Skip to content

GitLab

"vscode:/vscode.git/clone" did not exist on "28684f909dce1daa5674fb8f8dc9f1dad4638c0b"
Unverified Commit 56adf7e9 authored by Chao Liu's avatar Chao Liu Committed by GitHub
Browse files

GEMM with Multiple Source, GEMM+Bias+Add+FastGeLU example and ckProfiler (#241) 

* ad gelu and fast_gelu

* added GeLU and fast GeLU

* clean up

* add gemm+fastgelu example

* add gemm+gelu instances

* update profiler

* clean up

* clean up

* adding gemm+bias+activation

* clean

* adding bias

* clean

* adding gemm multiple d

* debugging

* add gemm bias add fastgelu

* rename, clean

* refactoring; add readme

* refactor

* refactor

* refactor

* refactor

* refactor

* refactor

* fix

* fix

* update example

* update example

* rename

* update example

* add ckProfiler

* clean

* clean

* clean

* clean

* add comment

* use type_convert

* clean

* clean element wise op
parent e4584d91
Show whitespace changes
Inline Side-by-side
Showing with 1138 additions and 57 deletions
include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v7.hpp 0 → 100644
View file @ 56adf7e9
#pragma once
#include "common_header.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "tensor_space_filling_curve.hpp"
namespace ck {
// Thread-level multi-source, multi-destination tensor slice data movement
// Assume:
// 1. All sources and destinations are DynamicBuffer
// 2. Same VectorDim and ScalerPerVector for all sources and destinations
// 3. DstInMemOps are per destination tensor
// 4. ThreadTransferSrcResetCoordinateAfterRunFlags are per source tensor
// 5. ThreadTransferDstResetCoordinateAfterRunFlags are per destination tensor
// 6. Does not need to know src_descs and dst_descs at compile-time
// 7. Does not need to know src_slice_origins and dst_slice_origins at compile-time,
//
// Does following things to avoid scratch memory issue
// 1. Use StaticallyIndexedArray or vector_type instead of C array for thread buffer
// 2. Pass tensor descritpors by reference (or tuple of references)
// 3. Does not keep reference to tensor descriptor
// 4. Does not construct new tensor coordinate when call Run()
template <typename SrcDatas,
typename DstDatas,
typename SrcDescs,
typename DstDescs,
typename ElementwiseOperation,
typename DstInMemOps, // Sequence<InMemoryDataOperationEnum ...>
typename SliceLengths,
typename DimAccessOrder,
index_t VectorDim,
index_t ScalarPerVector,
typename SrcResetCoordinateAfterRunFlags, // Sequence<bool ...>
typename DstResetCoordinateAfterRunFlags> // Sequence<bool ...>
struct ThreadwiseTensorSliceTransfer_v7
{
static constexpr auto I0 = Number<0>{};
static constexpr index_t nDim = SliceLengths::Size();
static constexpr index_t nSrc = SrcDescs::Size();
static constexpr index_t nDst = DstDescs::Size();
using Index = MultiIndex<nDim>;
// return a tuple of coordiantes for a tuple of tensor
template <typename Descs,
typename Indices,
enable_if_t<Descs::Size() == Indices::Size(), bool> = false>
static constexpr auto MakeCoordinates(const Descs& descs, const Indices& indices)
{
return generate_tuple([&](auto i) { return make_tensor_coordinate(descs[i], indices[i]); },
Number<Descs::Size()>{});
}
using SrcCoords = decltype(MakeCoordinates(SrcDescs{}, StaticallyIndexedArray<Index, nSrc>{}));
using DstCoords = decltype(MakeCoordinates(DstDescs{}, StaticallyIndexedArray<Index, nDst>{}));
// scalar per access on each dim
// FIXME: don't use lambda_scalar_per_access
static constexpr auto scalar_per_access = generate_sequence(
detail::lambda_scalar_per_access<VectorDim, ScalarPerVector>{}, Number<nDim>{});
using SpaceFillingCurve =
SpaceFillingCurve<SliceLengths, DimAccessOrder, remove_cv_t<decltype(scalar_per_access)>>;
__device__ constexpr ThreadwiseTensorSliceTransfer_v7(
const SrcDescs& src_descs,
const StaticallyIndexedArray<Index, nSrc>& src_slice_origins,
const DstDescs& dst_descs,
const StaticallyIndexedArray<Index, nDst>& dst_slice_origins,
const ElementwiseOperation& element_op)
: src_coords_(MakeCoordinates(src_descs, src_slice_origins)),
dst_coords_(MakeCoordinates(dst_descs, dst_slice_origins)),
element_op_(element_op)
{
static_assert(SliceLengths::At(Number<VectorDim>{}) % ScalarPerVector == 0,
"wrong! cannot evenly divide");
}
template <typename Indices, enable_if_t<SrcDescs::Size() == Indices::Size(), bool> = false>
__device__ void SetSrcSliceOrigins(const SrcDescs& src_descs,
const Indices& src_slice_origin_idxs)
{
static_for<0, nSrc, 1>{}([&](auto i) {
src_coords_(i) = make_tensor_coordinate(src_descs[i], src_slice_origin_idxs[i]);
});
}
template <typename Indices, enable_if_t<DstDescs::Size() == Indices::Size(), bool> = false>
__device__ void SetDstSliceOrigins(const DstDescs& dst_descs,
const Indices& dst_slice_origin_idxs)
{
static_for<0, nDst, 1>{}([&](auto i) {
dst_coords_(i) = make_tensor_coordinate(dst_descs[i], dst_slice_origin_idxs[i]);
});
}
// SrcDescs: Tuple<const SrcDesc0&, const SrcDesc1&, ...>
// SrcBuffers: Tuple<const SrcBuffer0&, const SrcBuffer1&, ...>
// DstDescs: Tuple<const DstDesc0&, const DstDesc1&, ...>
// DstBuffers: Tuple<const DstBuffer0&, const DstBuffer1&, ...>
template <typename SrcBuffers,
typename DstBuffers,
enable_if_t<SrcDescs::Size() == SrcBuffers::Size() &&
DstDescs::Size() == DstBuffers::Size(),
bool> = false>
__device__ void Run(const SrcDescs& src_descs,
const SrcBuffers& src_bufs,
const DstDescs& dst_descs,
DstBuffers dst_bufs)
{
auto generate_vectors = [&](auto data_types) {
constexpr index_t num = data_types.Size();
return generate_tuple(
[&](auto i) {
using DataType = remove_cvref_t<decltype(data_types[i])>;
return vector_type_maker_t<DataType, ScalarPerVector>{};
