Skip to content

GitLab

Unverified Commit dda3a0a1 authored by Qianfeng's avatar Qianfeng Committed by GitHub
Browse files

Update to the Reduction API and instances (#476) 

* 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
parent 6ea9257e
Show whitespace changes
Inline Side-by-side
Showing with 576 additions and 346 deletions
example/12_reduce/reduce_blockwise.cpp
View file @ dda3a0a1
......@@ -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;
});
......
example/12_reduce/reduce_blockwise_impl.hpp
View file @ dda3a0a1
......@@ -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,
......
example/12_reduce/reduce_blockwise_two_call.cpp
View file @ dda3a0a1
......@@ -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,
......
example/12_reduce/reduce_example_common.hpp
View file @ dda3a0a1
......@@ -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;
......
example/12_reduce/reduce_multiblock_atomic_add.cpp
View file @ dda3a0a1
......@@ -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;
});
......
example/12_reduce/reduce_multiblock_atomic_add_impl.hpp
View file @ dda3a0a1
......@@ -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,
......
include/ck/tensor_operation/gpu/device/device_reduce.hpp
View file @ dda3a0a1
......@@ -3,27 +3,30 @@
#pragma once
#include <vector>
#include <array>
#include <memory>
#include <iostream>
#include "ck/utility/common_header.hpp"
#include "ck/utility/reduction_enums.hpp"
#include "ck/ck.hpp"
#include "ck/tensor_operation/gpu/device/device_base.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
template <typename InElementwiseOperation, typename AccElementwiseOperation>
template <index_t Rank,
index_t NumReduceDim,
typename InElementwiseOperation,
typename AccElementwiseOperation>
struct DeviceReduce : public BaseOperator
{
static constexpr index_t NumOutDim = (Rank - NumReduceDim == 0) ? 1 : Rank - NumReduceDim;
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(const std::vector<index_t> inLengths,
const std::vector<index_t> inStrides,
const std::vector<index_t> outLengths,
const std::vector<index_t> outStrides,
const std::vector<int> reduceDims,
MakeArgumentPointer(const std::array<index_t, Rank> inLengths,
const std::array<index_t, Rank> inStrides,
const std::array<index_t, NumOutDim> outLengths,
const std::array<index_t, NumOutDim> outStrides,
const std::array<int, NumReduceDim> reduceDims,
float alpha,
float beta,
const void* in_dev,
......@@ -36,9 +39,12 @@ struct DeviceReduce : public BaseOperator
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
template <typename InElementwiseOperation, typename AccElementwiseOperation>
using DeviceReducePtr =
std::unique_ptr<DeviceReduce<InElementwiseOperation, AccElementwiseOperation>>;
template <index_t Rank,
index_t NumReduceDim,
typename InElementwiseOperation,
typename AccElementwiseOperation>
using DeviceReducePtr = std::unique_ptr<
DeviceReduce<Rank, NumReduceDim, InElementwiseOperation, AccElementwiseOperation>>;
} // namespace device
} // namespace tensor_operation
......
include/ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp
View file @ dda3a0a1
......@@ -5,9 +5,8 @@
#include <iostream>
#include <sstream>
#include <array>
#include "ck/utility/common_header.hpp"
#include "ck/utility/reduction_operator.hpp"
#include "ck/tensor_description/tensor_descriptor.hpp"
#include "ck/tensor_description/tensor_descriptor_helper.hpp"
#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
......@@ -41,7 +40,8 @@ template <typename InDataType,
index_t InSrcVectorDim,
index_t InSrcVectorSize,
index_t OutDstVectorSize>
struct DeviceReduceMultiBlock : public DeviceReduce<InElementwiseOperation, AccElementwiseOperation>
struct DeviceReduceMultiBlock
: public DeviceReduce<Rank, NumReduceDim, InElementwiseOperation, AccElementwiseOperation>
{
static_assert(Rank <= 6, "Bigger Rank size is not supported!");
static_assert(BlockSize == MThreadClusterSize * KThreadClusterSize,
......@@ -58,8 +58,8 @@ struct DeviceReduceMultiBlock : public DeviceReduce<InElementwiseOperation, AccE
static constexpr index_t NumInvariantDim = Rank - NumReduceDim;
static constexpr index_t numSrcDim = Rank;
static constexpr index_t numDstDim = (NumInvariantDim == 0) ? 1 : NumInvariantDim;
