Skip to content

GitLab

"backend/apps/vscode:/vscode.git/clone" did not exist on "a75a9c953af0c080f3f55ba2f4e1570c17cb815b"
Unverified Commit ccc4a1d3 authored by Chao Liu's avatar Chao Liu Committed by GitHub
Browse files

Merge pull request #8 from ROCmSoftwarePlatform/miopen_downstream_init_integration

parents 3b866461 16effa76
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 1046 additions and 1159 deletions
composable_kernel/include/tensor_description/tensor_adaptor.hpp
View file @ ccc4a1d3
......@@ -2,8 +2,8 @@
#define CK_TENSOR_ADAPTOR_HPP
#include "common_header.hpp"
#include "dynamic_tensor_descriptor.hpp"
#include "dynamic_tensor_descriptor_helper.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
namespace ck {
......@@ -454,9 +454,7 @@ __host__ __device__ constexpr auto make_single_stage_tensor_adaptor(const Transf
remove_cv_t<decltype(top_dim_hidden_ids)>>{transforms};
}
template <typename X,
typename... Xs,
typename std::enable_if<sizeof...(Xs) >= 2, bool>::type = false>
template <typename X, typename... Xs, typename enable_if<sizeof...(Xs) >= 2, bool>::type = false>
__host__ __device__ constexpr auto chain_tensor_adaptors(const X& x, const Xs&... xs)
{
return chain_tensor_adaptors(x, chain_tensor_adaptors(xs...));
......
composable_kernel/include/tensor_description/dynamic_tensor_descriptor.hpp → composable_kernel/include/tensor_description/tensor_descriptor.hpp
View file @ ccc4a1d3
#ifndef CK_DYNAMIC_TENSOR_DESCRIPTOR_HPP
#define CK_DYNAMIC_TENSOR_DESCRIPTOR_HPP
#ifndef CK_TENSOR_DESCRIPTOR_HPP
#define CK_TENSOR_DESCRIPTOR_HPP
#include "common_header.hpp"
#include "dynamic_multi_index_transform.hpp"
#include "multi_index_transform.hpp"
namespace ck {
template <index_t NDimHidden, typename VisibleDimensionIds>
struct DynamicTensorCoordinate;
struct TensorCoordinate;
template <index_t NTransform, index_t NDimVisible, typename UpdateLowerIndexHack>
struct DynamicTensorCoordinateIterator;
struct TensorCoordinateStep;
// Transforms: Tuple<transforms...>
// LowerDimensionIdss : Tuple<Sequence<...>, ...>
......@@ -21,7 +21,7 @@ template <typename Transforms,
typename UpperDimensionIdss,
typename VisibleDimensionIds,
typename ElementSpaceSize>
struct DynamicTensorDescriptor
struct TensorDescriptor
{
// TODO make these private
__host__ __device__ static constexpr index_t GetNumOfTransform() { return Transforms::Size(); }
......@@ -105,16 +105,16 @@ struct DynamicTensorDescriptor
using VisibleIndex = MultiIndex<ndim_visible_>;
using HiddenIndex = MultiIndex<ndim_hidden_>;
using Coordinate = DynamicTensorCoordinate<ndim_hidden_, VisibleDimensionIds>;
using Coordinate = TensorCoordinate<ndim_hidden_, VisibleDimensionIds>;
// may be index_t or Number<>
using ElementSize = remove_cv_t<decltype(InitializeElementSize(Transforms{}))>;
public:
__host__ __device__ constexpr DynamicTensorDescriptor() = default;
__host__ __device__ constexpr TensorDescriptor() = default;
__host__ __device__ constexpr DynamicTensorDescriptor(const Transforms& transforms,
ElementSpaceSize element_space_size)
__host__ __device__ constexpr TensorDescriptor(const Transforms& transforms,
ElementSpaceSize element_space_size)
: transforms_{transforms},
element_size_{InitializeElementSize(transforms)},
element_space_size_{element_space_size}
......@@ -159,7 +159,7 @@ struct DynamicTensorDescriptor
{
static_assert(Idx::Size() == GetNumOfDimension(), "wrong! inconsistent # of dimension");
return make_dynamic_tensor_coordinate(*this, idx).GetOffset();
return make_tensor_coordinate(*this, idx).GetOffset();
}
// TODO make these private
......@@ -196,7 +196,7 @@ struct DynamicTensorDescriptor
__host__ __device__ void Print() const
{
printf("{");
printf("DynamicTensorDescriptor, ");
printf("TensorDescriptor, ");
static_for<0, ntransform_, 1>{}([&](auto i) {
printf("transforms: ");
transforms_[i].Print();
......@@ -217,7 +217,7 @@ struct DynamicTensorDescriptor
};
template <index_t NDimHidden, typename VisibleDimensionIds>
struct DynamicTensorCoordinate
struct TensorCoordinate
{
// TODO make these private
static constexpr index_t ndim_visible_ = VisibleDimensionIds::Size();
......@@ -226,9 +226,9 @@ struct DynamicTensorCoordinate
using VisibleIndex = MultiIndex<ndim_visible_>;
public:
__host__ __device__ constexpr DynamicTensorCoordinate() = default;
__host__ __device__ constexpr TensorCoordinate() = default;
__host__ __device__ constexpr DynamicTensorCoordinate(const HiddenIndex& idx_hidden)
__host__ __device__ constexpr TensorCoordinate(const HiddenIndex& idx_hidden)
: idx_hidden_{idx_hidden}