},
Number<num>{});
};
constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
// loop over space-filling curve
static_for<0, num_access, 1>{}([&](auto iAccess) {
auto src_vectors = generate_vectors(SrcDatas{});
auto dst_vectors = generate_vectors(DstDatas{});
// copy data from src_bufs into src_vectors
static_for<0, nSrc, 1>{}([&](auto i) {
using src_vector_t = typename remove_cvref_t<decltype(src_vectors[i])>::type;
const bool is_src_valid =
coordinate_has_valid_offset_assuming_visible_index_is_valid(src_descs[i],
src_coords_[i]);
src_vectors(i).template AsType<src_vector_t>()(I0) =
src_bufs[i].template Get<src_vector_t>(src_coords_[i].GetOffset(),
is_src_valid);
});
// apply pointwise function
static_for<0, ScalarPerVector, 1>{}([&](auto i) {
// get reference to src data
const auto src_data_refs = generate_tie(
// return type should be lvalue
[&](auto iSrc) -> const auto& {
using SrcData = remove_cvref_t<tuple_element_t<iSrc.value, SrcDatas>>;
return src_vectors[iSrc].template AsType<SrcData>()[i];
},
Number<nSrc>{});
// get reference to dst data
auto dst_data_refs = generate_tie(
// return type should be lvalue
[&](auto iDst) -> auto& {
using DstData = remove_cvref_t<tuple_element_t<iDst.value, DstDatas>>;
return dst_vectors(iDst).template AsType<DstData>()(i);
},
Number<nDst>{});
// apply pointwise function
// pointwise function signature:
// element_op_(dst_data_refs[I0],
// dst_data_refs[I1],
// ...,
// src_data_refs[I0],
// src_data_refs[I1],
// ...)
unpack2(element_op_, dst_data_refs, src_data_refs);
});
// copy data from buf_vectors into dst_bufs
static_for<0, nDst, 1>{}([&](auto i) {
using dst_vector_t = typename remove_cvref_t<decltype(dst_vectors[i])>::type;
const bool is_dst_valid =
coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_descs[i],
dst_coords_[i]);
constexpr InMemoryDataOperationEnum DstInMemOp =
static_cast<InMemoryDataOperationEnum>(DstInMemOps::At(i.value));
dst_bufs(i).template Update<DstInMemOp, dst_vector_t>(
dst_coords_[i].GetOffset(),
is_dst_valid,
dst_vectors[i].template AsType<dst_vector_t>()[I0]);
});
// move coordinate
if constexpr(iAccess.value != num_access - 1)
{
constexpr auto forward_step = SpaceFillingCurve::GetForwardStep(iAccess);
static_for<0, nSrc, 1>{}([&](auto i) {
move_tensor_coordinate(src_descs[i],
src_coords_(i),
make_tensor_coordinate_step(src_descs[i], forward_step));
});
static_for<0, nDst, 1>{}([&](auto i) {
move_tensor_coordinate(dst_descs[i],
dst_coords_(i),
make_tensor_coordinate_step(dst_descs[i], forward_step));
});
}
});
// move coordinate back to slice origin (or not)
static_for<0, nSrc, 1>{}([&](auto i) {
if constexpr(SrcResetCoordinateAfterRunFlags::At(i))
{
const auto src_reset_step =
make_tensor_coordinate_step(src_descs[i], GetCoordinateResetStep());
move_tensor_coordinate(src_descs[i], src_coords_(i), src_reset_step);
}
});
static_for<0, nDst, 1>{}([&](auto i) {
if constexpr(DstResetCoordinateAfterRunFlags::At(i))
{
const auto dst_reset_step =
make_tensor_coordinate_step(dst_descs[i], GetCoordinateResetStep());
move_tensor_coordinate(dst_descs[i], dst_coords_(i), dst_reset_step);
}
});
}
__device__ static constexpr auto GetCoordinateResetStep()
{
constexpr auto num_access = SpaceFillingCurve::GetNumOfAccess();
if constexpr(num_access == 0)
{
return typename SpaceFillingCurve::Index{};
}
else
{
constexpr auto reset_step =
SpaceFillingCurve::GetStepBetween(Number<num_access - 1>{}, Number<0>{});
return reset_step;
}
}
// src_slice_origin_step_idx need to be known at compile-time, for performance reason
template <index_t ISrc>
__device__ void MoveSrcSliceWindow(const SrcDescs& src_descs,
Number<ISrc> iSrc,
const Index& src_slice_origin_step_idx)
{
// if src coord was not reset by RunRead(), then need to adjust the step here
const auto adjusted_step_idx = SrcResetCoordinateAfterRunFlags::At(iSrc)
? src_slice_origin_step_idx
: src_slice_origin_step_idx + GetCoordinateResetStep();
// is it OK to construct a new step every time?
const auto adjusted_step = make_tensor_coordinate_step(src_descs[iSrc], adjusted_step_idx);
move_tensor_coordinate(src_descs[iSrc], src_coords_(iSrc), adjusted_step);
}
// dst_slice_origin_step_idx need to be known at compile-time, for performance reason
template <index_t IDst>
__device__ void MoveDstSliceWindow(const DstDescs& dst_descs,
Number<IDst> iDst,
const Index& dst_slice_origin_step_idx)
{
// if dst coord was not reset by Run(), then need to adjust the step here
const auto adjusted_step_idx = DstResetCoordinateAfterRunFlags::At(iDst)
? dst_slice_origin_step_idx
: dst_slice_origin_step_idx + GetCoordinateResetStep();
// is it OK to construct a new step every time?
const auto adjusted_step = make_tensor_coordinate_step(dst_descs[iDst], adjusted_step_idx);
move_tensor_coordinate(dst_descs[iDst], dst_coords_(iDst), adjusted_step);
}
private:
SrcCoords src_coords_;
DstCoords dst_coords_;
const ElementwiseOperation element_op_;
};
} // namespace ck
include/ck/utility/amd_buffer_addressing.hpp
View file @ 56adf7e9
......@@ -6,6 +6,8 @@ namespace ck {
template <typename T>
union BufferResource
{
__device__ constexpr BufferResource() : content{} {}
// 128 bit SGPRs to supply buffer resource in buffer instructions
// https://rocm-documentation.readthedocs.io/en/latest/GCN_ISA_Manuals/testdocbook.html#vector-memory-buffer-instructions
int32x4_t content;
......
include/ck/utility/data_type.hpp
View file @ 56adf7e9
#pragma once
#include "statically_indexed_array.hpp"
namespace ck {
......
include/ck/utility/enable_if.hpp
View file @ 56adf7e9
#ifndef CK_ENABLE_IF_HPP
#define CK_ENABLE_IF_HPP
#pragma once
namespace ck {
......@@ -10,4 +9,3 @@ template <bool B, typename T = void>