static constexpr index_t NumSrcDim = Rank;
static constexpr index_t NumDstDim = (NumInvariantDim == 0) ? 1 : NumInvariantDim;
static constexpr bool reduceAllDim = (NumInvariantDim == 0);
// So far, only AtomicAdd is considered, other Atomic Operation like AtomicMax can be added
......@@ -81,13 +81,15 @@ struct DeviceReduceMultiBlock : public DeviceReduce<InElementwiseOperation, AccE
static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
static auto MakeSrc2dDescriptor(const std::vector<index_t>& inLengths,
const std::vector<index_t>& inStrides,
static auto MakeSrc2dDescriptor(const std::array<index_t, Rank>& inLengths,
const std::array<index_t, Rank>& inStrides,
int blkGroupSize,
int numBlockTileIteration)
{
const auto tupleSrcLengths = make_tuple_from_array(inLengths, Number<numSrcDim>{});
const auto tupleSrcStrides = make_tuple_from_array(inStrides, Number<numSrcDim>{});
const auto tupleSrcLengths =
generate_tuple([&](auto I) { return inLengths[I]; }, Number<Rank>{});
const auto tupleSrcStrides =
generate_tuple([&](auto I) { return inStrides[I]; }, Number<Rank>{});
const auto inDesc = make_naive_tensor_descriptor(tupleSrcLengths, tupleSrcStrides);
......@@ -97,7 +99,7 @@ struct DeviceReduceMultiBlock : public DeviceReduce<InElementwiseOperation, AccE
const auto one_dim_inDesc = transform_tensor_descriptor(
inDesc,
make_tuple(make_merge_transform(tupleSrcLengths)),
make_tuple(typename arithmetic_sequence_gen<0, numSrcDim, 1>::type{}),
make_tuple(typename arithmetic_sequence_gen<0, NumSrcDim, 1>::type{}),
make_tuple(Sequence<0>{}));
return transform_tensor_descriptor(one_dim_inDesc,
......@@ -111,10 +113,10 @@ struct DeviceReduceMultiBlock : public DeviceReduce<InElementwiseOperation, AccE
using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
using ReduceDims = typename arithmetic_sequence_gen<NumInvariantDim, Rank, 1>::type;
const auto reduceDimLengths =
make_tuple_from_array_and_index_seq(inLengths, ReduceDims{});
const auto reduceDimLengths = generate_tuple(
[&](auto I) { return inLengths[NumInvariantDim + I]; }, Number<NumReduceDim>{});
const auto invariantDimLengths =
make_tuple_from_array_and_index_seq(inLengths, InvariantDims{});
generate_tuple([&](auto I) { return inLengths[I]; }, Number<NumInvariantDim>{});
return transform_tensor_descriptor(
inDesc,
......@@ -143,18 +145,20 @@ struct DeviceReduceMultiBlock : public DeviceReduce<InElementwiseOperation, AccE
return (in_grid_desc_m_k_padded);
};
static auto MakeDst1dDescriptor(const std::vector<index_t>& outLengths,
const std::vector<index_t>& outStrides)
static auto MakeDst1dDescriptor(const std::array<index_t, NumDstDim>& outLengths,
const std::array<index_t, NumDstDim>& outStrides)
{
const auto tupleDstLengths = make_tuple_from_array(outLengths, Number<numDstDim>{});
const auto tupleDstStrides = make_tuple_from_array(outStrides, Number<numDstDim>{});
const auto tupleDstLengths =
generate_tuple([&](auto I) { return outLengths[I]; }, Number<NumDstDim>{});
const auto tupleDstStrides =
generate_tuple([&](auto I) { return outStrides[I]; }, Number<NumDstDim>{});
auto outDesc = make_naive_tensor_descriptor(tupleDstLengths, tupleDstStrides);
auto out_grid_desc_m = transform_tensor_descriptor(
outDesc,
make_tuple(make_merge_transform(tupleDstLengths)),
make_tuple(typename arithmetic_sequence_gen<0, numDstDim, 1>::type{}),
make_tuple(typename arithmetic_sequence_gen<0, NumDstDim, 1>::type{}),
make_tuple(Sequence<0>{}));
const auto invariantLength = out_grid_desc_m.GetLength(Number<0>{});
......@@ -170,18 +174,20 @@ struct DeviceReduceMultiBlock : public DeviceReduce<InElementwiseOperation, AccE
return (out_grid_desc_m_padded);
};
static auto MakeDst1dDescriptorForBufferSet(const std::vector<index_t>& outLengths,
const std::vector<index_t>& outStrides)
static auto MakeDst1dDescriptorForBufferSet(const std::array<index_t, NumDstDim>& outLengths,
const std::array<index_t, NumDstDim>& outStrides)
{
const auto tupleDstLengths = make_tuple_from_array(outLengths, Number<numDstDim>{});
const auto tupleDstStrides = make_tuple_from_array(outStrides, Number<numDstDim>{});
const auto tupleDstLengths =
generate_tuple([&](auto I) { return outLengths[I]; }, Number<NumDstDim>{});
