800, 1, 1}, block_dim {256, 1, 1}
Warm up 1 time
Start running 10 times...
launch_and_time_kernel: grid_dim {6400, 1, 1}, block_dim {256, 1, 1}
Warm up 1 time
Start running 10 times...
Perf: 2.1791 ms, 771.42 GB/s, DeviceReduceBlockWise<256,M_C32_S1,K_C8_S1,InSrcVectorDim_1_InSrcVectorSize_1_OutDstVectorSize_1> => DeviceReduceBlockWise<256,M_C256_S1,K_C1_S1,InSrcVectorDim_1_InSrcVectorSize_1_OutDstVectorSize_1>
```

188
example/12_reduce/reduce_blockwise.cpp
View File

@@ -12,8 +12,8 @@
#include "host_tensor_generator.hpp"
#include "device_tensor.hpp"
#include "device_base.hpp"
#include "device_reduce_blockwise.hpp"
#include "host_reduce_util.hpp"
#include "device_reduce_multiblock.hpp"
#include "host_common_util.hpp"
#include "host_reduction.hpp"
#include "reduction_enums.hpp"
@@ -30,9 +30,8 @@ constexpr int Rank = 4;
constexpr int NumReduceDim = 3;
constexpr ReduceTensorOp ReduceOpId = ReduceTensorOp::NORM2;
constexpr NanPropagation NanOpt = NanPropagation::PROPAGATE_NAN;
constexpr bool PropagateNan = (NanOpt == NanPropagation::NOT_PROPAGATE_NAN) ? false : true;
constexpr ReduceTensorIndices IndicesOpt = ReduceTensorIndices::NO_INDICES;
constexpr bool PropagateNan = true;
constexpr bool OutputIndex = false;
using ReduceOperation = typename reduce_binary_operator<AccDataType, ReduceOpId>::opType;
using InElementwiseOperation =
@@ -40,85 +39,44 @@ using InElementwiseOperation =
using AccElementwiseOperation =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::AccElementwiseOperation;
using DeviceReduceInstance = DeviceReduceBlockWise<InDataType,
AccDataType,
OutDataType,
Rank,
NumReduceDim,
ReduceOperation,
InElementwiseOperation,
AccElementwiseOperation,
PropagateNan,
false,
256,
4,
64,
1,
1,
0,
1,
1>;
using DeviceReduceInstance = DeviceReduceMultiBlock<InDataType,
AccDataType,
OutDataType,
Rank,
NumReduceDim,
ReduceOperation,
InElementwiseOperation,
AccElementwiseOperation,
InMemoryDataOperationEnum::Set,
PropagateNan,
OutputIndex,
false, // HaveIndexInputIfOutputIndex
256,
4,
64,
1,
1,
0,
1,
1>;
static struct option long_options[] = {{"inLengths", required_argument, nullptr, 'D'},
{"scales", required_argument, nullptr, 'S'},
{"verify", required_argument, nullptr, 'v'},
{"help", no_argument, nullptr, '?'},
{nullptr, 0, nullptr, 0}};
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;
public:
std::vector<size_t> inLengths;
std::vector<float> scales;
std::vector<size_t> inLengths = {16, 64, 32, 960};
std::vector<float> scales = {1.0f, 0.0f};
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
bool time_kernel = true;
public:
void show_usage(const char* cmd)
@@ -126,24 +84,24 @@ class SimpleAppArgs
std::cout << "Usage of " << cmd << std::endl;
std::cout << "--inLengths or -D, comma separated list of input tensor dimension lengths"
<< std::endl;
std::cout << "--scales or -S, comma separated two float values for alpha and beta"
<< 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 -- init method (0=no init, 1=single integer value, 2=scope integer "
"value, 3=decimal value)"
<< std::endl;
std::cout << "Arg2 -- time kernel (0=n0, 1=yes)" << std::endl;
std::cout << "Arg2 -- 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:S:v:l:", long_options, &option_index);
ch = getopt_long(argc, argv, "D:v:l:", long_options, &option_index);
if(ch == -1)
break;
switch(ch)
@@ -154,12 +112,6 @@ class SimpleAppArgs
inLengths = getTypeValuesFromString<size_t>(optarg);
break;
case 'S':
if(!optarg)
throw std::runtime_error("Invalid option format!");