{
}
......@@ -252,16 +252,16 @@ struct DynamicTensorCoordinate
};
template <index_t NTransform, index_t NDimVisible, typename UpdateLowerIndexHack>
struct DynamicTensorCoordinateIterator
struct TensorCoordinateStep
{
// TODO make these private
using VisibleIndex = MultiIndex<NDimVisible>;
public:
__host__ __device__ constexpr DynamicTensorCoordinateIterator() = default;
__host__ __device__ constexpr TensorCoordinateStep() = default;
__host__ __device__ constexpr DynamicTensorCoordinateIterator(
const VisibleIndex& idx_diff_visible, const MultiIndex<NTransform>& do_transforms)
__host__ __device__ constexpr TensorCoordinateStep(const VisibleIndex& idx_diff_visible,
const MultiIndex<NTransform>& do_transforms)
: idx_diff_visible_{idx_diff_visible}, do_transforms_{do_transforms}
{
}
......@@ -283,7 +283,7 @@ struct DynamicTensorCoordinateIterator
// TODO: How to fix this? It uses an struct instead of lambda because lambda
// doesn't have constructor, and to put it outside the scope where it is used
// (transform_dynamic_tensor_descriptor) because template cannot be defined inside a function
// (transform_tensor_descriptor) because template cannot be defined inside a function
// template
template <typename NewTransforms>
struct lambda_get_up_dim_num
......@@ -301,10 +301,10 @@ template <typename OldTensorDescriptor,
typename NewLowerDimensionOldVisibleIdss,
typename NewUpperDimensionNewVisibleIdss>
__host__ __device__ constexpr auto
transform_dynamic_tensor_descriptor(const OldTensorDescriptor& old_tensor_desc,
const NewTransforms& new_transforms,
NewLowerDimensionOldVisibleIdss,
NewUpperDimensionNewVisibleIdss)
transform_tensor_descriptor(const OldTensorDescriptor& old_tensor_desc,
const NewTransforms& new_transforms,
NewLowerDimensionOldVisibleIdss,
NewUpperDimensionNewVisibleIdss)
{
// sanity check
{
......@@ -376,17 +376,17 @@ transform_dynamic_tensor_descriptor(const OldTensorDescriptor& old_tensor_desc,
const auto element_space_size = old_tensor_desc.GetElementSpaceSize();
return DynamicTensorDescriptor<remove_cv_t<decltype(all_transforms)>,
remove_cv_t<decltype(all_low_dim_hidden_idss)>,
remove_cv_t<decltype(all_up_dim_hidden_idss)>,
remove_cv_t<decltype(new_visible_dim_hidden_ids)>,
remove_cv_t<decltype(element_space_size)>>{all_transforms,
element_space_size};
return TensorDescriptor<remove_cv_t<decltype(all_transforms)>,
remove_cv_t<decltype(all_low_dim_hidden_idss)>,
remove_cv_t<decltype(all_up_dim_hidden_idss)>,
remove_cv_t<decltype(new_visible_dim_hidden_ids)>,
remove_cv_t<decltype(element_space_size)>>{all_transforms,
element_space_size};
}
template <typename TensorDesc, typename VisibleIndex>
__host__ __device__ constexpr auto make_dynamic_tensor_coordinate(const TensorDesc& tensor_desc,
const VisibleIndex& idx_visible)
__host__ __device__ constexpr auto make_tensor_coordinate(const TensorDesc& tensor_desc,
const VisibleIndex& idx_visible)
{
static_assert(TensorDesc::GetNumOfDimension() == VisibleIndex::Size(),
"wrong! # of dimension inconsistent");
......@@ -416,14 +416,15 @@ __host__ __device__ constexpr auto make_dynamic_tensor_coordinate(const TensorDe
set_container_subset(idx_hidden, dims_low, idx_low);
});
return DynamicTensorCoordinate<ndim_hidden, decltype(visible_dim_ids)>{idx_hidden};
return TensorCoordinate<ndim_hidden, decltype(visible_dim_ids)>{idx_hidden};
}
// UpdateLowerIndexHack: Sequence<...>
// HACK: control UpdateLowerIndex
template <typename TensorDesc, typename VisibleIndex, typename UpdateLowerIndexHack>
__host__ __device__ constexpr auto make_dynamic_tensor_coordinate_iterator(
const TensorDesc&, const VisibleIndex& idx_diff_visible, UpdateLowerIndexHack)
__host__ __device__ constexpr auto make_tensor_coordinate_step(const TensorDesc&,
const VisibleIndex& idx_diff_visible,
UpdateLowerIndexHack)
{
static_assert(TensorDesc::GetNumOfDimension() == VisibleIndex::Size(),
"wrong! # of dimension inconsistent");
......@@ -470,23 +471,24 @@ __host__ __device__ constexpr auto make_dynamic_tensor_coordinate_iterator(
set_container_subset(is_non_zero_diff, dims_low, non_zero_diff_pick_low);
});
return DynamicTensorCoordinateIterator<ntransform, ndim_visible, UpdateLowerIndexHack>{
idx_diff_visible, do_transforms};
return TensorCoordinateStep<ntransform, ndim_visible, UpdateLowerIndexHack>{idx_diff_visible,
do_transforms};