using enable_if_t = typename std::enable_if<B, T>::type;
} // namespace ck
#endif
include/ck/utility/sequence.hpp
View file @ 56adf7e9
#ifndef CK_SEQUENCE_HPP
#define CK_SEQUENCE_HPP
#pragma once
#include "integral_constant.hpp"
#include "type.hpp"
......@@ -241,7 +240,13 @@ struct arithmetic_sequence_gen
}
};
using type = typename sequence_gen<(IEnd - IBegin) / Increment, F>::type;
using type0 = typename sequence_gen<(IEnd - IBegin) / Increment, F>::type;
using type1 = Sequence<>;
static constexpr bool kHasContent =
(Increment > 0 && IBegin < IEnd) || (Increment < 0 && IBegin > IEnd);
using type = typename conditional<kHasContent, type0, type1>::type;
};
// uniform sequence
......@@ -882,5 +887,10 @@ __host__ __device__ constexpr bool sequence_all_of(Seq, F f)
return flag;
}
template <typename Sx, typename Sy>
using sequence_merge_t = typename sequence_merge<Sx, Sy>::type;
template <index_t NSize, index_t I>
using uniform_sequence_gen_t = typename uniform_sequence_gen<NSize, I>::type;
} // namespace ck
#endif
include/ck/utility/tuple.hpp
View file @ 56adf7e9
#ifndef CK_TUPLE_HPP
#define CK_TUPLE_HPP
#pragma once
#include "integral_constant.hpp"
#include "sequence.hpp"
......@@ -17,14 +16,18 @@ struct TupleElementKey
};
template <typename Key, typename Data>
struct TupleElement
struct TupleElementKeyData
{
__host__ __device__ constexpr TupleElement() = default;
#if 0 // workaround compiler complaint about implicitly-deleted default constructor
__host__ __device__ constexpr TupleElementKeyData() = default;
#else
__host__ __device__ constexpr TupleElementKeyData() : mData{} {}
#endif
template <
typename T,
typename enable_if<!is_same<remove_cvref_t<T>, TupleElement>::value, bool>::type = false>
__host__ __device__ constexpr TupleElement(T&& v) : mData(std::forward<T>(v))
template <typename T,
typename enable_if<!is_same<remove_cvref_t<T>, TupleElementKeyData>::value,
bool>::type = false>
__host__ __device__ constexpr TupleElementKeyData(T&& v) : mData(std::forward<T>(v))
{
}
......@@ -32,20 +35,21 @@ struct TupleElement
};
template <typename Key, typename Data>
__host__ __device__ constexpr const Data& get_tuple_element(const TupleElement<Key, Data>& x)
__host__ __device__ constexpr const Data&
get_tuple_element_data(const TupleElementKeyData<Key, Data>& x)
{
return static_cast<const Data&>(x.mData);
}
template <typename Key, typename Data>
__host__ __device__ constexpr Data& get_tuple_element(TupleElement<Key, Data>& x)
__host__ __device__ constexpr Data& get_tuple_element_data(TupleElementKeyData<Key, Data>& x)
{
return x.mData;
}
// TODO: not sure the use of reference is correct
template <typename Key, typename Data>
__host__ __device__ constexpr Data&& get_tuple_element(TupleElement<Key, Data>&& x)
__host__ __device__ constexpr Data&& get_tuple_element_data(TupleElementKeyData<Key, Data>&& x)
{
return static_cast<Data&&>(x.mData);
}
......@@ -54,7 +58,7 @@ template <typename Indices, typename... Xs>
struct TupleImpl;
template <index_t... Is, typename... Xs>
struct TupleImpl<Sequence<Is...>, Xs...> : TupleElement<TupleElementKey<Is>, Xs>...
struct TupleImpl<Sequence<Is...>, Xs...> : TupleElementKeyData<TupleElementKey<Is>, Xs>...
{
__host__ __device__ constexpr TupleImpl() = default;
......@@ -63,13 +67,13 @@ struct TupleImpl<Sequence<Is...>, Xs...> : TupleElement<TupleElementKey<Is>, Xs>
!is_same<remove_cvref_t<Y>, TupleImpl>::value,
bool>::type = false>
__host__ __device__ constexpr TupleImpl(Y&& y)
: TupleElement<TupleElementKey<Is>, Xs>(std::forward<Y>(y))...
: TupleElementKeyData<TupleElementKey<Is>, Xs>(std::forward<Y>(y))...
{
}
template <typename... Ys, typename enable_if<sizeof...(Ys) >= 2, bool>::type = false>
__host__ __device__ constexpr TupleImpl(Ys&&... ys)
: TupleElement<TupleElementKey<Is>, Xs>(std::forward<Ys>(ys))...
: TupleElementKeyData<TupleElementKey<Is>, Xs>(std::forward<Ys>(ys))...
{
static_assert(sizeof...(Is) == sizeof...(Xs) && sizeof...(Is) == sizeof...(Ys),
"wrong! inconsistent size");
......@@ -78,15 +82,15 @@ struct TupleImpl<Sequence<Is...>, Xs...> : TupleElement<TupleElementKey<Is>, Xs>
__host__ __device__ static constexpr index_t Size() { return sizeof...(Xs); }
template <index_t I>
__host__ __device__ constexpr const auto& GetElementByKey(TupleElementKey<I>) const
__host__ __device__ constexpr const auto& GetElementDataByKey(TupleElementKey<I>) const
{
return get_tuple_element<TupleElementKey<I>>(*this);
return get_tuple_element_data<TupleElementKey<I>>(*this);
}
template <index_t I>
__host__ __device__ constexpr auto& GetElementByKey(TupleElementKey<I>)
__host__ __device__ constexpr auto& GetElementDataByKey(TupleElementKey<I>)
{
return get_tuple_element<TupleElementKey<I>>(*this);
return get_tuple_element_data<TupleElementKey<I>>(*this);
}
};
......@@ -121,7 +125,7 @@ struct Tuple : detail::TupleImpl<typename arithmetic_sequence_gen<0, sizeof...(X
__host__ __device__ constexpr const auto& At(Number<I>) const
{
static_assert(I < base::Size(), "wrong! out of range");
return base::GetElementByKey(detail::TupleElementKey<I>{});
return base::GetElementDataByKey(detail::TupleElementKey<I>{});
}
// write access
......@@ -129,7 +133,7 @@ struct Tuple : detail::TupleImpl<typename arithmetic_sequence_gen<0, sizeof...(X
__host__ __device__ constexpr auto& At(Number<I>)
{
static_assert(I < base::Size(), "wrong! out of range");
return base::GetElementByKey(detail::TupleElementKey<I>{});
return base::GetElementDataByKey(detail::TupleElementKey<I>{});
}
// read access
......@@ -159,6 +163,31 @@ struct Tuple : detail::TupleImpl<typename arithmetic_sequence_gen<0, sizeof...(X