const auto tupleDstStrides =
generate_tuple([&](auto I) { return outStrides[I]; }, Number<NumDstDim>{});
auto outDesc = make_naive_tensor_descriptor(tupleDstLengths, tupleDstStrides);
auto out_grid_desc_m = transform_tensor_descriptor(
outDesc,
make_tuple(make_merge_transform(tupleDstLengths)),
make_tuple(typename arithmetic_sequence_gen<0, numDstDim, 1>::type{}),
make_tuple(typename arithmetic_sequence_gen<0, NumDstDim, 1>::type{}),
make_tuple(Sequence<0>{}));
const auto length = out_grid_desc_m.GetLength(Number<0>{});
......@@ -198,11 +204,11 @@ struct DeviceReduceMultiBlock : public DeviceReduce<InElementwiseOperation, AccE
struct Argument : public BaseArgument
{
Argument(const std::vector<index_t> inLengths,
const std::vector<index_t> inStrides,
const std::vector<index_t> outLengths,
const std::vector<index_t> outStrides,
const std::vector<int> reduceDims,
Argument(const std::array<index_t, Rank> inLengths,
const std::array<index_t, Rank> inStrides,
const std::array<index_t, NumDstDim> outLengths,
const std::array<index_t, NumDstDim> outStrides,
const std::array<int, NumReduceDim> reduceDims,
float alpha,
float beta,
const InDataType* in_dev,
......@@ -272,10 +278,10 @@ struct DeviceReduceMultiBlock : public DeviceReduce<InElementwiseOperation, AccE
math::integer_least_multiple(invariant_total_length, BlockSize) / BlockSize;
}
std::vector<index_t> inLengths_;
std::vector<index_t> inStrides_;
std::vector<index_t> outLengths_;
std::vector<index_t> outStrides_;
std::array<index_t, Rank> inLengths_;
std::array<index_t, Rank> inStrides_;
std::array<index_t, NumDstDim> outLengths_;
std::array<index_t, NumDstDim> outStrides_;
AccDataType alpha_;
AccDataType beta_;
......@@ -459,11 +465,11 @@ struct DeviceReduceMultiBlock : public DeviceReduce<InElementwiseOperation, AccE
};
std::unique_ptr<BaseArgument>
MakeArgumentPointer(const std::vector<index_t> inLengths,
const std::vector<index_t> inStrides,
const std::vector<index_t> outLengths,
const std::vector<index_t> outStrides,
const std::vector<int> reduceDims,
MakeArgumentPointer(const std::array<index_t, Rank> inLengths,
const std::array<index_t, Rank> inStrides,
const std::array<index_t, NumDstDim> outLengths,
const std::array<index_t, NumDstDim> outStrides,
const std::array<int, NumReduceDim> reduceDims,
float alpha,
float beta,
const void* in_dev,
......
include/ck/tensor_operation/gpu/device/impl/device_reduce_threadwise.hpp
View file @ dda3a0a1
......@@ -5,6 +5,7 @@
#include <iostream>
#include <sstream>
#include <array>
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
......@@ -34,7 +35,8 @@ template <typename InDataType,
index_t InSrcVectorDim,
index_t InSrcVectorSize,
index_t OutDstVectorSize>
struct DeviceReduceThreadWise : public DeviceReduce<InElementwiseOperation, AccElementwiseOperation>
struct DeviceReduceThreadWise
: public DeviceReduce<Rank, NumReduceDim, InElementwiseOperation, AccElementwiseOperation>
{
static_assert(Rank <= 6, "Bigger Rank size is not supported!");
......@@ -49,18 +51,20 @@ struct DeviceReduceThreadWise : public DeviceReduce<InElementwiseOperation, AccE
static constexpr index_t NumInvariantDim = Rank - NumReduceDim;
static constexpr index_t numSrcDim = Rank;
static constexpr index_t numDstDim = (NumInvariantDim == 0) ? 1 : NumInvariantDim;
static constexpr index_t NumSrcDim = Rank;
static constexpr index_t NumDstDim = (NumInvariantDim == 0) ? 1 : NumInvariantDim;
static constexpr bool reduceAllDim = (NumInvariantDim == 0);
static constexpr index_t M_BlockTileSize = BlockSize * MThreadSliceSize;
static constexpr index_t K_BlockTileSize = 1 * KThreadSliceSize;
static auto MakeSrc2dDescriptor(const std::vector<index_t>& inLengths,
const std::vector<index_t>& inStrides)
static auto MakeSrc2dDescriptor(const std::array<index_t, Rank>& inLengths,
const std::array<index_t, Rank>& inStrides)
{
const auto tupleSrcLengths = make_tuple_from_array(inLengths, Number<numSrcDim>{});
const auto tupleSrcStrides = make_tuple_from_array(inStrides, Number<numSrcDim>{});
const auto tupleSrcLengths =
generate_tuple([&](auto I) { return inLengths[I]; }, Number<Rank>{});
const auto tupleSrcStrides =
generate_tuple([&](auto I) { return inStrides[I]; }, Number<Rank>{});