scales = getTypeValuesFromString<float>(optarg);
break;
case 'v':
if(!optarg)
throw std::runtime_error("Invalid option format!");
@@ -181,7 +133,7 @@ class SimpleAppArgs
throw std::runtime_error("Invalid cmd-line arguments, more argumetns are needed!");
init_method = std::atoi(argv[optind++]);
time_kernel = std::atoi(argv[optind]);
time_kernel = static_cast<bool>(std::atoi(argv[optind]));
if(scales.empty())
{
@@ -202,16 +154,16 @@ int main(int argc, char* argv[])
SimpleAppArgs args;
if(args.processArgs(argc, argv) < 0)
return (-1);
if(argc > 1)
{
if(args.processArgs(argc, argv) < 0)
return (-1);
};
constexpr bool op_support_indices =
(ReduceOpId == ReduceTensorOp::MIN || ReduceOpId == ReduceTensorOp::MAX ||
ReduceOpId == ReduceTensorOp::AMAX);
constexpr bool NeedIndices =
(op_support_indices && (IndicesOpt != ReduceTensorIndices::NO_INDICES));
// if input is half type, no reason to use float for indiced reduction operation and must use
// float for non-indiced reduction operation for accuracy
constexpr bool invalid_reduce_1 =
@@ -225,8 +177,7 @@ int main(int argc, char* argv[])
(op_support_indices && !std::is_same<AccDataType, float>::value);
// indices option can only be used when it is really needed
constexpr bool invalid_reduce_3 =
(!op_support_indices && IndicesOpt != ReduceTensorIndices::NO_INDICES);
constexpr bool invalid_reduce_3 = (!op_support_indices && OutputIndex);
constexpr bool invalid_reduce = (invalid_reduce_1 || invalid_reduce_2 || invalid_reduce_3);
@@ -294,9 +245,9 @@ int main(int argc, char* argv[])
if(beta != 0.0f)
out_dev.ToDevice(out.mData.data());
size_t indicesSizeInBytes = NeedIndices ? out.mDesc.GetElementSize() * sizeof(int32_t) : 0;
size_t indicesSizeInBytes = OutputIndex ? out.mDesc.GetElementSize() * sizeof(int32_t) : 0;
DeviceMem out_indices_dev(indicesSizeInBytes);
DeviceMem out_index_dev(indicesSizeInBytes);
if(args.do_verification)
{
@@ -307,38 +258,39 @@ int main(int argc, char* argv[])
Rank,
NumReduceDim,
PropagateNan,
NeedIndices>
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());
};
const auto i_inLengths = to_int_vector(args.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);
std::vector<ck::index_t> i_inLengths;
std::vector<ck::index_t> i_inStrides;
std::vector<ck::index_t> i_outLengths;
std::vector<ck::index_t> i_outStrides;
i_inLengths.assign(args.inLengths.begin(), args.inLengths.end());
i_inStrides.assign(inStrides.begin(), inStrides.end());
i_outLengths.assign(outLengths.begin(), outLengths.end());
i_outStrides.assign(outStrides.begin(), outStrides.end());
auto reduce = DeviceReduceInstance{};
auto wsSizeInBytes = reduce.GetWorkspaceSizeInBytes(i_inLengths, reduceDims);
DeviceMem ws_dev(wsSizeInBytes);
auto argument_ptr =
reduce.MakeArgumentPointer(i_inLengths,
i_inStrides,
i_outLengths,
i_outStrides,
reduceDims,
alpha,
beta,
in_dev.GetDeviceBuffer(),
out_dev.GetDeviceBuffer(),
out_indices_dev.GetDeviceBuffer(),
ws_dev.GetDeviceBuffer(),
InElementwiseOperation{static_cast<int>(reduce_total_length)},
AccElementwiseOperation{static_cast<int>(reduce_total_length)});
auto argument_ptr = reduce.MakeArgumentPointer(
i_inLengths,
i_inStrides,
i_outLengths,
i_outStrides,
reduceDims,
alpha,
beta,
in_dev.GetDeviceBuffer(),
nullptr,
out_dev.GetDeviceBuffer(),
out_index_dev.GetDeviceBuffer(),
InElementwiseOperation{static_cast<int32_t>(reduce_total_length)},
AccElementwiseOperation{static_cast<int32_t>(reduce_total_length)});
if(!reduce.IsSupportedArgument(argument_ptr.get()))