__host__ __device__ static constexpr bool IsStaticBuffer() { return true; }
};
template <>
struct Tuple<>
{
__host__ __device__ constexpr Tuple() = default;
__host__ __device__ static constexpr index_t Size() { return 0; }
template <typename T>
__host__ __device__ constexpr auto operator=(const T&)
{
return *this;
}
__host__ __device__ static constexpr bool IsStaticBuffer() { return true; }
};
template <index_t I, typename TTuple>
struct tuple_element
{
using type = decltype(TTuple{}.At(Number<I>{}));
};
template <index_t I, typename TTuple>
using tuple_element_t = typename tuple_element<I, TTuple>::type;
template <typename... Xs>
__host__ __device__ constexpr auto make_tuple(Xs&&... xs)
{
......@@ -173,4 +202,3 @@ constexpr Tuple<Args&...> tie(Args&... args) noexcept
}
} // namespace ck
#endif
include/ck/utility/tuple_helper.hpp
View file @ 56adf7e9
#ifndef CK_TUPLE_HELPER_HPP
#define CK_TUPLE_HELPER_HPP
#pragma once
#include "functional4.hpp"
#include "tuple.hpp"
......@@ -20,6 +19,17 @@ __host__ __device__ constexpr auto generate_tie(F&& f, Number<N>)
typename arithmetic_sequence_gen<0, N, 1>::type{});
}
// tx and ty are tuple of references, return type of will tuple of referennce (not rvalue)
template <typename... X, typename... Y>
__host__ __device__ constexpr auto concat_tuple_of_reference(const Tuple<X&...>& tx,
const Tuple<Y&...>& ty)
{
return unpack2(
[&](auto&&... zs) { return Tuple<decltype(zs)...>{std::forward<decltype(zs)>(zs)...}; },
tx,
ty);
}
namespace detail {
template <typename F, typename X, index_t... Is>
......@@ -66,4 +76,3 @@ __host__ __device__ constexpr auto transform_tuples(F f, const X& x, const Y& y,
}
} // namespace ck
#endif
library/include/ck/library/reference_tensor_operation/cpu/reference_gemm_bias_2d.hpp
View file @ 56adf7e9
......@@ -66,8 +66,8 @@ struct ReferenceGemmBias2D : public device::BaseOperator
for(int k = 0; k < K; ++k)
{
arg.a_element_op_(a, static_cast<AccDataType>(arg.a_m_k_(m, k)));
arg.b_element_op_(b, static_cast<AccDataType>(arg.b_k_n_(k, n)));
arg.a_element_op_(a, ck::type_convert<AccDataType>(arg.a_m_k_(m, k)));
arg.b_element_op_(b, ck::type_convert<AccDataType>(arg.b_k_n_(k, n)));
acc += a * b;
}
......
library/include/ck/library/tensor_operation_instance/device_operation_instance.hpp
View file @ 56adf7e9
#ifndef CK_DEVICE_OPERATION_INSTANCE_HPP
#define CK_DEVICE_OPERATION_INSTANCE_HPP
#pragma once
#include <stdlib.h>
#include <vector>
namespace ck {
namespace tensor_operation {
......@@ -23,4 +22,3 @@ void add_device_operation_instances(std::vector<std::unique_ptr<OpInstance>>& op
} // namespace device
} // namespace tensor_operation
} // namespace ck
#endif
library/src/tensor_operation_instance/gpu/CMakeLists.txt
View file @ 56adf7e9
......@@ -44,6 +44,7 @@ add_subdirectory(convnd_bwd_data)
add_subdirectory(grouped_gemm)
add_subdirectory(conv2d_bwd_weight)
add_subdirectory(batched_gemm_reduce)
add_subdirectory(gemm_add_add_fastgelu)
add_library(device_operations STATIC
$<TARGET_OBJECTS:device_conv1d_fwd_instance>
......@@ -63,6 +64,7 @@ add_library(device_operations STATIC
$<TARGET_OBJECTS:device_conv2d_bwd_weight_instance>
$<TARGET_OBJECTS:device_batched_gemm_reduce_instance>
$<TARGET_OBJECTS:device_conv3d_fwd_instance>
$<TARGET_OBJECTS:device_gemm_add_add_fastgelu_instance>
device_conv2d.cpp
)
add_library(composablekernels::device_operations ALIAS device_operations)
......@@ -97,9 +99,11 @@ target_include_directories(device_operations PUBLIC
#once new arches are enabled make this an option on the main cmake file
# and pass down here to be exported
target_compile_options(device_operations
PRIVATE --offload-arch=gfx908
target_compile_options(device_operations PRIVATE
--offload-arch=gfx908
--offload-arch=gfx90a
)
# install(TARGETS device_operations LIBRARY DESTINATION lib)
install(TARGETS device_operations
EXPORT device_operationsTargets
......
library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/CMakeLists.txt 0 → 100644
View file @ 56adf7e9
# device_gemm_add_add_fastgelu_instance
set(DEVICE_GEMM_ADD_ADD_FASTGELU_INSTANCE_SOURCE
device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp;
device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp;
device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp;
device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp;
)
add_library(device_gemm_add_add_fastgelu_instance OBJECT ${DEVICE_GEMM_ADD_ADD_FASTGELU_INSTANCE_SOURCE})
target_compile_features(device_gemm_add_add_fastgelu_instance PUBLIC)
set_target_properties(device_gemm_add_add_fastgelu_instance PROPERTIES POSITION_INDEPENDENT_CODE ON)
clang_tidy_check(device_gemm_add_add_fastgelu_instance)
library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instance.cpp 0 → 100644
View file @ 56adf7e9
#include <stdlib.h>
#include "config.hpp"
#include "element_wise_operation.hpp"
#include "device_operation_instance.hpp"
#include "device_gemm_multiple_d_xdl_cshuffle.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using F16_F16 = ck::Tuple<F16, F16>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// e = elementwise((a * b), d)
// outout: e[m, n]
// input: a[k, m], b[k, n], d[m, n]
using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances = std::tuple<
// clang-format off
//##############################| ALayout| BLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##############################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//##############################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 256, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 128, 256, 32, 2, 2, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 128, 128, 32, 2, 2, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 128, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 128, 64, 32, 2, 2, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 64, 128, 32, 2, 2, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 128, 64, 32, 2, 2, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 64, 128, 32, 2, 2, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>