}
template <typename TensorDesc, typename VisibleIndex>
__host__ __device__ constexpr auto
make_dynamic_tensor_coordinate_iterator(const TensorDesc&, const VisibleIndex& idx_diff_visible)
__host__ __device__ constexpr auto make_tensor_coordinate_step(const TensorDesc&,
const VisibleIndex& idx_diff_visible)
{
constexpr index_t ntransform = TensorDesc::GetNumOfTransform();
return make_dynamic_tensor_coordinate_iterator(
return make_tensor_coordinate_step(
TensorDesc{}, idx_diff_visible, typename uniform_sequence_gen<ntransform, 0>::type{});
}
template <typename TensorDesc, typename TensorCoord, typename TensorCoordIterator>
__host__ __device__ constexpr void move_dynamic_tensor_coordinate(
const TensorDesc& tensor_desc, TensorCoord& coord, const TensorCoordIterator& coord_iterator)
template <typename TensorDesc, typename TensorCoord, typename TensorCoordStep>
__host__ __device__ constexpr void move_tensor_coordinate(const TensorDesc& tensor_desc,
TensorCoord& coord,
const TensorCoordStep& coord_step)
{
constexpr index_t ndim_hidden = TensorDesc::GetNumOfHiddenDimension();
constexpr index_t ntransform = TensorDesc::GetNumOfTransform();
......@@ -495,9 +497,8 @@ __host__ __device__ constexpr void move_dynamic_tensor_coordinate(
auto idx_diff_hidden = make_zero_multi_index<ndim_hidden>();
// initialize visible index diff
set_container_subset(idx_diff_hidden,
TensorDesc::GetVisibleDimensionIds(),
coord_iterator.GetVisibleIndexDiff());
set_container_subset(
idx_diff_hidden, TensorDesc::GetVisibleDimensionIds(), coord_step.GetVisibleIndexDiff());
// this is what needs to be updated
auto& idx_hidden = coord.GetHiddenIndex();
......@@ -506,13 +507,13 @@ __host__ __device__ constexpr void move_dynamic_tensor_coordinate(
auto idx_hidden_pick_visible =
get_container_subset(idx_hidden, TensorDesc::GetVisibleDimensionIds());
idx_hidden_pick_visible += coord_iterator.GetIndexDiff();
idx_hidden_pick_visible += coord_step.GetIndexDiff();
set_container_subset(idx_hidden, TensorDesc::GetVisibleDimensionIds(), idx_hidden_pick_visible);
// update rest of hidden index
static_for<ntransform - 1, -1, -1>{}([&](auto itran) {
if(coord_iterator.do_transforms_[itran])
if(coord_step.do_transforms_[itran])
{
const auto& tran = tensor_desc.GetTransforms().At(itran);
constexpr auto dims_low = TensorDesc::GetLowerDimensionIdss().At(itran);
......@@ -524,8 +525,8 @@ __host__ __device__ constexpr void move_dynamic_tensor_coordinate(
MultiIndex<dims_low.Size()> idx_diff_low;
// HACK: control UpdateLowerIndex for DynamicMerge using hack
constexpr index_t Hack = decltype(coord_iterator.update_lower_index_hack_)::At(itran);
// HACK: control UpdateLowerIndex for Merge using hack
constexpr index_t Hack = decltype(coord_step.update_lower_index_hack_)::At(itran);
tran.UpdateLowerIndex(idx_diff_low, idx_diff_up, idx_low, idx_up_new, Number<Hack>{});
......@@ -585,11 +586,11 @@ __host__ __device__ constexpr bool coordinate_has_valid_offset(const TensorDesc&
}
template <typename TensorDesc>
using DynamicTensorCoordinate_t = decltype(make_dynamic_tensor_coordinate(
using TensorCoordinate_t = decltype(make_tensor_coordinate(
TensorDesc{}, MultiIndex<remove_cv_t<remove_reference_t<TensorDesc>>::GetNumOfDimension()>{}));
template <typename TensorDesc>
using DynamicTensorCoordinateIterator_t = decltype(make_dynamic_tensor_coordinate_iterator(
using TensorCoordinateStep_t = decltype(make_tensor_coordinate_step(
TensorDesc{}, MultiIndex<remove_cv_t<remove_reference_t<TensorDesc>>::GetNumOfDimension()>{}));
} // namespace ck
......
composable_kernel/include/tensor_description/dynamic_tensor_descriptor_helper.hpp → composable_kernel/include/tensor_description/tensor_descriptor_helper.hpp
View file @ ccc4a1d3
#ifndef CK_DYNAMIC_TENSOR_DESCRIPTOR_HELPER_HPP
#define CK_DYNAMIC_TENSOR_DESCRIPTOR_HELPER_HPP
#ifndef CK_TENSOR_DESCRIPTOR_HELPER_HPP
#define CK_TENSOR_DESCRIPTOR_HELPER_HPP
#include "common_header.hpp"
#include "dynamic_tensor_descriptor.hpp"
#include "dynamic_multi_index_transform_helper.hpp"
#include "tensor_descriptor.hpp"
#include "multi_index_transform_helper.hpp"
namespace ck {
......@@ -37,10 +37,9 @@ __host__ __device__ constexpr auto calculate_element_space_size_impl(const Lengt
template <typename... Lengths,
typename... Strides,
typename std::enable_if<sizeof...(Lengths) == sizeof...(Strides), bool>::type = false>
__host__ __device__ constexpr auto
make_dynamic_naive_tensor_descriptor_v2(const Tuple<Lengths...>& lengths,