{
@@ -362,16 +314,18 @@ int main(int argc, char* argv[])
<< std::endl;
bool pass = true;
if(args.do_verification)
{
out_dev.FromDevice(out.mData.data());
pass &= ck::utils::check_err(out.mData, out_ref.mData);
pass = pass && ck::utils::check_err(out.mData, out_ref.mData);
if(NeedIndices)
if(OutputIndex)
{
out_indices_dev.FromDevice(out_indices.mData.data());
pass &= ck::utils::check_err(out_indices.mData, out_indices_ref.mData);
out_index_dev.FromDevice(out_indices.mData.data());
pass = pass && ck::utils::check_err(out_indices.mData, out_indices_ref.mData);
};
};
return pass ? 0 : 1;
return (pass ? 0 : 1);
}

290
example/12_reduce/reduce_blockwise_two_call.cpp Normal file
View File

@@ -0,0 +1,290 @@
#include <iostream>
#include <numeric>
#include <sstream>
#include <initializer_list>
#include <cstdlib>
#include <getopt.h>
#include "check_err.hpp"
#include "config.hpp"
#include "print.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "device_tensor.hpp"
#include "device_base.hpp"
#include "device_reduce_multiblock.hpp"
#include "host_common_util.hpp"
#include "host_reduction.hpp"
#include "reduction_enums.hpp"
#include "reduction_operator_mapping.hpp"
using namespace ck;
using namespace ck::tensor_operation::device;
using InOutDataType = ck::half_t;
using InOutDataType = ck::half_t;
using AccDataType = float;
constexpr ReduceTensorOp ReduceOpId = ReduceTensorOp::NORM2;
constexpr bool PropagateNan = true;
constexpr bool OutputIndex = false;
using ReduceOperation = typename reduce_binary_operator<AccDataType, ReduceOpId>::opType;
using InElementwiseOperation =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::InElementwiseOperation;
using AccElementwiseOperation =
typename reduce_unary_operator<AccDataType, ReduceOpId, true, true>::AccElementwiseOperation;
using PassThroughOp = tensor_operation::element_wise::UnaryIdentic<AccDataType, AccDataType>;
using DeviceReduceInstance_1 = DeviceReduceMultiBlock<InOutDataType,
AccDataType,
InOutDataType,
5, // Rank
1, // NumReduceDim
ReduceOperation,
InElementwiseOperation,
PassThroughOp,
InMemoryDataOperationEnum::Set,
PropagateNan,
OutputIndex,
false, // HaveIndexInputIfOutputIndex
256,
32,
8,
1,
1,
1, // vector dim
1,
1>;
using DeviceReduceInstance_2 = DeviceReduceMultiBlock<InOutDataType,
AccDataType,
InOutDataType,
4, // Rank
1, // NumReduceDim
ReduceOperation,
PassThroughOp,
AccElementwiseOperation,
InMemoryDataOperationEnum::Set,
PropagateNan,
OutputIndex,
false, // HaveIndexInputIfOutputIndex
256,
128,
2,
1,
1,
1, // vector dim
1,
1>;
static bool do_verify;
static int init_method;
static float alpha;
static float beta;
static bool time_kernel;
int main(int argc, char* argv[])
{
// used by the device reduction
const std::vector<int> reduceDims_1 = {4};
const std::vector<int> invariantDims_1 = {0, 1, 2, 3};
const std::vector<int> reduceDims_2 = {3};
const std::vector<int> invariantDims_2 = {0, 1, 2};
// used by the host reduction
const std::vector<int> reduceDims = {3, 4};
const std::vector<int> invariantDims = {0, 1, 2};
const std::vector<size_t> inLengths_1 = {64, 320, 80, 4, 128};
// input lengths of the second reduction, which is also the output lengths of the first
// reduction
const std::vector<size_t> inLengths_2 = {64, 320, 80, 4};
const std::vector<size_t> outLengths = {64, 320, 80};
using namespace ck::host_reduce;
if(argc == 1)
{
do_verify = true;
init_method = 2;
time_kernel = true;
}
else if(argc == 4)
{
do_verify = static_cast<bool>(argv[1]);