// clang-format on
>;
void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(
std::vector<DeviceGemmMultipleDPtr<2, PassThrough, PassThrough, AddAddFastGelu>>& instances)
{
add_device_operation_instances(
instances, device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances{});
}
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instance.cpp 0 → 100644
View file @ 56adf7e9
#include <stdlib.h>
#include "config.hpp"
#include "element_wise_operation.hpp"
#include "device_operation_instance.hpp"
#include "device_gemm_multiple_d_xdl_cshuffle.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using F16_F16 = ck::Tuple<F16, F16>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// e = elementwise((a * b), d)
// outout: e[m, n]
// input: a[k, m], b[n, k], d[m, n]
using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances = std::tuple<
// clang-format off
//##############################| ALayout| BLayout| ELayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##############################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//##############################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 256, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 128, 256, 32, 2, 8, 32, 32, 2, 4, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 128, 128, 32, 2, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 128, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 128, 64, 32, 2, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 64, 128, 32, 2, 8, 32, 32, 2, 2, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 128, 64, 32, 2, 8, 32, 32, 2, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 64, 128, 32, 2, 8, 32, 32, 1, 2, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Col, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>
// clang-format on
>;
void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(
std::vector<DeviceGemmMultipleDPtr<2, PassThrough, PassThrough, AddAddFastGelu>>& instances)
{
add_device_operation_instances(
instances, device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances{});
}
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instance.cpp 0 → 100644
View file @ 56adf7e9
#include <stdlib.h>
#include "config.hpp"
#include "element_wise_operation.hpp"
#include "device_operation_instance.hpp"
#include "device_gemm_multiple_d_xdl_cshuffle.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using F16_F16 = ck::Tuple<F16, F16>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// e = elementwise((a * b), d)
// outout: e[m, n]
// input: a[m, k], b[k, n], d[m, n]
using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances = std::tuple<
// clang-format off
//##############################| ALayout| BLayout| CLayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##############################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//##############################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 256, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 128, 256, 32, 8, 2, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 128, 128, 32, 8, 2, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 128, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 128, 64, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 64, 128, 32, 8, 2, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 128, 64, 32, 8, 2, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<16,16, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 1, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 64, 128, 32, 8, 2, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<8, 32, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 4, 2, 0, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Row, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<0, 2, 1>, S<0, 2, 1>, 1, 2, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>
// clang-format on
>;
void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(
std::vector<DeviceGemmMultipleDPtr<2, PassThrough, PassThrough, AddAddFastGelu>>& instances)
{
add_device_operation_instances(
instances, device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances{});
}
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
library/src/tensor_operation_instance/gpu/gemm_add_add_fastgelu/device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instance.cpp 0 → 100644
View file @ 56adf7e9
#include <stdlib.h>
#include "config.hpp"
#include "element_wise_operation.hpp"
#include "device_operation_instance.hpp"
#include "device_gemm_multiple_d_xdl_cshuffle.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using F16 = ck::half_t;
using F32 = float;
using F16_F16 = ck::Tuple<F16, F16>;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
static constexpr auto GemmDefault = ck::tensor_operation::device::GemmSpecialization::Default;
// e = elementwise((a * b), d)
// outout: e[m, n]
// input: a[m, k], b[n, k], d[m, n]
using device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances = std::tuple<
// clang-format off