const Tuple<Strides...>& strides)
typename enable_if<sizeof...(Lengths) == sizeof...(Strides), bool>::type = false>
__host__ __device__ constexpr auto make_naive_tensor_descriptor_v2(const Tuple<Lengths...>& lengths,
const Tuple<Strides...>& strides)
{
constexpr index_t N = sizeof...(Lengths);
......@@ -75,12 +74,12 @@ make_dynamic_naive_tensor_descriptor_v2(const Tuple<Lengths...>& lengths,
calculate_element_space_size_impl(lengths, strides, Number<0>{}, Number<1>{});
#endif
return DynamicTensorDescriptor<remove_cv_t<decltype(transforms)>,
remove_cv_t<decltype(low_dim_hidden_idss)>,
remove_cv_t<decltype(up_dim_hidden_idss)>,
remove_cv_t<decltype(visible_dim_hidden_ids)>,
remove_cv_t<decltype(element_space_size)>>{transforms,
element_space_size};
return TensorDescriptor<remove_cv_t<decltype(transforms)>,
remove_cv_t<decltype(low_dim_hidden_idss)>,
remove_cv_t<decltype(up_dim_hidden_idss)>,
remove_cv_t<decltype(visible_dim_hidden_ids)>,
remove_cv_t<decltype(element_space_size)>>{transforms,
element_space_size};
}
// Lengths... can be:
......@@ -88,7 +87,7 @@ make_dynamic_naive_tensor_descriptor_v2(const Tuple<Lengths...>& lengths,
// 2) Number<>, which is known at compile-time
template <typename... Lengths>
__host__ __device__ constexpr auto
make_dynamic_naive_tensor_descriptor_packed_v2(const Tuple<Lengths...>& lengths)
make_naive_tensor_descriptor_packed(const Tuple<Lengths...>& lengths)
{
constexpr index_t N = sizeof...(Lengths);
......@@ -103,17 +102,17 @@ make_dynamic_naive_tensor_descriptor_packed_v2(const Tuple<Lengths...>& lengths)
const auto element_space_size = container_reduce(lengths, math::multiplies_v2{}, Number<1>{});
return DynamicTensorDescriptor<remove_cv_t<decltype(transforms)>,
remove_cv_t<decltype(low_dim_hidden_idss)>,
remove_cv_t<decltype(up_dim_hidden_idss)>,
remove_cv_t<decltype(visible_dim_hidden_ids)>,
remove_cv_t<decltype(element_space_size)>>{transforms,
element_space_size};
return TensorDescriptor<remove_cv_t<decltype(transforms)>,
remove_cv_t<decltype(low_dim_hidden_idss)>,
remove_cv_t<decltype(up_dim_hidden_idss)>,
remove_cv_t<decltype(visible_dim_hidden_ids)>,
remove_cv_t<decltype(element_space_size)>>{transforms,
element_space_size};
}
template <typename... Lengths, typename Align>
__host__ __device__ constexpr auto
make_dynamic_naive_tensor_descriptor_aligned_v2(const Tuple<Lengths...>& lengths, Align align)
make_naive_tensor_descriptor_aligned_v2(const Tuple<Lengths...>& lengths, Align align)
{
constexpr auto I1 = Number<1>{};
......@@ -143,7 +142,7 @@ make_dynamic_naive_tensor_descriptor_aligned_v2(const Tuple<Lengths...>& lengths
},
Number<N>{});
return make_dynamic_naive_tensor_descriptor_v2(lengths, strides);
return make_naive_tensor_descriptor_v2(lengths, strides);
}
} // namespace ck
......
composable_kernel/include/tensor_operation/blockwise_gemm_dlops_v2r2.hpp
View file @ ccc4a1d3
......@@ -3,7 +3,7 @@
#include "common_header.hpp"
#include "tensor_adaptor.hpp"
#include "threadwise_dynamic_tensor_slice_transfer.hpp"
#include "threadwise_tensor_slice_transfer.hpp"
#include "threadwise_contraction_dlops.hpp"
namespace ck {
......@@ -22,24 +22,24 @@ namespace ck {
// 2. CThreadBuffer is StaticBuffer
// Also assume:
// M0 = N0 = 2. It will do 2x2 pipelined read and fma (ABBA optimization)
template <index_t BlockSize,
typename FloatA,
typename FloatB,
typename FloatC,
typename AKMBlockDesc,
typename BKNBlockDesc,
index_t M1PerThreadM11,
index_t N1PerThreadN11,
index_t KPerThread,
index_t M1N1ThreadClusterM100,
index_t M1N1ThreadClusterN100,
index_t M1N1ThreadClusterM101,
index_t M1N1ThreadClusterN101,
index_t AThreadCopyScalarPerVector_M11,
index_t BThreadCopyScalarPerVector_N11,
typename std::enable_if<AKMBlockDesc::IsKnownAtCompileTime() &&
BKNBlockDesc::IsKnownAtCompileTime(),
bool>::type = false>
template <
index_t BlockSize,
typename FloatA,
typename FloatB,
typename FloatC,
typename AKMBlockDesc,
typename BKNBlockDesc,
index_t M1PerThreadM11,
index_t N1PerThreadN11,
index_t KPerThread,
index_t M1N1ThreadClusterM100,
index_t M1N1ThreadClusterN100,
index_t M1N1ThreadClusterM101,
index_t M1N1ThreadClusterN101,
index_t AThreadCopyScalarPerVector_M11,
index_t BThreadCopyScalarPerVector_N11,
typename enable_if<AKMBlockDesc::IsKnownAtCompileTime() && BKNBlockDesc::IsKnownAtCompileTime(),
bool>::type = false>
struct BlockwiseGemmDlops_km_kn_m0m1n0n1_v2r2_pipeline_2x2
{