init_method = atoi(argv[2]);
time_kernel = static_cast<bool>(atoi(argv[3]));
}
else
{
std::ostringstream ostr;
ostr << "Wrong parameter! " << std::endl
<< "Usage: " << argv[0] << "[verify 0/1] init_method time_kernel" << std::endl;
throw std::runtime_error(ostr.str());
};
alpha = 1.0f;
beta = 0.0f;
Tensor<InOutDataType> in_1(inLengths_1);
Tensor<InOutDataType> out_ref(outLengths);
Tensor<InOutDataType> in_2(inLengths_2); // also the output tensor of the first reduction
Tensor<InOutDataType> out(outLengths);
auto inStrides_1 = in_1.mDesc.GetStrides();
auto inStrides_2 = in_2.mDesc.GetStrides();
auto outStrides = out.mDesc.GetStrides();
size_t invariant_total_length = out.mDesc.GetElementSize();
size_t reduce_total_length = in_1.mDesc.GetElementSize() / invariant_total_length;
std::size_t num_thread = 1;
if(do_verify)
{
switch(init_method)
{
case 0: break;
case 1:
in_1.GenerateTensorValue(GeneratorTensor_1<InOutDataType>{1}, num_thread);
if(beta != 0.0f)
out_ref.GenerateTensorValue(GeneratorTensor_1<InOutDataType>{1}, num_thread);
break;
case 2:
in_1.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_1.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.GetElementSpace(); i++)
out.mData[i] = out_ref.mData[i];
};
DeviceMem in_1_dev(sizeof(InOutDataType) * in_1.mDesc.GetElementSpace());
DeviceMem in_2_dev(sizeof(InOutDataType) * in_2.mDesc.GetElementSpace());
DeviceMem out_dev(sizeof(InOutDataType) * out.mDesc.GetElementSpace());
in_1_dev.ToDevice(in_1.mData.data());
if(beta != 0.0f)
out_dev.ToDevice(out.mData.data());
if(do_verify)
{
ReductionHost<InOutDataType,
AccDataType,
InOutDataType,
ReduceOpId,
5, // Rank
2, // NumReduceDim
PropagateNan,
OutputIndex>
hostReduce(in_1.mDesc, out_ref.mDesc, invariantDims, reduceDims);
hostReduce.Run(alpha, in_1.mData.data(), beta, out_ref.mData.data(), nullptr);
};
std::vector<ck::index_t> i_inLengths_1;
std::vector<ck::index_t> i_inStrides_1;
std::vector<ck::index_t> i_inLengths_2;
std::vector<ck::index_t> i_inStrides_2;
std::vector<ck::index_t> i_outLengths;
std::vector<ck::index_t> i_outStrides;
i_inLengths_1.assign(inLengths_1.begin(), inLengths_1.end());
i_inStrides_1.assign(inStrides_1.begin(), inStrides_1.end());
i_inLengths_2.assign(inLengths_2.begin(), inLengths_2.end());
i_inStrides_2.assign(inStrides_2.begin(), inStrides_2.end());
i_outLengths.assign(outLengths.begin(), outLengths.end());
i_outStrides.assign(outStrides.begin(), outStrides.end());
auto reduce_1 = DeviceReduceInstance_1{};
auto argument_ptr_1 = reduce_1.MakeArgumentPointer(
i_inLengths_1,
i_inStrides_1,
i_inLengths_2,
i_inStrides_2,
reduceDims_1,
1.0f,
0.0f,
in_1_dev.GetDeviceBuffer(),
nullptr,
in_2_dev.GetDeviceBuffer(),
nullptr,
InElementwiseOperation{static_cast<int32_t>(reduce_total_length)},
PassThroughOp{});
if(!reduce_1.IsSupportedArgument(argument_ptr_1.get()))
{
std::cout
<< "The runtime parameters seems not supported by the DeviceReduce instance, exiting!"
<< std::endl;
};
auto invoker_ptr_1 = reduce_1.MakeInvokerPointer();
auto reduce_2 = DeviceReduceInstance_2{};
auto argument_ptr_2 = reduce_2.MakeArgumentPointer(
i_inLengths_2,
i_inStrides_2,
i_outLengths,
i_outStrides,
reduceDims_2,
alpha,
beta,
in_2_dev.GetDeviceBuffer(),
nullptr,
out_dev.GetDeviceBuffer(),
nullptr,
PassThroughOp{},
AccElementwiseOperation{static_cast<int32_t>(reduce_total_length)});
if(!reduce_2.IsSupportedArgument(argument_ptr_2.get()))
{
std::cout
<< "The runtime parameters seems not supported by the DeviceReduce instance, exiting!"