//##############################| ALayout| BLayout| CLayout| AData| BData| AccData| CShuffle| DsData| EData| A| B| CDE| GEMM| NumGemmK| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##############################| | | | Type| Type| Type| DataType| Type| Type| Elementwise| Elementwise| Elementwise| Spacialization| Prefetch| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MWaveMPerXdl| ScalarPerVector|
//##############################| | | | | | | | | | Operation| Operation| Operation| | Stage| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NWaveNPerXdl| _NWaveNPerXdl|
//##############################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 32, 1, 4>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 8>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>,
DeviceGemmMultipleD_Xdl_CShuffle< Row, Col, Row, F16, F16, F32, F32, F16_F16, F16, PassThrough, PassThrough, AddAddFastGelu, GemmDefault, 1, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, 1, 1, 1, S<1, 16, 1, 4>, 8>
// clang-format on
>;
void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(
std::vector<DeviceGemmMultipleDPtr<2, PassThrough, PassThrough, AddAddFastGelu>>& instances)
{
add_device_operation_instances(
instances, device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances{});
}
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
profiler/CMakeLists.txt
View file @ 56adf7e9
......@@ -40,6 +40,7 @@ set(PROFILER_SOURCE
src/profile_grouped_gemm.cpp
src/profile_conv_bwd_weight.cpp
src/profile_batched_gemm_reduce.cpp
src/profile_gemm_add_add_fastgelu.cpp
)
add_executable(ckProfiler ${PROFILER_SOURCE})
......@@ -64,3 +65,4 @@ target_link_libraries(ckProfiler PRIVATE device_reduce_instance)
target_link_libraries(ckProfiler PRIVATE device_grouped_gemm_instance)
target_link_libraries(ckProfiler PRIVATE device_conv2d_bwd_weight_instance)
target_link_libraries(ckProfiler PRIVATE device_batched_gemm_reduce_instance)
target_link_libraries(ckProfiler PRIVATE device_gemm_add_add_fastgelu_instance)
profiler/include/profile_gemm_add_add_fastgelu_impl.hpp 0 → 100644
View file @ 56adf7e9
#pragma once
#include <iomanip>
#include "check_err.hpp"
#include "config.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "host_conv.hpp"
#include "tensor_layout.hpp"
#include "device_tensor.hpp"
#include "element_wise_operation.hpp"
#include "reference_gemm.hpp"
#include "device_gemm_multiple_d.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
namespace device_gemm_instance {
using DeviceGemmAddAddFastGeluPtr = ck::tensor_operation::device::DeviceGemmMultipleDPtr<
2,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::PassThrough,
ck::tensor_operation::element_wise::AddAddFastGelu>;
void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(
std::vector<DeviceGemmAddAddFastGeluPtr>&);
void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(
std::vector<DeviceGemmAddAddFastGeluPtr>&);
void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(
std::vector<DeviceGemmAddAddFastGeluPtr>&);
void add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(
std::vector<DeviceGemmAddAddFastGeluPtr>&);
} // namespace device_gemm_instance
} // namespace device
} // namespace tensor_operation
} // namespace ck
namespace ck {
namespace profiler {
template <typename ADataType,
typename BDataType,
typename AccDataType,
typename D0DataType,
typename D1DataType,
typename EDataType,
typename ALayout,
typename BLayout,
typename D0Layout,
typename D1Layout,
typename ELayout>
int profile_gemm_add_add_fastgelu_impl(int do_verification,
int init_method,
bool /*do_log*/,
bool time_kernel,
int M,
int N,
int K,
int StrideA,
int StrideB,
int StrideD0,
int StrideD1,
int StrideE)
{
auto f_host_tensor_descriptor =
[](std::size_t row, std::size_t col, std::size_t stride, auto layout) {
if(is_same<decltype(layout), tensor_layout::gemm::RowMajor>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({stride, 1}));
}
else
{
return HostTensorDescriptor(std::vector<std::size_t>({row, col}),
std::vector<std::size_t>({1, stride}));
}
};
Tensor<ADataType> a_m_k(f_host_tensor_descriptor(M, K, StrideA, ALayout{}));
Tensor<BDataType> b_k_n(f_host_tensor_descriptor(K, N, StrideB, BLayout{}));
Tensor<D0DataType> d0_m_n(f_host_tensor_descriptor(M, N, StrideD0, D0Layout{}));
Tensor<D1DataType> d1_m_n(f_host_tensor_descriptor(M, N, StrideD1, D1Layout{}));
Tensor<EDataType> e_m_n_device_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
Tensor<EDataType> e_m_n_host_result(f_host_tensor_descriptor(M, N, StrideE, ELayout{}));
std::cout << "a_m_k: " << a_m_k.mDesc << std::endl;
std::cout << "b_k_n: " << b_k_n.mDesc << std::endl;
std::cout << "d0_m_n: " << d0_m_n.mDesc << std::endl;
std::cout << "d1_m_n: " << d1_m_n.mDesc << std::endl;
std::cout << "e_m_n: " << e_m_n_device_result.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
a_m_k.GenerateTensorValue(GeneratorTensor_2<ADataType>{-5, 5});
b_k_n.GenerateTensorValue(GeneratorTensor_2<BDataType>{-5, 5});
d0_m_n.GenerateTensorValue(GeneratorTensor_2<D0DataType>{-5, 5});
d1_m_n.GenerateTensorValue(GeneratorTensor_2<D1DataType>{-5, 5});
break;
default:
a_m_k.GenerateTensorValue(GeneratorTensor_3<ADataType>{0.0, 1.0});
b_k_n.GenerateTensorValue(GeneratorTensor_3<BDataType>{-0.5, 0.5});
d0_m_n.GenerateTensorValue(GeneratorTensor_3<D0DataType>{0.0, 1.0});
d1_m_n.GenerateTensorValue(GeneratorTensor_3<D1DataType>{0.0, 1.0});
}
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using AddAddFastGelu = ck::tensor_operation::element_wise::AddAddFastGelu;
using AElementOp = PassThrough;
using BElementOp = PassThrough;
using CDEElementOp = AddAddFastGelu;
const auto a_element_op = AElementOp{};
const auto b_element_op = BElementOp{};
const auto cde_element_op = CDEElementOp{};
// add device GEMM instances
std::vector<ck::tensor_operation::device::device_gemm_instance::DeviceGemmAddAddFastGeluPtr>