using AIndex = MultiIndex<3>;
......@@ -71,9 +71,9 @@ struct BlockwiseGemmDlops_km_kn_m0m1n0n1_v2r2_pipeline_2x2
static constexpr index_t N0 = N / N1;
__host__ __device__ static constexpr auto
MakeAKM0M1BlockDescriptor(const AKMBlockDesc& a_k_m_block_desc)
MakeAKM0M1BlockDescriptor(const AKMBlockDesc& /* a_k_m_block_desc */)
{
const auto a_k_m0_m1_block_desc = transform_dynamic_tensor_descriptor(
const auto a_k_m0_m1_block_desc = transform_tensor_descriptor(
AKMBlockDesc{},
make_tuple(make_pass_through_transform(Number<K>{}),
make_unmerge_transform(make_tuple(Number<M0>{}, Number<M1>{}))),
......@@ -84,9 +84,9 @@ struct BlockwiseGemmDlops_km_kn_m0m1n0n1_v2r2_pipeline_2x2
}
__host__ __device__ static constexpr auto
MakeBKN0N1BlockDescriptor(const BKNBlockDesc& b_k_n_block_desc)
MakeBKN0N1BlockDescriptor(const BKNBlockDesc& /* b_k_n_block_desc */)
{
const auto b_k_n0_n1_block_desc = transform_dynamic_tensor_descriptor(
const auto b_k_n0_n1_block_desc = transform_tensor_descriptor(
BKNBlockDesc{},
make_tuple(make_pass_through_transform(Number<K>{}),
make_unmerge_transform(make_tuple(Number<N0>{}, Number<N1>{}))),
......@@ -194,7 +194,7 @@ struct BlockwiseGemmDlops_km_kn_m0m1n0n1_v2r2_pipeline_2x2
typename ABlockBuffer,
typename BBlockBuffer,
typename CThreadBuffer>
__device__ void Run(const CM0M1N0N1ThreadDesc& c_m0_m1_n0_n1_thread_desc,
__device__ void Run(const CM0M1N0N1ThreadDesc& /* c_m0_m1_n0_n1_thread_desc */,
const ABlockBuffer& a_block_buf,
const BBlockBuffer& b_block_buf,
CThreadBuffer& c_thread_buf) const
......@@ -357,34 +357,32 @@ struct BlockwiseGemmDlops_km_kn_m0m1n0n1_v2r2_pipeline_2x2
private:
// A[K, M0, M1]
static constexpr auto a_k_m0_m1_thread_desc_ = make_dynamic_naive_tensor_descriptor_packed_v2(
static constexpr auto a_k_m0_m1_thread_desc_ = make_naive_tensor_descriptor_packed(
make_tuple(Number<KPerThread>{}, Number<M0>{}, Number<M1PerThreadM11>{}));
// B[K, N0, N1]
static constexpr auto b_k_n0_n1_thread_desc_ = make_dynamic_naive_tensor_descriptor_packed_v2(
static constexpr auto b_k_n0_n1_thread_desc_ = make_naive_tensor_descriptor_packed(
make_tuple(Number<KPerThread>{}, Number<N0>{}, Number<N1PerThreadN11>{}));
using AThreadCopy =
ThreadwiseDynamicTensorSliceTransfer_v4<FloatA,
FloatA,
decltype(a_k_m0_m1_block_desc_),
decltype(a_k_m0_m1_thread_desc_),
Sequence<KPerThread, 1, M1PerThreadM11>,
Sequence<0, 1, 2>,
2,
AThreadCopyScalarPerVector_M11,
1>;
using BThreadCopy =
ThreadwiseDynamicTensorSliceTransfer_v4<FloatB,
FloatB,
decltype(b_k_n0_n1_block_desc_),
decltype(b_k_n0_n1_thread_desc_),
Sequence<KPerThread, 1, N1PerThreadN11>,
Sequence<0, 1, 2>,
2,
BThreadCopyScalarPerVector_N11,
1>;
using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatA,
FloatA,
decltype(a_k_m0_m1_block_desc_),
decltype(a_k_m0_m1_thread_desc_),
Sequence<KPerThread, 1, M1PerThreadM11>,
Sequence<0, 1, 2>,
2,
AThreadCopyScalarPerVector_M11,
1>;
using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatB,
FloatB,
decltype(b_k_n0_n1_block_desc_),
decltype(b_k_n0_n1_thread_desc_),
Sequence<KPerThread, 1, N1PerThreadN11>,
Sequence<0, 1, 2>,
2,
BThreadCopyScalarPerVector_N11,
1>;
CIndex c_thread_origin_data_idx_;
......
composable_kernel/include/tensor_operation/blockwise_gemm_dlops_v2r3.hpp
View file @ ccc4a1d3
......@@ -3,7 +3,7 @@
#include "common_header.hpp"
#include "tensor_adaptor.hpp"
#include "threadwise_dynamic_tensor_slice_transfer_v2.hpp"
#include "threadwise_tensor_slice_transfer_v2.hpp"
#include "threadwise_contraction_dlops.hpp"
namespace ck {
......@@ -38,9 +38,9 @@ template <index_t BlockSize,
// BM10BN10ThreadClusterBN101, ...>
index_t AThreadCopyScalarPerVector_BM11,
index_t BThreadCopyScalarPerVector_BN11,
typename std::enable_if<ABlockDesc_BK0_BM_BK1::IsKnownAtCompileTime() &&
BBlockDesc_BK0_BN_BK1::IsKnownAtCompileTime(),
bool>::type = false>
typename enable_if<ABlockDesc_BK0_BM_BK1::IsKnownAtCompileTime() &&
BBlockDesc_BK0_BN_BK1::IsKnownAtCompileTime(),
bool>::type = false>
struct BlockwiseGemmDlops_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_BM0_2_BN0_2
{
using AIndex = MultiIndex<3>;
......@@ -75,7 +75,7 @@ struct BlockwiseGemmDlops_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_B
__host__ __device__ static constexpr auto
MakeABlockDescriptor_BK0_BM0_BM1_BK1(const ABlockDesc_BK0_BM_BK1& a_block_desc_bk0_bm_bk1)
{
const auto a_block_bk0_bm0_bm1_bk1 = transform_dynamic_tensor_descriptor(
const auto a_block_bk0_bm0_bm1_bk1 = transform_tensor_descriptor(