const auto inDesc = make_naive_tensor_descriptor(tupleSrcLengths, tupleSrcStrides);
......@@ -70,7 +74,7 @@ struct DeviceReduceThreadWise : public DeviceReduce<InElementwiseOperation, AccE
const auto one_dim_inDesc = transform_tensor_descriptor(
inDesc,
make_tuple(make_merge_transform(tupleSrcLengths)),
make_tuple(typename arithmetic_sequence_gen<0, numSrcDim, 1>::type{}),
make_tuple(typename arithmetic_sequence_gen<0, NumSrcDim, 1>::type{}),
make_tuple(Sequence<0>{}));
return transform_tensor_descriptor(one_dim_inDesc,
......@@ -84,10 +88,10 @@ struct DeviceReduceThreadWise : public DeviceReduce<InElementwiseOperation, AccE
using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
using ReduceDims = typename arithmetic_sequence_gen<NumInvariantDim, Rank, 1>::type;
const auto reduceDimLengths =
make_tuple_from_array_and_index_seq(inLengths, ReduceDims{});
const auto reduceDimLengths = generate_tuple(
[&](auto I) { return inLengths[NumInvariantDim + I]; }, Number<NumReduceDim>{});
const auto invariantDimLengths =
make_tuple_from_array_and_index_seq(inLengths, InvariantDims{});
generate_tuple([&](auto I) { return inLengths[I]; }, Number<NumInvariantDim>{});
return transform_tensor_descriptor(
inDesc,
......@@ -116,18 +120,20 @@ struct DeviceReduceThreadWise : public DeviceReduce<InElementwiseOperation, AccE
return (in_grid_desc_m_k_padded);
};
static auto MakeDst1dDescriptor(const std::vector<index_t>& outLengths,
const std::vector<index_t>& outStrides)
static auto MakeDst1dDescriptor(const std::array<index_t, NumDstDim>& outLengths,
const std::array<index_t, NumDstDim>& outStrides)
{
const auto tupleDstLengths = make_tuple_from_array(outLengths, Number<numDstDim>{});
const auto tupleDstStrides = make_tuple_from_array(outStrides, Number<numDstDim>{});
const auto tupleDstLengths =
generate_tuple([&](auto I) { return outLengths[I]; }, Number<NumDstDim>{});
const auto tupleDstStrides =
generate_tuple([&](auto I) { return outStrides[I]; }, Number<NumDstDim>{});
auto outDesc = make_naive_tensor_descriptor(tupleDstLengths, tupleDstStrides);
auto out_grid_desc_m = transform_tensor_descriptor(
outDesc,
make_tuple(make_merge_transform(tupleDstLengths)),
make_tuple(typename arithmetic_sequence_gen<0, numDstDim, 1>::type{}),
make_tuple(typename arithmetic_sequence_gen<0, NumDstDim, 1>::type{}),
make_tuple(Sequence<0>{}));
const auto invariantLength = out_grid_desc_m.GetLength(Number<0>{});
......@@ -145,11 +151,11 @@ struct DeviceReduceThreadWise : public DeviceReduce<InElementwiseOperation, AccE
struct Argument : public BaseArgument
{
Argument(const std::vector<index_t> inLengths,
const std::vector<index_t> inStrides,
const std::vector<index_t> outLengths,
const std::vector<index_t> outStrides,
const std::vector<int> reduceDims,
Argument(const std::array<index_t, Rank> inLengths,
const std::array<index_t, Rank> inStrides,
const std::array<index_t, NumDstDim> outLengths,
const std::array<index_t, NumDstDim> outStrides,
const std::array<int, NumReduceDim> reduceDims,
float alpha,
float beta,
const InDataType* in_dev,
......@@ -187,10 +193,10 @@ struct DeviceReduceThreadWise : public DeviceReduce<InElementwiseOperation, AccE
M_BlockTileSize;
}
std::vector<index_t> inLengths_;
std::vector<index_t> inStrides_;
std::vector<index_t> outLengths_;
std::vector<index_t> outStrides_;
std::array<index_t, Rank> inLengths_;
std::array<index_t, Rank> inStrides_;
std::array<index_t, NumDstDim> outLengths_;
std::array<index_t, NumDstDim> outStrides_;
AccDataType alpha_;
AccDataType beta_;
......@@ -321,11 +327,11 @@ struct DeviceReduceThreadWise : public DeviceReduce<InElementwiseOperation, AccE
};
std::unique_ptr<BaseArgument>
MakeArgumentPointer(const std::vector<index_t> inLengths,
const std::vector<index_t> inStrides,
const std::vector<index_t> outLengths,
const std::vector<index_t> outStrides,
const std::vector<int> reduceDims,
MakeArgumentPointer(const std::array<index_t, Rank> inLengths,
const std::array<index_t, Rank> inStrides,
const std::array<index_t, NumDstDim> outLengths,
const std::array<index_t, NumDstDim> outStrides,
const std::array<int, NumReduceDim> reduceDims,
float alpha,
float beta,
const void* in_dev,
......