device_op_ptrs;
if constexpr(is_same_v<ADataType, half_t> && is_same_v<BDataType, half_t> &&
is_same_v<EDataType, half_t>)
{
if constexpr(is_same_v<ALayout, tensor_layout::gemm::RowMajor> &&
is_same_v<BLayout, tensor_layout::gemm::RowMajor> &&
is_same_v<ELayout, tensor_layout::gemm::RowMajor>)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_kn_mn_instances(
device_op_ptrs);
}
else if constexpr(is_same_v<ALayout, tensor_layout::gemm::RowMajor> &&
is_same_v<BLayout, tensor_layout::gemm::ColumnMajor> &&
is_same_v<ELayout, tensor_layout::gemm::RowMajor>)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_mk_nk_mn_instances(
device_op_ptrs);
}
else if constexpr(is_same_v<ALayout, tensor_layout::gemm::ColumnMajor> &&
is_same_v<BLayout, tensor_layout::gemm::RowMajor> &&
is_same_v<ELayout, tensor_layout::gemm::RowMajor>)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_km_kn_mn_instances(
device_op_ptrs);
}
else if constexpr(is_same_v<ALayout, tensor_layout::gemm::ColumnMajor> &&
is_same_v<BLayout, tensor_layout::gemm::ColumnMajor> &&
is_same_v<ELayout, tensor_layout::gemm::RowMajor>)
{
ck::tensor_operation::device::device_gemm_instance::
add_device_gemm_add_add_fastgelu_xdl_c_shuffle_f16_f16_f16_km_nk_mn_instances(
device_op_ptrs);
}
}
std::cout << "found " << device_op_ptrs.size() << " instances" << std::endl;
// run reference
if(do_verification)
{
Tensor<AccDataType> c_m_n(HostTensorDescriptor(
std::vector<std::size_t>{static_cast<std::size_t>(M), static_cast<std::size_t>(N)}));
using ReferenceGemmInstance = ck::tensor_operation::host::ReferenceGemm<ADataType,
BDataType,
AccDataType,
AccDataType,
AElementOp,
BElementOp,
PassThrough>;
auto ref_gemm = ReferenceGemmInstance{};
auto ref_invoker = ref_gemm.MakeInvoker();
auto ref_argument =
ref_gemm.MakeArgument(a_m_k, b_k_n, c_m_n, a_element_op, b_element_op, PassThrough{});
ref_invoker.Run(ref_argument);
for(int m = 0; m < M; ++m)
{
for(int n = 0; n < N; ++n)
{
cde_element_op(e_m_n_host_result(m, n), c_m_n(m, n), d0_m_n(m, n), d1_m_n(m, n));
}
}
}
DeviceMem a_device_buf(sizeof(ADataType) * a_m_k.mDesc.GetElementSpace());
DeviceMem b_device_buf(sizeof(BDataType) * b_k_n.mDesc.GetElementSpace());
DeviceMem d0_m_n_device_buf(sizeof(D0DataType) * d0_m_n.mDesc.GetElementSpace());
DeviceMem d1_m_n_device_buf(sizeof(D1DataType) * d1_m_n.mDesc.GetElementSpace());
DeviceMem e_device_buf(sizeof(EDataType) * e_m_n_device_result.mDesc.GetElementSpace());
a_device_buf.ToDevice(a_m_k.mData.data());
b_device_buf.ToDevice(b_k_n.mData.data());
d0_m_n_device_buf.ToDevice(d0_m_n.mData.data());
d1_m_n_device_buf.ToDevice(d1_m_n.mData.data());
std::string best_device_op_name;
float best_ave_time = 0;
float best_tflops = 0;
float best_gb_per_sec = 0;
bool pass = true;
// profile device operation instances
for(auto& device_op_ptr : device_op_ptrs)
{
auto argument_ptr = device_op_ptr->MakeArgumentPointer(
a_device_buf.GetDeviceBuffer(),
b_device_buf.GetDeviceBuffer(),
std::array<const void*, 2>{d0_m_n_device_buf.GetDeviceBuffer(),
d1_m_n_device_buf.GetDeviceBuffer()},
static_cast<EDataType*>(e_device_buf.GetDeviceBuffer()),
M,
N,
K,
StrideA,
StrideB,
std::array<ck::index_t, 2>{StrideD0, StrideD1},
StrideE,
a_element_op,
b_element_op,
cde_element_op);
auto invoker_ptr = device_op_ptr->MakeInvokerPointer();
std::string device_op_name = device_op_ptr->GetTypeString();
if(device_op_ptr->IsSupportedArgument(argument_ptr.get()))
{
// re-init E to zero before profiling a kernel
e_device_buf.SetZero();
float ave_time =
invoker_ptr->Run(argument_ptr.get(), StreamConfig{nullptr, time_kernel});
std::size_t flop = std::size_t(2) * M * N * K;
std::size_t num_btype =
sizeof(ADataType) * M * K + sizeof(BDataType) * K * N + sizeof(EDataType) * M * N;
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << std::setw(10) << ave_time << " ms, " << tflops << " TFlops, "
<< gb_per_sec << " GB/s, " << device_op_name << std::endl;
if(tflops > best_tflops)
{
best_device_op_name = device_op_name;
best_tflops = tflops;
best_ave_time = ave_time;
best_gb_per_sec = gb_per_sec;
}
if(do_verification)
{
e_device_buf.FromDevice(e_m_n_device_result.mData.data());
pass = pass &&
ck::utils::check_err(e_m_n_device_result.mData, e_m_n_host_result.mData);
}
}
else
{
std::cout << device_op_name << " does not support this problem" << std::endl;
}
}
std::cout << "Best Perf: " << best_ave_time << " ms, " << best_tflops << " TFlops, "
<< best_gb_per_sec << " GB/s, " << best_device_op_name << std::endl;
return pass ? 0 : 1;
}
} // namespace profiler
} // namespace ck
profiler/src/profile_gemm_add_add_fastgelu.cpp 0 → 100644
View file @ 56adf7e9
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include "profile_gemm_add_add_fastgelu_impl.hpp"
int profile_gemm_add_add_fastgelu(int argc, char* argv[])
{
enum struct MatrixLayout
{
MK_KN_MN_MN_MN, // 0
MK_NK_MN_MN_MN, // 1
KM_KN_MN_MN_MN, // 2
KM_NK_MN_MN_MN, // 3
MK_KN_NM_MN_MN, // 4
MK_NK_NM_MN_MN, // 5
KM_KN_NM_MN_MN, // 6
KM_NK_NM_MN_MN, // 7
};
enum struct MatrixDataType
{
F32_F32_F32_F32_F32, // 0
F16_F16_F16_F16_F16, // 1
BF16_BF16_BF16_BF16_BF16, // 2
INT8_INT8_INT8_INT8_INT8, // 3
};
if(argc != 16)
{
// clang-format off
printf("arg1: tensor operation (gemm_add_add_fastgelu: GEMM+Add+Add+GeLU)\n");
printf("arg2: data type (0: fp32; 1: fp16; 2: bf16; 3: int8)\n");
printf("arg3: matrix layout (0: E[m, n] = FastGeLU(A[m, k] * B[k, n] + D0[m, n] + D1[m, n]);\n");
printf(" 1: E[m, n] = FastGeLU(A[m, k] * B[n, k] + D0[m, n] + D1[m, n]);\n");
printf(" 2: E[m, n] = FastGeLU(A[k, m] * B[k, n] + D0[m, n] + D1[m, n]);\n");
printf(" 3: E[m, n] = FastGeLU(A[k, m] * B[n, k] + D0[m, n] + D1[m, n]))\n");