a_block_desc_bk0_bm_bk1,
make_tuple(make_pass_through_transform(Number<BK0>{}),
make_unmerge_transform(make_tuple(Number<BM0>{}, Number<BM1>{})),
......@@ -89,7 +89,7 @@ struct BlockwiseGemmDlops_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_B
__host__ __device__ static constexpr auto
MakeBBlockDescriptor_BK0_BN0_BN1_BK1(const BBlockDesc_BK0_BN_BK1& b_block_desc_bk0_bn_bk1)
{
const auto b_block_desc_bk0_bn0_bn1_bk1 = transform_dynamic_tensor_descriptor(
const auto b_block_desc_bk0_bn0_bn1_bk1 = transform_tensor_descriptor(
b_block_desc_bk0_bn_bk1,
make_tuple(make_pass_through_transform(Number<BK0>{}),
make_unmerge_transform(make_tuple(Number<BN0>{}, Number<BN1>{})),
......@@ -372,15 +372,15 @@ struct BlockwiseGemmDlops_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_B
private:
// A[BK0, BM0, BM1, BK1]
static constexpr auto a_thread_desc_bk0_bm0_bm1_bk1_ =
make_dynamic_naive_tensor_descriptor_packed_v2(make_tuple(
make_naive_tensor_descriptor_packed(make_tuple(
Number<BK0PerThread>{}, Number<BM0>{}, Number<BM1PerThreadBM11>{}, Number<BK1>{}));
// B[BK0, BN0, BN1, BK1]
static constexpr auto b_thread_desc_bk0_bn0_bn1_bk1_ =
make_dynamic_naive_tensor_descriptor_packed_v2(make_tuple(
make_naive_tensor_descriptor_packed(make_tuple(
Number<BK0PerThread>{}, Number<BN0>{}, Number<BN1PerThreadBN11>{}, Number<BK1>{}));
using AThreadCopy = ThreadwiseDynamicTensorSliceTransfer_v4r1<
using AThreadCopy = ThreadwiseTensorSliceTransfer_v4r1<
FloatA,
FloatA,
decltype(a_block_desc_bk0_bm0_bm1_bk1_),
......@@ -390,7 +390,7 @@ struct BlockwiseGemmDlops_A_BK0_BM_BK1_B_BK0_BN_BK1_C_BM0_BM1_BN0_BN1_pipeline_B
Sequence<1, 1, BM1PerThreadBM11, BK1>, // SrcVectorTensorLengths
Sequence<0, 1, 2, 3>>; // SrcVectorTensorContiguousDimOrder
using BThreadCopy = ThreadwiseDynamicTensorSliceTransfer_v4r1<
using BThreadCopy = ThreadwiseTensorSliceTransfer_v4r1<
FloatB,
FloatB,
decltype(b_block_desc_bk0_bn0_bn1_bk1_),
......
composable_kernel/include/tensor_operation/blockwise_gemm_dlops_v3.hpp
View file @ ccc4a1d3
......@@ -31,25 +31,24 @@ struct BlockwiseGemmDlops_km_kn_m0m1n0n1_v3
// HACK: fix this @Jing Zhang
static constexpr index_t KPerThreadSubC = 4;
static constexpr auto a_thread_mtx_ = make_dynamic_naive_tensor_descriptor_packed_v2(
static constexpr auto a_thread_mtx_ = make_naive_tensor_descriptor_packed(
make_tuple(Number<EPerThreadLoop>{}, Number<KPerThreadSubC>{}));
static constexpr auto b_thread_mtx_ = make_dynamic_naive_tensor_descriptor_packed_v2(make_tuple(
static constexpr auto b_thread_mtx_ = make_naive_tensor_descriptor_packed(make_tuple(
Number<EPerThreadLoop>{}, Number<1>{}, Number<HPerThread>{}, Number<WPerThread>{}));
static constexpr auto c_thread_mtx_ = make_dynamic_naive_tensor_descriptor_packed_v2(make_tuple(
static constexpr auto c_thread_mtx_ = make_naive_tensor_descriptor_packed(make_tuple(
Number<KPerThreadSubC>{}, Number<1>{}, Number<HPerThread>{}, Number<WPerThread>{}));
using AThreadCopy =
ThreadwiseDynamicTensorSliceTransfer_v4<FloatA,
FloatA,
BlockMatrixA,
decltype(a_thread_mtx_),
Sequence<EPerThreadLoop, KPerThreadSubC>,
Sequence<0, 1>,
1,
ThreadGemmADataPerRead_K,
1>;
using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatA,
FloatA,
BlockMatrixA,
decltype(a_thread_mtx_),
Sequence<EPerThreadLoop, KPerThreadSubC>,
Sequence<0, 1>,
1,
ThreadGemmADataPerRead_K,
1>;
__device__ BlockwiseGemmDlops_km_kn_m0m1n0n1_v3()
: c_thread_begin_mtx_idx_{GetBeginOfThreadMatrixC(get_thread_local_1d_id())},
......@@ -69,7 +68,6 @@ struct BlockwiseGemmDlops_km_kn_m0m1n0n1_v3
"wrong! K dimension not consistent\n");
constexpr index_t K = BlockMatrixA{}.GetLength(I1); // A is transposed
constexpr index_t N = BlockMatrixB{}.GetLength(I1);
constexpr index_t H = BlockMatrixB{}.GetLength(I2);
constexpr index_t W = BlockMatrixB{}.GetLength(I3);
......@@ -121,9 +119,6 @@ struct BlockwiseGemmDlops_km_kn_m0m1n0n1_v3
"wrong! inconsistent type");
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto a_block_mtx = BlockMatrixA{};
......@@ -138,7 +133,7 @@ struct BlockwiseGemmDlops_km_kn_m0m1n0n1_v3
static_assert(WPerThread % WoPerThreadSubC == 0, "");
// thread A buffer for GEMM
StaticBuffer<AddressSpaceEnum_t::Vgpr, FloatA, a_thread_mtx_.GetElementSpaceSize()>
StaticBuffer<AddressSpaceEnum_t::Vgpr, FloatA, a_thread_mtx_.GetElementSpaceSize(), true>
a_thread_buf;
constexpr auto threadwise_gemm = ThreadwiseGemmDlops_km_kn_mn_v3<FloatA,
......