<< std::endl;
};
auto invoker_ptr_2 = reduce_2.MakeInvokerPointer();
float avg_time_1 = invoker_ptr_1->Run(argument_ptr_1.get(), StreamConfig{nullptr, time_kernel});
float avg_time_2 = invoker_ptr_2->Run(argument_ptr_2.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_1 + avg_time_2);
std::cout << "Perf: " << avg_time_1 + avg_time_2 << " ms, " << gb_per_sec << " GB/s, "
<< reduce_1.GetTypeString() << " => " << reduce_2.GetTypeString() << std::endl;
bool pass = true;
if(do_verify)
{
out_dev.FromDevice(out.mData.data());
pass = pass && ck::utils::check_err(out.mData, out_ref.mData);
};
return (pass ? 0 : 1);
}

10
example/13_pool2d_fwd/README.md
View File

@@ -4,9 +4,9 @@
```bash
#arg1: verification (0=no, 1=yes)
#arg2: initialization (0=no init, 1=single integer value, 2=scope integer value, 3=decimal value)
#arg3: run kernel # of times (>1)
#arg3: time kernel (0=no, 1=yes)
#arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, LeftPy, LeftPx, RightPy, RightPx
./bin/example_pool2d_fwd 1 1 10
./bin/example_pool2d_fwd 1 1 1
```
Result
@@ -14,9 +14,7 @@ Result
in_n_c_hi_wi: dim 4, lengths {128, 192, 71, 71}, strides {967872, 1, 13632, 192}
out_n_c_ho_wo: dim 4, lengths {128, 192, 36, 36}, strides {248832, 1, 6912, 192}
launch_and_time_kernel: grid_dim {124416, 1, 1}, block_dim {64, 1, 1}
Warm up
Warm up 1 time
Start running 10 times...
Perf: 0.415453 ms, 1.37996 TFlops, 749.726 GB/s
error: 0
max_diff: 0, 1, 1
Perf: 0.397436 ms, 1.44252 TFlops, 783.713 GB/s
```

114
example/13_pool2d_fwd/pool2d_fwd.cpp
View File

@@ -20,6 +20,8 @@ using InDataType = ck::half_t;
using OutDataType = ck::half_t;
using AccDataType = float;
using IndexDataType = int32_t;
using InLayout = ck::tensor_layout::convolution::NHWC;
using OutLayout = ck::tensor_layout::convolution::NHWC;
@@ -29,7 +31,7 @@ static constexpr auto ReduceOpId = ck::ReduceTensorOp::MAX;
static constexpr auto ReduceOpId = ck::ReduceTensorOp::AVG;
#endif
static constexpr bool NeedIndices = false;
static constexpr bool OutputIndex = false;
static constexpr bool PropagateNan = false;
using DevicePoolFwdInstance =
@@ -38,7 +40,7 @@ using DevicePoolFwdInstance =
OutDataType, // OutDataType
AccDataType, // AccDataType
ReduceOpId,
NeedIndices,
OutputIndex,
64, // BlockSize
64, // ReduceMThreadClusterSize
1, // ReduceKThreadClusterSize
@@ -51,10 +53,10 @@ template <typename InDataType,
typename AccDataType,
ck::ReduceTensorOp ReduceOpId,
bool PropagateNan,
bool NeedIndices>
bool OutputIndex>
static void pool_host_verify(const Tensor<InDataType>& in,
Tensor<OutDataType>& out,
Tensor<int>& out_indices,
Tensor<IndexDataType>& out_indices,
const std::array<ck::index_t, 2>& window_spatial_lengths,
const std::array<ck::index_t, 2>& window_strides,
const std::array<ck::index_t, 2>& in_left_pads,
@@ -62,26 +64,26 @@ static void pool_host_verify(const Tensor<InDataType>& in,
{
using namespace ck::host_reduce;
const int divider = window_spatial_lengths[0] * window_spatial_lengths[1];
const int32_t divider = window_spatial_lengths[0] * window_spatial_lengths[1];
const auto PreUnaryOp = PreUnaryOpFn<AccDataType, ReduceOpId>(divider);
const auto PosUnaryOp = PosUnaryOpFn<AccDataType, ReduceOpId>(divider);
if constexpr(!NeedIndices)
if constexpr(!OutputIndex)
{
auto opReduce = ReduceOpFn<AccDataType, ReduceOpId>();
