ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-17 11:30:02 +00:00
Code Issues Packages Projects Releases Wiki Activity

Update to the Reduction API and instances (#476)

Browse Source 
* Simplify the macros for declaring and defining the add_device_reduce_instance_xxxx() instances

* Change the types of lengths and strides from std::vector to std::array for the reduction device interfaces

* Remove DeviceSoftmaxImpl's depending on DeviceReduceMultiblock

* Split the cpp and hpp files for reduction instances to enable more parallel compiling

* Remove the using of macros for declaring reduction instances and instance references

* Update to add_device_reduce_instance_xxxx templated functions

* Use ReduceOperation+InElementwiseOp+AccElementwiseOp to repace the ReduceOpId in defining add_reduce_instance_xxxx() templates

* Change return format

[ROCm/composable_kernel commit: dda3a0a10b]
This commit is contained in:
Qianfeng
2022-10-25 23:39:11 +08:00
committed by GitHub
parent 838ad0ef94
commit 2bc994b462
209 changed files with 4646 additions and 2279 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
example/12_reduce/reduce_blockwise.cpp
View File

@@ -140,6 +140,10 @@ bool reduce_blockwise_test(bool do_verification,
if(ShapeType::Rank_ != inLengths.size() || ShapeType::NumReduceDim_ != reduceDims.size())
return;
std::array<int, ShapeType::NumReduceDim_> arrReduceDims;
std::copy(reduceDims.begin(), reduceDims.end(), arrReduceDims.begin());
result = reduce_blockwise_impl<InOutDataType,
AccDataType,
ReduceOpId,
@@ -147,7 +151,7 @@ bool reduce_blockwise_test(bool do_verification,
ShapeType::NumReduceDim_,
PropagateNan,
OutputIndex>(
do_verification, init_method, time_kernel, inLengths, reduceDims, alpha, beta);
do_verification, init_method, time_kernel, inLengths, arrReduceDims, alpha, beta);
matched = true;
});

30
example/12_reduce/reduce_blockwise_impl.hpp
View File

@@ -30,7 +30,7 @@ int reduce_blockwise_impl(bool do_verification,
int init_method,
bool time_kernel,
const std::vector<size_t>& inLengths,
const std::vector<int>& reduceDims,
const std::array<int, NumReduceDim>& reduceDims,
float alpha,
float beta)
@@ -38,6 +38,8 @@ int reduce_blockwise_impl(bool do_verification,
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);
@@ -143,7 +145,7 @@ int reduce_blockwise_impl(bool do_verification,
std::vector<size_t> outLengths;
std::vector<int> invariantDims = get_invariant_dims<Rank, NumReduceDim>(reduceDims);
auto invariantDims = get_invariant_dims<Rank, NumReduceDim>(reduceDims);
if(invariantDims.empty())
outLengths.push_back(1);
@@ -256,22 +258,22 @@ int reduce_blockwise_impl(bool do_verification,
acc_elementwise_op);
};
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;
std::array<index_t, Rank> arrInLengths;
std::array<index_t, Rank> arrInStrides;
std::array<index_t, NumOutDim> arrOutLengths;
std::array<index_t, NumOutDim> arrOutStrides;
i_inLengths.assign(inLengths.begin(), inLengths.end());
i_inStrides.assign(inStrides.begin(), inStrides.end());
i_outLengths.assign(outLengths.begin(), outLengths.end());
i_outStrides.assign(outStrides.begin(), outStrides.end());
std::copy(inLengths.begin(), inLengths.end(), arrInLengths.begin());
std::copy(inStrides.begin(), inStrides.end(), arrInStrides.begin());
std::copy(outLengths.begin(), outLengths.end(), arrOutLengths.begin());
std::copy(outStrides.begin(), outStrides.end(), arrOutStrides.begin());
auto reduce = DeviceReduceInstance{};
auto argument_ptr = reduce.MakeArgumentPointer(i_inLengths,
i_inStrides,
i_outLengths,
i_outStrides,
auto argument_ptr = reduce.MakeArgumentPointer(arrInLengths,
arrInStrides,
arrOutLengths,
arrOutStrides,
reduceDims,
alpha,
beta,

52
example/12_reduce/reduce_blockwise_two_call.cpp
View File

@@ -90,15 +90,15 @@ 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::array<int, 1> reduceDims_1 = {4};
// const std::array<int, 4> invariantDims_1 = {0, 1, 2, 3};
const std::vector<int> reduceDims_2 = {3};
const std::vector<int> invariantDims_2 = {0, 1, 2};
const std::array<int, 1> reduceDims_2 = {3};
// const std::array<int, 3> 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::array<int, 2> reduceDims = {3, 4};
const std::array<int, 3> invariantDims = {0, 1, 2};
const std::vector<size_t> inLengths_1 = {64, 320, 80, 4, 128};
@@ -214,26 +214,26 @@ int main(int argc, char* argv[])
acc_elementwise_op);
};
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;
std::array<index_t, 5> arrInLengths_1;
std::array<index_t, 5> arrInStrides_1;
std::array<index_t, 4> arrInLengths_2;
std::array<index_t, 4> arrInStrides_2;
std::array<index_t, 3> arrOutLengths;
std::array<index_t, 3> arrOutStrides;
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());
std::copy(inLengths_1.begin(), inLengths_1.end(), arrInLengths_1.begin());
std::copy(inStrides_1.begin(), inStrides_1.end(), arrInStrides_1.begin());
std::copy(inLengths_2.begin(), inLengths_2.end(), arrInLengths_2.begin());
std::copy(inStrides_2.begin(), inStrides_2.end(), arrInStrides_2.begin());
std::copy(outLengths.begin(), outLengths.end(), arrOutLengths.begin());
std::copy(outStrides.begin(), outStrides.end(), arrOutStrides.begin());
auto reduce_1 = DeviceReduceInstance_1{};
auto argument_ptr_1 = reduce_1.MakeArgumentPointer(i_inLengths_1,
i_inStrides_1,
i_inLengths_2,
i_inStrides_2,
auto argument_ptr_1 = reduce_1.MakeArgumentPointer(arrInLengths_1,
arrInStrides_1,
arrInLengths_2,
arrInStrides_2,
reduceDims_1,
1.0f,
0.0f,
@@ -255,10 +255,10 @@ int main(int argc, char* argv[])
auto reduce_2 = DeviceReduceInstance_2{};
auto argument_ptr_2 = reduce_2.MakeArgumentPointer(i_inLengths_2,
i_inStrides_2,
i_outLengths,
i_outStrides,
auto argument_ptr_2 = reduce_2.MakeArgumentPointer(arrInLengths_2,
arrInStrides_2,
arrOutLengths,
arrOutStrides,
reduceDims_2,
alpha,
beta,