include/ck/tensor_operation/gpu/device/impl/device_softmax_impl.hpp
View file @ dda3a0a1
......@@ -8,12 +8,9 @@
#include "ck/utility/reduction_operator.hpp"
#include "ck/tensor_operation/gpu/device/device_base.hpp"
#include "ck/tensor_operation/gpu/device/device_reduce.hpp"
#include "ck/tensor_operation/gpu/device/device_softmax.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_reduce_common.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_softmax.hpp"
#include "ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp"
#include "ck/host_utility/device_prop.hpp"
#include "ck/host_utility/kernel_launch.hpp"
......@@ -50,29 +47,80 @@ struct DeviceSoftmaxImpl : public DeviceSoftmax<InDataType,
virtual index_t GetNumReduceDim() const override { return kNumReduceDim; }
// Used for freeloading of some handy functions from DeviceReduceMultiBlock
using Reduction = DeviceReduceMultiBlock<InDataType,
AccDataType,
OutDataType,
Rank,
NumReduceDim,
reduce::Add,
InElementwiseOp,
AccElementwiseOp,
InMemoryDataOperationEnum::Set,
false, // PropagateNan
false, // OutputIndex
false, // HaveIndexInputIfOutputIndex
BlockSize,
MThreadClusterSize,
KThreadClusterSize,
MThreadSliceSize,
KThreadSliceSize,
InSrcVectorDim,
InSrcVectorSize,
1>; // OutDstVectorSize
static constexpr index_t NumInvariantDim = Rank - NumReduceDim;
static constexpr index_t NumSrcDim = Rank;
static constexpr index_t NumDstDim = (NumInvariantDim == 0) ? 1 : NumInvariantDim;
static constexpr bool reduceAllDim = (NumInvariantDim == 0);
static constexpr index_t M_BlockTileSize = MThreadClusterSize * MThreadSliceSize;
static constexpr index_t K_BlockTileSize = KThreadClusterSize * KThreadSliceSize;
static auto MakeSrc2dDescriptor(const std::vector<index_t>& inLengths,
const std::vector<index_t>& inStrides,
int blkGroupSize,
int numBlockTileIteration)
{
const auto tupleSrcLengths =
generate_tuple([&](auto I) { return inLengths[I]; }, Number<Rank>{});
const auto tupleSrcStrides =
generate_tuple([&](auto I) { return inStrides[I]; }, Number<Rank>{});
const auto inDesc = make_naive_tensor_descriptor(tupleSrcLengths, tupleSrcStrides);
const auto in_grid_desc_m_k = [&]() {
if constexpr(reduceAllDim)
{
const auto one_dim_inDesc = transform_tensor_descriptor(
inDesc,
make_tuple(make_merge_transform(tupleSrcLengths)),
make_tuple(typename arithmetic_sequence_gen<0, NumSrcDim, 1>::type{}),
make_tuple(Sequence<0>{}));
return transform_tensor_descriptor(one_dim_inDesc,
make_tuple(make_unmerge_transform(make_tuple(
1, one_dim_inDesc.GetLength(Number<0>{})))),
make_tuple(Sequence<0>{}),
make_tuple(Sequence<0, 1>{}));
}
else
{
using InvariantDims = typename arithmetic_sequence_gen<0, NumInvariantDim, 1>::type;
using ReduceDims = typename arithmetic_sequence_gen<NumInvariantDim, Rank, 1>::type;
using GridDesc_M_K = decltype(Reduction::MakeSrc2dDescriptor({1}, {1}, 1, 1));
const auto reduceDimLengths = generate_tuple(
[&](auto I) { return inLengths[NumInvariantDim + I]; }, Number<NumReduceDim>{});
const auto invariantDimLengths =
generate_tuple([&](auto I) { return inLengths[I]; }, Number<NumInvariantDim>{});
return transform_tensor_descriptor(
inDesc,
make_tuple(make_merge_transform(invariantDimLengths),
make_merge_transform(reduceDimLengths)),
make_tuple(InvariantDims{}, ReduceDims{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
}
}();
const auto invariantLength = in_grid_desc_m_k.GetLength(Number<0>{});
const auto reduceLength = in_grid_desc_m_k.GetLength(Number<1>{});
const int reduceSizePerBlock = K_BlockTileSize * numBlockTileIteration;
const auto inPad_M =
math::integer_least_multiple(invariantLength, M_BlockTileSize) - invariantLength;
const auto inPad_K = reduceSizePerBlock * blkGroupSize - reduceLength;
auto in_grid_desc_m_k_padded = transform_tensor_descriptor(
in_grid_desc_m_k,
make_tuple(make_right_pad_transform(invariantLength, inPad_M),
make_right_pad_transform(reduceLength, inPad_K)),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}));
return (in_grid_desc_m_k_padded);
};
using GridDesc_M_K = decltype(MakeSrc2dDescriptor({1}, {1}, 1, 1));
using GridwiseSoftmaxGeneric = GridwiseSoftmax_mk_to_mk<InDataType,
OutDataType,
......@@ -102,7 +150,7 @@ struct DeviceSoftmaxImpl : public DeviceSoftmax<InDataType,
OutDstVectorSize,
true>;
struct Argument : public Reduction::Argument
struct Argument : public BaseArgument
{
Argument(const std::vector<index_t> inLengths,
const std::vector<index_t> inStrides,
......@@ -113,42 +161,60 @@ struct DeviceSoftmaxImpl : public DeviceSoftmax<InDataType,
OutDataType* out_dev,
InElementwiseOp in_elementwise_op,
AccElementwiseOp acc_elementwise_op)
: Reduction::Argument(inLengths,
inStrides,
{},
{},
reduceDims,
0.0f, // alpha
0.0f, // beta
in_dev,
nullptr,
out_dev,
nullptr,
in_elementwise_op,
acc_elementwise_op),
// FIXME: The base class DeviceReduceMultiBlock::Argument only supports alpha/beta of
// float32 precision. Make it support any data type so the fields can be removed.