composable_kernel/include/tensor_operation/blockwise_gemm_xdlops.hpp
View file @ ccc4a1d3
......@@ -2,7 +2,7 @@
#define CK_BLOCKWISE_GEMM_XDLOPS_HPP
#include "common_header.hpp"
#include "threadwise_dynamic_tensor_slice_transfer.hpp"
#include "threadwise_tensor_slice_transfer.hpp"
#include "xdlops_gemm.hpp"
namespace ck {
......@@ -52,7 +52,6 @@ struct BlockwiseGemmXdlops_km_kn_m0m1m2n_v1
const index_t waveId = thread_id / WaveSize;
const index_t laneId = thread_id % WaveSize;
const index_t waveId_m = waveId / NWaves;
const index_t waveId_n = waveId % NWaves;
if constexpr(xdlops_gemm.IsKReduction)
{
......@@ -73,7 +72,6 @@ struct BlockwiseGemmXdlops_km_kn_m0m1m2n_v1
const index_t thread_id = get_thread_local_1d_id();
const index_t waveId = thread_id / WaveSize;
const index_t laneId = thread_id % WaveSize;
const index_t waveId_m = waveId / NWaves;
const index_t waveId_n = waveId % NWaves;
if constexpr(xdlops_gemm.IsKReduction)
......@@ -193,35 +191,35 @@ struct BlockwiseGemmXdlops_km_kn_m0m1m2n_v1
private:
// A[K, M]
static constexpr auto a_thread_desc_ = make_dynamic_naive_tensor_descriptor_packed_v2(
make_tuple(I1, Number<MRepeat>{}, I1, Number<K1>{}));
static constexpr auto a_thread_desc_ =
make_naive_tensor_descriptor_packed(make_tuple(I1, Number<MRepeat>{}, I1, Number<K1>{}));
// B[K, N]
static constexpr auto b_thread_desc_ = make_dynamic_naive_tensor_descriptor_packed_v2(
make_tuple(I1, Number<NRepeat>{}, I1, Number<K1>{}));
static constexpr auto c_thread_desc_ = make_dynamic_naive_tensor_descriptor_packed_v2(
make_tuple(Number<MRepeat>{}, Number<NRepeat>{}));
using AThreadCopy = ThreadwiseDynamicTensorSliceTransfer_v4<FloatAB,
FloatAB,
ABlockDesc,
decltype(a_thread_desc_),
Sequence<1, MRepeat, 1, K1>,
Sequence<0, 1, 2, 3>,
3,
K1,
1>;
using BThreadCopy = ThreadwiseDynamicTensorSliceTransfer_v4<FloatAB,
FloatAB,
BBlockDesc,
decltype(b_thread_desc_),
Sequence<1, NRepeat, 1, K1>,
Sequence<0, 1, 2, 3>,
3,
K1,
1>;
static constexpr auto b_thread_desc_ =
make_naive_tensor_descriptor_packed(make_tuple(I1, Number<NRepeat>{}, I1, Number<K1>{}));
static constexpr auto c_thread_desc_ =
make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{}, Number<NRepeat>{}));
using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
FloatAB,
ABlockDesc,
decltype(a_thread_desc_),
Sequence<1, MRepeat, 1, K1>,
Sequence<0, 1, 2, 3>,
3,
K1,
1>;
using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
FloatAB,
BBlockDesc,
decltype(b_thread_desc_),
Sequence<1, NRepeat, 1, K1>,
Sequence<0, 1, 2, 3>,
3,
K1,
1>;
AThreadCopy a_thread_copy_;
BThreadCopy b_thread_copy_;
......@@ -272,7 +270,6 @@ struct BlockwiseGemmXdlops_km_kn_m0m1m2n_v1_2x2pipeline
const index_t waveId = thread_id / WaveSize;
const index_t laneId = thread_id % WaveSize;
const index_t waveId_m = waveId / NWaves;
const index_t waveId_n = waveId % NWaves;
if constexpr(xdlops_gemm.IsKReduction)
{
......@@ -293,7 +290,6 @@ struct BlockwiseGemmXdlops_km_kn_m0m1m2n_v1_2x2pipeline
const index_t thread_id = get_thread_local_1d_id();
const index_t waveId = thread_id / WaveSize;
const index_t laneId = thread_id % WaveSize;
const index_t waveId_m = waveId / NWaves;
const index_t waveId_n = waveId % NWaves;
if constexpr(xdlops_gemm.IsKReduction)
......@@ -490,35 +486,35 @@ struct BlockwiseGemmXdlops_km_kn_m0m1m2n_v1_2x2pipeline
private:
// A[K, M]
static constexpr auto a_thread_desc_ = make_dynamic_naive_tensor_descriptor_packed_v2(
make_tuple(I1, Number<MRepeat>{}, I1, Number<K1>{}));
static constexpr auto a_thread_desc_ =
make_naive_tensor_descriptor_packed(make_tuple(I1, Number<MRepeat>{}, I1, Number<K1>{}));
// B[K, N]
static constexpr auto b_thread_desc_ = make_dynamic_naive_tensor_descriptor_packed_v2(
make_tuple(I1, Number<NRepeat>{}, I1, Number<K1>{}));
static constexpr auto c_thread_desc_ = make_dynamic_naive_tensor_descriptor_packed_v2(
make_tuple(Number<MRepeat>{}, Number<NRepeat>{}));
using AThreadCopy = ThreadwiseDynamicTensorSliceTransfer_v4<FloatAB,
FloatAB,
ABlockDesc,
decltype(a_thread_desc_),
Sequence<1, 1, 1, K1>,
Sequence<0, 1, 2, 3>,
3,
1, // K1,
1>;
using BThreadCopy = ThreadwiseDynamicTensorSliceTransfer_v4<FloatAB,
FloatAB,
BBlockDesc,
decltype(b_thread_desc_),
Sequence<1, 1, 1, K1>,
Sequence<0, 1, 2, 3>,
3,
1, // K1,
1>;
static constexpr auto b_thread_desc_ =
make_naive_tensor_descriptor_packed(make_tuple(I1, Number<NRepeat>{}, I1, Number<K1>{}));
static constexpr auto c_thread_desc_ =
make_naive_tensor_descriptor_packed(make_tuple(Number<MRepeat>{}, Number<NRepeat>{}));
using AThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
FloatAB,
ABlockDesc,
decltype(a_thread_desc_),
Sequence<1, 1, 1, K1>,
Sequence<0, 1, 2, 3>,
3,
1, // K1,
1>;
using BThreadCopy = ThreadwiseTensorSliceTransfer_v4<FloatAB,
FloatAB,
BBlockDesc,
decltype(b_thread_desc_),
Sequence<1, 1, 1, K1>,
Sequence<0, 1, 2, 3>,
3,
1, // K1,
1>;
AThreadCopy a_thread_copy_;
BThreadCopy b_thread_copy_;
......