auto f_nchw = [&](auto n, auto c, auto ho, auto wo) {
auto accuVal = ReduceOpZeroVal<AccDataType, ReduceOpId>();
for(int y = 0; y < window_spatial_lengths[0]; ++y)
for(ck::index_t y = 0; y < window_spatial_lengths[0]; ++y)
{
int hi = ho * window_strides[0] + y - in_left_pads[0];
for(int x = 0; x < window_spatial_lengths[1]; ++x)
ck::index_t hi = ho * window_strides[0] + y - in_left_pads[0];
for(ck::index_t x = 0; x < window_spatial_lengths[1]; ++x)
{
int wi = wo * window_strides[1] + x - in_left_pads[1];
if(hi >= 0 && hi < ck::type_convert<int>(in.mDesc.GetLengths()[2]) && wi >= 0 &&
wi < ck::type_convert<int>(in.mDesc.GetLengths()[3]))
ck::index_t wi = wo * window_strides[1] + x - in_left_pads[1];
if(hi >= 0 && hi < static_cast<ck::index_t>(in.mDesc.GetLengths()[2]) &&
wi >= 0 && wi < static_cast<ck::index_t>(in.mDesc.GetLengths()[3]))
{
AccDataType currVal = static_cast<AccDataType>(in(n, c, hi, wi));
@@ -108,24 +110,24 @@ static void pool_host_verify(const Tensor<InDataType>& in,
auto opReduce = ReduceOpFn2<AccDataType, ReduceOpId>();
auto f_nchw = [&](auto n, auto c, auto ho, auto wo) {
auto accuVal = ReduceOpZeroVal<AccDataType, ReduceOpId>();
int accuIndex = 0;
auto accuVal = ReduceOpZeroVal<AccDataType, ReduceOpId>();
IndexDataType accuIndex = 0;
for(int y = 0; y < window_spatial_lengths[0]; ++y)
for(ck::index_t y = 0; y < window_spatial_lengths[0]; ++y)
{
int hi = ho * window_strides[0] + y - in_left_pads[0];
for(int x = 0; x < window_spatial_lengths[1]; ++x)
ck::index_t hi = ho * window_strides[0] + y - in_left_pads[0];
for(ck::index_t x = 0; x < window_spatial_lengths[1]; ++x)
{
int wi = wo * window_strides[1] + x - in_left_pads[1];
ck::index_t wi = wo * window_strides[1] + x - in_left_pads[1];
if(hi >= 0 && hi < in.mDesc.GetLengths()[2] && wi >= 0 &&
wi < in.mDesc.GetLengths()[3])
{
AccDataType currVal = static_cast<AccDataType>(in(n, c, hi, wi));
int currIndex = y * window_spatial_lengths[1] + x;
AccDataType currVal = static_cast<AccDataType>(in(n, c, hi, wi));
IndexDataType currIndex = y * window_spatial_lengths[1] + x;
PreUnaryOp(currVal);
binop_with_nan_check2<AccDataType, PropagateNan>(
binop_with_index_and_nan_check<AccDataType, IndexDataType, PropagateNan>(
opReduce, accuVal, currVal, accuIndex, currIndex);
}
}
@@ -149,9 +151,9 @@ int main(int argc, char* argv[])
{
using namespace ck::host_reduce;
bool do_verification = true;
int init_method = 1;
bool time_kernel = false;
bool do_verification;
int init_method;
bool time_kernel;
// Pool shape
ck::index_t N = 128;
@@ -167,17 +169,23 @@ int main(int argc, char* argv[])
ck::index_t in_right_pad_h = 1;
ck::index_t in_right_pad_w = 1;
if(argc == 4)
if(argc == 1)
{
do_verification = true;
init_method = 1;
time_kernel = true;
}
else if(argc == 4)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
time_kernel = static_cast<bool>(std::stoi(argv[3]));
}
else if(argc == 16)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
time_kernel = std::stoi(argv[3]);
time_kernel = static_cast<bool>(std::stoi(argv[3]));
N = std::stoi(argv[4]);
C = std::stoi(argv[5]);
@@ -196,7 +204,7 @@ int main(int argc, char* argv[])
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: time kernel (0=n0, 1=yes)\n");
printf("arg3: time kernel (0=no, 1=yes)\n");