alpha_(alpha),
beta_(beta)
: alpha_{alpha},
beta_{beta},
in_dev_{in_dev},
out_dev_{out_dev},
in_elementwise_op_{in_elementwise_op},
acc_elementwise_op_{acc_elementwise_op}
{
// std::cout << "blkGroupSize= " << this->blkGroupSize
// << ", numBlockTileIteration= " << this->numBlockTileIteration
// << ", gridSize=" << this->gridSize
// << ", invariant_total_length=" << this->invariant_total_length <<
// std::endl;
inLengths_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(inLengths, reduceDims);
inStrides_ = shuffle_tensor_dimensions<Rank, NumReduceDim>(inStrides, reduceDims);
long_index_t invariant_total_length;
long_index_t reduce_total_length;
std::tie(invariant_total_length, reduce_total_length) =
get_2d_lengths<Rank, NumReduceDim>(inLengths_);
if constexpr(NumInvariantDim == 0)
invariant_lowest_length_ = 1;
else
invariant_lowest_length_ = inLengths_[NumInvariantDim - 1];
blkGroupSize = 1;
numBlockTileIteration = (reduce_total_length + K_BlockTileSize - 1) / K_BlockTileSize;
gridSize = math::integer_least_multiple(invariant_total_length, M_BlockTileSize) /
M_BlockTileSize * blkGroupSize;
}
std::vector<index_t> inLengths_;
std::vector<index_t> inStrides_;
AccDataType alpha_;
AccDataType beta_;
const InDataType* in_dev_;
OutDataType* out_dev_;
InElementwiseOp in_elementwise_op_;
AccElementwiseOp acc_elementwise_op_;
index_t invariant_lowest_length_;
int blkGroupSize;
int numBlockTileIteration;
size_t gridSize;
};
struct Invoker : public BaseInvoker
{
float Run(const Argument& arg, const StreamConfig& stream_config = StreamConfig{})
{
const auto in_grid_desc_m_k = Reduction::MakeSrc2dDescriptor(
const auto in_grid_desc_m_k = DeviceSoftmaxImpl::MakeSrc2dDescriptor(
arg.inLengths_, arg.inStrides_, arg.blkGroupSize, arg.numBlockTileIteration);
const auto out_grid_desc_m_k = Reduction::MakeSrc2dDescriptor(
const auto out_grid_desc_m_k = DeviceSoftmaxImpl::MakeSrc2dDescriptor(
arg.inLengths_, arg.inStrides_, arg.blkGroupSize, arg.numBlockTileIteration);
bool sweep_once =
......@@ -195,15 +261,32 @@ struct DeviceSoftmaxImpl : public DeviceSoftmax<InDataType,
{
const Argument* p_arg_ = dynamic_cast<const Argument*>(p_arg);
if(!Reduction::IsSupportedArgument(p_arg_))
if constexpr(InSrcVectorDim == 0)
{
if constexpr(NumInvariantDim == 0)
{
return false;
}
if(p_arg_->inLengths_[Rank - 1] % OutDstVectorSize != 0)
else
{
if(p_arg_->inStrides_[NumInvariantDim - 1] != 1)
return false;
if(p_arg_->invariant_lowest_length_ % InSrcVectorSize != 0)
return false;
};
}
else
{
if(p_arg_->inStrides_[Rank - 1] != 1)
return false;
if(p_arg_->inLengths_[Rank - 1] % InSrcVectorSize != 0)
return false;
};
if(p_arg_->invariant_lowest_length_ % OutDstVectorSize != 0)
return false;
return true;
};
......
library/include/ck/library/tensor_operation_instance/device_operation_instance_factory.hpp
View file @ dda3a0a1
......@@ -3,7 +3,10 @@
#pragma once
#include <cstdlib>
#include "ck/utility/data_type.hpp"
#include "ck/utility/tuple.hpp"
#include "ck/tensor_operation/gpu/device/tensor_layout.hpp"
#include "ck/tensor_operation/gpu/element/element_wise_operation.hpp"
namespace ck {
namespace tensor_operation {
......@@ -15,6 +18,8 @@ using F64 = double;
using F32 = float;
using F16 = ck::half_t;
using BF16 = ck::bhalf_t;
using I8 = int8_t;
using I32 = int32_t;
using Empty_Tuple = ck::Tuple<>;
......