printf("arg4: verification (0: no; 1: yes)\n");
printf("arg5: initialization (0: no init; 1: integer value; 2: decimal value)\n");
printf("arg6: print tensor value (0: no; 1: yes)\n");
printf("arg7: time kernel (0=no, 1=yes)\n");
printf("arg8 to 13: M, N, K, StrideA, StrideB, StrideD0, StrideD1, StrideE\n");
// clang-format on
exit(1);
}
const auto data_type = static_cast<MatrixDataType>(std::stoi(argv[2]));
const auto layout = static_cast<MatrixLayout>(std::stoi(argv[3]));
const bool do_verification = std::stoi(argv[4]);
const int init_method = std::stoi(argv[5]);
const bool do_log = std::stoi(argv[6]);
const bool time_kernel = std::stoi(argv[7]);
const int M = std::stoi(argv[8]);
const int N = std::stoi(argv[9]);
const int K = std::stoi(argv[10]);
const int StrideA = std::stoi(argv[11]);
const int StrideB = std::stoi(argv[12]);
const int StrideD0 = std::stoi(argv[13]);
const int StrideD1 = std::stoi(argv[14]);
const int StrideE = std::stoi(argv[15]);
using F16 = ck::half_t;
using F32 = float;
using Row = ck::tensor_layout::gemm::RowMajor;
using Col = ck::tensor_layout::gemm::ColumnMajor;
auto profile = [&](auto a_type,
auto b_type,
auto acc_type,
auto d0_type,
auto d1_type,
auto e_type,
auto a_layout,
auto b_layout,
auto d0_layout,
auto d1_layout,
auto e_layout) {
using ADataType = decltype(a_type);
using BDataType = decltype(b_type);
using AccDataType = decltype(acc_type);
using D0DataType = decltype(d0_type);
using D1DataType = decltype(d1_type);
using EDataType = decltype(e_type);
using ALayout = decltype(a_layout);
using BLayout = decltype(b_layout);
using D0Layout = decltype(d0_layout);
using D1Layout = decltype(d1_layout);
using ELayout = decltype(e_layout);
const int DefaultStrideA = ck::is_same_v<ALayout, Row> ? K : M;
const int DefaultStrideB = ck::is_same_v<BLayout, Row> ? N : K;
const int DefaultStrideD0 = ck::is_same_v<D0Layout, Row> ? N : M;
const int DefaultStrideD1 = ck::is_same_v<D1Layout, Row> ? N : M;
const int DefaultStrideE = ck::is_same_v<ELayout, Row> ? N : M;
return ck::profiler::profile_gemm_add_add_fastgelu_impl<ADataType,
BDataType,
AccDataType,
D0DataType,
D1DataType,
EDataType,
ALayout,
BLayout,
D0Layout,
D1Layout,
ELayout>(
do_verification,
init_method,
do_log,
time_kernel,
M,
N,
K,
(StrideA < 0) ? DefaultStrideA : StrideA,
(StrideB < 0) ? DefaultStrideB : StrideB,
(StrideD0 < 0) ? DefaultStrideD0 : StrideD0,
(StrideD1 < 0) ? DefaultStrideD1 : StrideD1,
(StrideE < 0) ? DefaultStrideE : StrideE);
};
if(data_type == MatrixDataType::F16_F16_F16_F16_F16 && layout == MatrixLayout::MK_KN_MN_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Row{}, Row{}, Row{}, Row{}, Row{});
}
else if(data_type == MatrixDataType::F16_F16_F16_F16_F16 &&
layout == MatrixLayout::MK_NK_MN_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Row{}, Col{}, Row{}, Row{}, Row{});
}
else if(data_type == MatrixDataType::F16_F16_F16_F16_F16 &&
layout == MatrixLayout::KM_KN_MN_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Col{}, Row{}, Row{}, Row{}, Row{});
}
else if(data_type == MatrixDataType::F16_F16_F16_F16_F16 &&
layout == MatrixLayout::KM_NK_MN_MN_MN)
{
return profile(F16{}, F16{}, F32{}, F16{}, F16{}, F16{}, Col{}, Col{}, Row{}, Row{}, Row{});
}
else
{
std::cout << "this data_type & layout is not implemented" << std::endl;
return 0;
}
}
profiler/src/profiler.cpp
View file @ 56adf7e9
......@@ -22,9 +22,39 @@ int profile_convnd_bwd_data(int, char*[], int);
int profile_reduce(int, char*[]);
int profile_conv_bwd_weight(int, char*[]);
int profile_batched_gemm_reduce(int, char*[]);
int profile_gemm_add_add_fastgelu(int, char*[]);
static void print_helper_message()
{
// clang-format off
printf("arg1: tensor operation (gemm: GEMM\n"
" gemm_bias_2d: GEMM+Bias(2D)\n"
" gemm_bias_relu: GEMM+Bias+ReLU\n"
" gemm_bias_relu_add: GEMM+Bias+ReLU+Add\n"
" gemm_reduce: GEMM+Reduce\n"
" grouped_gemm: Grouped GEMM\n"
" conv_fwd: ForwardConvolution\n"
" conv_fwd_bias_relu: ForwardConvolution+Bias+ReLU\n"
" conv_fwd_bias_relu_add: ForwardConvolution+Bias+ReLU+Add\n"
" conv_fwd_bias_relu_atomic_add: ForwardConvolution+Bias+ReLU+AtomicAdd\n"
" conv1d_bwd_data: BackwardConvolution data 1 dim\n"
" conv2d_bwd_data: BackwardConvolution data 2 dim\n"
" conv3d_bwd_data: BackwardConvolution data 3 dim\n"
" reduce: Reduce\n"
" conv2d_bwd_weight: Backward Weight Convolution 2d\n"
" gemm_add_add_fastgelu: GEMM+Add+Add+FastGeLU\n");
// clang-format on
}
int main(int argc, char* argv[])
{
if(argc == 1)
{
print_helper_message();
return 0;
}
if(strcmp(argv[1], "gemm") == 0)
{
return profile_gemm(argc, argv);
......@@ -97,25 +127,14 @@ int main(int argc, char* argv[])
{
return profile_conv_bwd_weight(argc, argv);
}
else
else if(strcmp(argv[1], "gemm_add_add_fastgelu") == 0)
{
// clang-format off
printf("arg1: tensor operation (gemm: GEMM\n"
" gemm_bias_2d: GEMM+Bias(2D)\n"
" gemm_bias_relu: GEMM+Bias+ReLU\n"
" gemm_bias_relu_add: GEMM+Bias+ReLU+Add\n"
" gemm_reduce: GEMM+Reduce\n"
" grouped_gemm: Grouped GEMM\n"
" conv_fwd: ForwardConvolution\n"
" conv_fwd_bias_relu: ForwardConvolution+Bias+ReLU\n"
" conv_fwd_bias_relu_add: ForwardConvolution+Bias+ReLU+Add\n"
" conv_fwd_bias_relu_atomic_add: ForwardConvolution+Bias+ReLU+AtomicAdd\n"
" conv1d_bwd_data: BackwardConvolution data 1 dim\n"
" conv2d_bwd_data: BackwardConvolution data 2 dim\n"
" conv3d_bwd_data: BackwardConvolution data 3 dim\n"
" reduce: Reduce\n"
" conv2d_bwd_weight: Backward Weight Convolution 2d\n");
// clang-format on
return profile_gemm_add_add_fastgelu(argc, argv);
}
else
{
print_helper_message();
return 0;
}
}
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