printf("arg4 to 15: N, C, Y, X, Hi, Wi, Sy, Sx, LeftPy, LeftPx, RightPy, "
"RightPx\n");
exit(0);
@@ -228,9 +236,11 @@ int main(int argc, char* argv[])
Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi, InLayout{}));
Tensor<OutDataType> out_n_c_ho_wo_host(f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
Tensor<int> out_indices_n_c_ho_wo_host(f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
Tensor<IndexDataType> out_indices_n_c_ho_wo_host(
f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
Tensor<OutDataType> out_n_c_ho_wo_device(f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
Tensor<int> out_indices_n_c_ho_wo_device(f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
Tensor<IndexDataType> out_indices_n_c_ho_wo_device(
f_host_tensor_descriptor(N, C, Ho, Wo, OutLayout{}));
std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
std::cout << "out_n_c_ho_wo: " << out_n_c_ho_wo_host.mDesc << std::endl;
@@ -245,25 +255,25 @@ int main(int argc, char* argv[])
DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpace());
DeviceMem out_device_buf(sizeof(OutDataType) * out_n_c_ho_wo_device.mDesc.GetElementSpace());
DeviceMem out_indices_device_buf(sizeof(int) *
DeviceMem out_indices_device_buf(sizeof(IndexDataType) *
out_indices_n_c_ho_wo_device.mDesc.GetElementSpace());
in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
auto pool = DevicePoolFwdInstance{};
auto invoker_ptr = pool.MakeInvokerPointer();
auto argument_ptr =
pool.MakeArgumentPointer(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
static_cast<int*>(out_indices_device_buf.GetDeviceBuffer()),
N,
C,
std::array<ck::index_t, 2>{{Hi, Wi}},
std::array<ck::index_t, 2>{{Y, X}},
std::array<ck::index_t, 2>{{Ho, Wo}},
window_strides,
input_left_pads,
input_right_pads);
auto pool = DevicePoolFwdInstance{};
auto invoker_ptr = pool.MakeInvokerPointer();
auto argument_ptr = pool.MakeArgumentPointer(
static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
static_cast<IndexDataType*>(out_indices_device_buf.GetDeviceBuffer()),
N,
C,
std::array<ck::index_t, 2>{{Hi, Wi}},
std::array<ck::index_t, 2>{{Y, X}},
std::array<ck::index_t, 2>{{Ho, Wo}},
window_strides,
input_left_pads,
input_right_pads);
if(!pool.IsSupportedArgument(argument_ptr.get()))
{
@@ -286,6 +296,7 @@ int main(int argc, char* argv[])
<< std::endl;
bool pass = true;
if(do_verification)
{
pool_host_verify<InDataType,
@@ -293,7 +304,7 @@ int main(int argc, char* argv[])
AccDataType,
ReduceOpId,
PropagateNan,
NeedIndices>(in_n_c_hi_wi,
OutputIndex>(in_n_c_hi_wi,
out_n_c_ho_wo_host,
out_indices_n_c_ho_wo_host,
window_spatial_lengths,
@@ -303,15 +314,16 @@ int main(int argc, char* argv[])
out_device_buf.FromDevice(out_n_c_ho_wo_device.mData.data());
pass &= ck::utils::check_err(out_n_c_ho_wo_device.mData, out_n_c_ho_wo_host.mData);
pass = pass && ck::utils::check_err(out_n_c_ho_wo_device.mData, out_n_c_ho_wo_host.mData);
if constexpr(NeedIndices)
if constexpr(OutputIndex)
{
out_indices_device_buf.FromDevice(out_indices_n_c_ho_wo_device.mData.data());
pass &= ck::utils::check_err(out_indices_n_c_ho_wo_device.mData,
out_indices_n_c_ho_wo_host.mData);
pass = pass && ck::utils::check_err(out_indices_n_c_ho_wo_device.mData,
out_indices_n_c_ho_wo_host.mData);
};
}
return pass ? 0 : 1;
return (pass ? 0 : 1);
}