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance.hpp
View file @ dda3a0a1
......@@ -3,24 +3,77 @@
#pragma once
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_min.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_max.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f16_f16_amax.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f16_f32_f16_norm2.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_norm2.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_min.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_max.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f32_f32_amax.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f32_f64_f32_norm2.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_norm2.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_min.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_max.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_f64_f64_f64_amax.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_min.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_max.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i8_i8_amax.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_i8_i32_i8_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f16_f32_f32_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f32_f32_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f32_f64_f32_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_f64_f64_f64_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_multiblock_atomic_add_b16_f32_f32_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_min.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_max.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f16_f16_amax.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f16_f32_f16_norm2.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_norm2.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_min.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_max.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f32_f32_amax.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f32_f64_f32_norm2.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_norm2.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_min.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_max.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_f64_f64_f64_amax.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_min.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_max.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i8_i8_amax.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_i8_i32_i8_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_add.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_avg.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_norm2.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_min.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_max.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_threadwise_b16_f32_b16_amax.hpp"
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp
View file @ dda3a0a1
......@@ -5,6 +5,8 @@
#include "ck/tensor_operation/gpu/device/reduction_operator_mapping.hpp"
#include "ck/tensor_operation/gpu/device/impl/device_reduce_multiblock.hpp"
#include "ck/library/tensor_operation_instance/device_operation_instance_factory.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_impl_common.hpp"
namespace ck {
......@@ -63,33 +65,20 @@ using reduce_configuration_2_instances_blockwise = std::tuple<
>;
#endif
template <ReduceTensorOp ReduceOpId>
using deviceReduceBlockWisePtrType = DeviceReducePtr<
typename reduce_unary_operator<ReduceOpId, true, true>::InElementwiseOperation,
typename reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation>;
template <typename InDataType,
typename AccDataType,
typename OutDataType,
int Rank,
int NumReduceDim,
ReduceTensorOp ReduceOpId,
typename ReduceOperation,
typename InElementwiseOp,
typename AccElementwiseOp,
bool PropagateNan,
bool UseIndex>
bool OutputIndex>
void add_device_reduce_instance_blockwise(
std::vector<deviceReduceBlockWisePtrType<ReduceOpId>>& device_op_instances)
std::vector<DeviceReducePtr<Rank, NumReduceDim, InElementwiseOp, AccElementwiseOp>>&
device_op_instances)
{
using ReduceOperation = typename reduce_binary_operator<ReduceOpId>::opType;
using InElementwiseOperation =
typename reduce_unary_operator<ReduceOpId, true, true>::InElementwiseOperation;
using AccElementwiseOperation =
typename reduce_unary_operator<ReduceOpId, true, true>::AccElementwiseOperation;
constexpr bool Indexable =
(ReduceOpId == ReduceTensorOp::MIN || ReduceOpId == ReduceTensorOp::MAX ||
ReduceOpId == ReduceTensorOp::AMAX);
constexpr bool OutputIndex = Indexable && UseIndex;
static_for<0, std::tuple_size<reduce_configuration_1_instances_blockwise>::value, 1>{}(
[&](auto i) {
using cfg1 = remove_cvref_t<decltype(
......@@ -107,8 +96,8 @@ void add_device_reduce_instance_blockwise(
Rank,
NumReduceDim,
ReduceOperation,
InElementwiseOperation,
AccElementwiseOperation,
InElementwiseOp,
AccElementwiseOp,
InMemoryDataOperationEnum::Set,
PropagateNan,
OutputIndex,
......@@ -128,52 +117,6 @@ void add_device_reduce_instance_blockwise(
});
};
#define ADD_BLOCKWISE_INST_BY_TYPE( \
inT, compT, outT, ReduceOpId, PropagateNan, UseIndex, Rank, NumReduceDim) \
template void add_device_reduce_instance_blockwise<inT, \
compT, \
outT, \
Rank, \
NumReduceDim, \
ReduceOpId, \
PropagateNan, \
UseIndex>( \
std::vector<deviceReduceBlockWisePtrType<ReduceOpId>> & device_op_instances)
#define ADD_BLOCKWISE_INST_BY_ID( \
inT, compT, outT, ReduceOpId, NanOpt, IndicesOpt, Rank, NumReduceDim) \
ADD_BLOCKWISE_INST_BY_TYPE(inT, \
compT, \
outT, \
static_cast<ReduceTensorOp>(ReduceOpId), \
static_cast<bool>(NanOpt), \
static_cast<bool>(IndicesOpt), \
Rank, \
NumReduceDim)
#define ADD_BLOCKWISE_INST_REF_BY_TYPE( \
inT, compT, outT, ReduceOpId, PropagateNan, UseIndex, Rank, NumReduceDim) \
extern template void add_device_reduce_instance_blockwise<inT, \
compT, \
outT, \
Rank, \
NumReduceDim, \
ReduceOpId, \
PropagateNan, \
UseIndex>( \
std::vector<deviceReduceBlockWisePtrType<ReduceOpId>> & device_op_instances)
#define ADD_BLOCKWISE_INST_REF_BY_ID( \
inT, compT, outT, ReduceOpId, NanOpt, IndicesOpt, Rank, NumReduceDim) \
ADD_BLOCKWISE_INST_REF_BY_TYPE(inT, \
compT, \
outT, \
static_cast<ReduceTensorOp>(ReduceOpId), \
static_cast<bool>(NanOpt), \
static_cast<bool>(IndicesOpt), \
Rank, \
NumReduceDim)
} // namespace instance
} // namespace device
} // namespace tensor_operation
......