13
example/12_reduce/reduce_example_common.hpp
View File

@@ -5,11 +5,10 @@
#include "ck/ck.hpp"
template <ck::index_t Rank, ck::index_t NumReduceDim>
std::vector<int> get_invariant_dims(const std::vector<int>& reduceDims)
template <int Rank, int NumReduceDim>
static inline std::array<int, Rank - NumReduceDim>
get_invariant_dims(const std::array<int, NumReduceDim>& reduceDims)
{
assert(NumReduceDim == reduceDims.size());
int reduceFlag = 0;
// flag the bits for the reduceDims
@@ -18,13 +17,15 @@ std::vector<int> get_invariant_dims(const std::vector<int>& reduceDims)
reduceFlag |= 1 << reduceDims[i];
};
std::vector<int> invariantDims;
std::array<int, Rank - NumReduceDim> invariantDims;
// collect invariant dimensions
int dim = 0;
for(int i = 0; i < Rank; i++)
if((reduceFlag & (1 << i)) == 0)
{
invariantDims.push_back(i);
invariantDims[dim] = i;
dim++;
};
return invariantDims;

6
example/12_reduce/reduce_multiblock_atomic_add.cpp
View File

@@ -138,13 +138,17 @@ bool reduce_multiblock_atomic_add_test(bool do_verification,
if(ShapeType::Rank_ != inLengths.size() || ShapeType::NumReduceDim_ != reduceDims.size())
return;
std::array<int, ShapeType::NumReduceDim_> a_reduceDims;
std::copy(reduceDims.begin(), reduceDims.end(), a_reduceDims.begin());
result = reduce_multiblock_atomic_add_impl<InOutDataType,
AccDataType,
ReduceOpId,
ShapeType::Rank_,
ShapeType::NumReduceDim_,
PropagateNan>(
do_verification, init_method, time_kernel, inLengths, reduceDims, alpha, beta);
do_verification, init_method, time_kernel, inLengths, a_reduceDims, alpha, beta);
matched = true;
});

30
example/12_reduce/reduce_multiblock_atomic_add_impl.hpp
View File

@@ -29,7 +29,7 @@ int reduce_multiblock_atomic_add_impl(bool do_verification,
int init_method,
bool time_kernel,
const std::vector<size_t>& inLengths,
const std::vector<int>& reduceDims,
const std::array<int, NumReduceDim>& reduceDims,
float alpha,
float beta)
@@ -37,6 +37,8 @@ int reduce_multiblock_atomic_add_impl(bool do_verification,
using namespace ck;
using namespace ck::tensor_operation::device;
constexpr index_t NumOutDim = (Rank - NumReduceDim == 0) ? 1 : Rank - NumReduceDim;
constexpr bool op_support_atomic_add =
(ReduceOpId == ReduceTensorOp::ADD || ReduceOpId == ReduceTensorOp::AVG);
@@ -84,7 +86,7 @@ int reduce_multiblock_atomic_add_impl(bool do_verification,
std::vector<size_t> outLengths;
std::vector<int> invariantDims = get_invariant_dims<Rank, NumReduceDim>(reduceDims);
auto invariantDims = get_invariant_dims<Rank, NumReduceDim>(reduceDims);
if(invariantDims.empty())
outLengths.push_back(1);
@@ -169,22 +171,22 @@ int reduce_multiblock_atomic_add_impl(bool do_verification,
acc_elementwise_op);
};
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;
std::array<index_t, Rank> arrInLengths;
std::array<index_t, Rank> arrInStrides;
std::array<index_t, NumOutDim> arrOutLengths;
std::array<index_t, NumOutDim> arrOutStrides;
i_inLengths.assign(inLengths.begin(), inLengths.end());
i_inStrides.assign(inStrides.begin(), inStrides.end());
i_outLengths.assign(outLengths.begin(), outLengths.end());
i_outStrides.assign(outStrides.begin(), outStrides.end());
std::copy(inLengths.begin(), inLengths.end(), arrInLengths.begin());
std::copy(inStrides.begin(), inStrides.end(), arrInStrides.begin());
std::copy(outLengths.begin(), outLengths.end(), arrOutLengths.begin());
std::copy(outStrides.begin(), outStrides.end(), arrOutStrides.begin());
auto reduce = DeviceReduceInstance{};
auto argument_ptr = reduce.MakeArgumentPointer(i_inLengths,
i_inStrides,
i_outLengths,
i_outStrides,
auto argument_ptr = reduce.MakeArgumentPointer(arrInLengths,
arrInStrides,
arrOutLengths,
arrOutStrides,
reduceDims,
alpha,
beta,