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16.hpp deleted 100644 → 0
View file @ 6ea9257e
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/data_type.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// clang-format off
// InDataType | AccDataType | OutDataType | ReduceOpId | NanPropaOpt | IndicesOpt | Rank | NumReduceDim
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 0, 0, 0, 4, 3); // for ADD
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 0, 0, 0, 4, 4);
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 0, 0, 0, 4, 1);
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 0, 0, 0, 2, 1);
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 5, 0, 0, 4, 3); // for AVG
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 5, 0, 0, 4, 4);
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 5, 0, 0, 4, 1);
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 5, 0, 0, 2, 1);
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 7, 0, 0, 4, 3); // for NORM2
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 7, 0, 0, 4, 4);
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 7, 0, 0, 4, 1);
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 7, 0, 0, 2, 1);
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 0, 4, 3); // for MIN
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 0, 4, 4);
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 0, 4, 1);
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 0, 2, 1);
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 0, 4, 3); // for MAX
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 0, 4, 4);
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 0, 4, 1);
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 0, 2, 1);
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 0, 4, 3); // for AMAX
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 0, 4, 4);
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 0, 4, 1);
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 0, 2, 1);
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 1, 4, 3); // for MIN
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 1, 4, 4);
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 1, 4, 1);
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 2, 0, 1, 2, 1);
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 1, 4, 3); // for MAX
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 1, 4, 4);
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 1, 4, 1);
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 3, 0, 1, 2, 1);
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 1, 4, 3); // for AMAX
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 1, 4, 4);
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 1, 4, 1);
ADD_BLOCKWISE_INST_REF_BY_ID(bhalf_t, float, bhalf_t, 4, 0, 1, 2, 1);
// clang-format on
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_add.hpp 0 → 100644
View file @ dda3a0a1
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/data_type.hpp"
#include "ck/utility/reduction_enums.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// clang-format off
// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
// clang-format on
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_amax.hpp 0 → 100644
View file @ dda3a0a1
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/data_type.hpp"
#include "ck/utility/reduction_enums.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// clang-format off
// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, UnaryAbs, PassThrough>>&);
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, UnaryAbs, PassThrough>>&);
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, UnaryAbs, PassThrough>>&);
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAMax, UnaryAbs, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, UnaryAbs, PassThrough>>&);
// clang-format on
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_avg.hpp 0 → 100644
View file @ dda3a0a1
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/data_type.hpp"
#include "ck/utility/reduction_enums.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// clang-format off
// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, UnaryDivide>>&);
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, UnaryDivide>>&);
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, UnaryDivide>>&);
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAdd, PassThrough, UnaryDivide, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, UnaryDivide>>&);
// clang-format on
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_max.hpp 0 → 100644
View file @ dda3a0a1
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/data_type.hpp"
#include "ck/utility/reduction_enums.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// clang-format off
// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceMax, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
// clang-format on
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_min.hpp 0 → 100644
View file @ dda3a0a1
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/data_type.hpp"
#include "ck/utility/reduction_enums.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// clang-format off
// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, false>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 3, PassThrough, PassThrough>>&);
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 4, PassThrough, PassThrough>>&);
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<4, 1, PassThrough, PassThrough>>&);
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceMin, PassThrough, PassThrough, false, true>(std::vector<DeviceReducePtr<2, 1, PassThrough, PassThrough>>&);
// clang-format on
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/include/ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise_b16_f32_b16_norm2.hpp 0 → 100644
View file @ dda3a0a1
// SPDX-License-Identifier: MIT
// Copyright (c) 2018-2022, Advanced Micro Devices, Inc. All rights reserved.
#pragma once
#include "ck/utility/data_type.hpp"
#include "ck/utility/reduction_enums.hpp"
#include "ck/library/tensor_operation_instance/gpu/reduce/device_reduce_instance_blockwise.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace instance {
// clang-format off
// InDataType | AccDataType | OutDataType | Rank | NumReduceDim | ReduceOperation | InElementwiseOp | AccElementwiseOp | PropagateNan | UseIndex
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 3, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 3, UnarySquare, UnarySqrt>>&);
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 4, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 4, UnarySquare, UnarySqrt>>&);
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 4, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<4, 1, UnarySquare, UnarySqrt>>&);
extern template void add_device_reduce_instance_blockwise<BF16, F32, BF16, 2, 1, ReduceAdd, UnarySquare, UnarySqrt, false, false>(std::vector<DeviceReducePtr<2, 1, UnarySquare, UnarySqrt>>&);
// clang-format on
} // namespace instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
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