composable_kernel/include/tensor_operation/blockwise_dynamic_tensor_slice_transfer.hpp → composable_kernel/include/tensor_operation/blockwise_tensor_slice_transfer.hpp
View file @ ccc4a1d3
#ifndef CK_BLOCKWISE_DYNAMIC_TENSOR_SLICE_TRANSFER_HPP
#define CK_BLOCKWISE_DYNAMIC_TENSOR_SLICE_TRANSFER_HPP
#ifndef CK_BLOCKWISE_TENSOR_SLICE_TRANSFER_HPP
#define CK_BLOCKWISE_TENSOR_SLICE_TRANSFER_HPP
#include "common_header.hpp"
#include "dynamic_tensor_descriptor.hpp"
#include "dynamic_tensor_descriptor_helper.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "cluster_descriptor.hpp"
#include "threadwise_dynamic_tensor_slice_transfer.hpp"
#include "threadwise_tensor_slice_transfer.hpp"
namespace ck {
// this version does following things to avoid scratch memory issue
// 1. Use StaticallyIndexedArray instead of C array for thread buffer
// 2. ThreadwiseDynamicTensorSliceTransfer_v3 does not keep reference to tensor descriptor
// 3. ThreadwiseDynamicTensorSliceTransfer_v3::Run() does not construct new tensor coordinate
// 2. ThreadwiseTensorSliceTransfer_v3 does not keep reference to tensor descriptor
// 3. ThreadwiseTensorSliceTransfer_v3::Run() does not construct new tensor coordinate
template <index_t BlockSize,
InMemoryDataOperationEnum_t DstInMemOp,
typename BlockSliceLengths,
......@@ -33,16 +33,16 @@ template <index_t BlockSize,
index_t DstScalarStrideInVector,
bool ThreadTransferSrcResetCoordinateAfterRun,
bool ThreadTransferDstResetCoordinateAfterRun>
struct BlockwiseDynamicTensorSliceTransfer_v4
struct BlockwiseTensorSliceTransfer_v4
{
static constexpr index_t nDim = remove_reference_t<SrcDesc>::GetNumOfDimension();
using Index = MultiIndex<nDim>;
__device__ constexpr BlockwiseDynamicTensorSliceTransfer_v4(const SrcDesc& src_desc,
const Index& src_block_slice_origin,
const DstDesc& dst_desc,
const Index& dst_block_slice_origin)
__device__ constexpr BlockwiseTensorSliceTransfer_v4(const SrcDesc& src_desc,
const Index& src_block_slice_origin,
const DstDesc& dst_desc,
const Index& dst_block_slice_origin)
: threadwise_transfer_(
src_desc, make_zero_multi_index<nDim>(), dst_desc, make_zero_multi_index<nDim>())
......@@ -77,15 +77,14 @@ struct BlockwiseDynamicTensorSliceTransfer_v4
}
}
template <typename SrcBuffer, typename SrcIteratorHacks>
__device__ void RunRead(const SrcDesc& src_desc,
const SrcBuffer& src_buf,
const SrcIteratorHacks& src_iterator_hacks)
template <typename SrcBuffer, typename SrcStepHacks>
__device__ void
RunRead(const SrcDesc& src_desc, const SrcBuffer& src_buf, const SrcStepHacks& src_step_hacks)
{
if(BlockSize == thread_cluster_desc_.GetElementSize() or
get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.RunRead(src_desc, src_buf, src_iterator_hacks);
threadwise_transfer_.RunRead(src_desc, src_buf, src_step_hacks);
}
}
......@@ -118,18 +117,18 @@ struct BlockwiseDynamicTensorSliceTransfer_v4
}
}
// SrcMoveSliceWindowIteratorHack to control index calculation move slice window
template <typename SrcMoveSliceWindowIteratorHack>
// SrcMoveSliceWindowStepHack to control index calculation move slice window
template <typename SrcMoveSliceWindowStepHack>
__device__ void
MoveSrcSliceWindow(const SrcDesc& src_desc,
const Index& step,
const SrcMoveSliceWindowIteratorHack& src_move_slice_window_iterator_hack)
const SrcMoveSliceWindowStepHack& src_move_slice_window_step_hack)
{
if(BlockSize == thread_cluster_desc_.GetElementSize() or
get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.MoveSrcSliceWindow(
src_desc, step, src_move_slice_window_iterator_hack);
src_desc, step, src_move_slice_window_step_hack);
}
}
......@@ -147,22 +146,22 @@ struct BlockwiseDynamicTensorSliceTransfer_v4
make_cluster_descriptor_v2(ThreadClusterLengths{}, ThreadClusterArrangeOrder{});
using ThreadwiseTransfer =
ThreadwiseDynamicTensorSliceTransfer_v3<ThreadSliceLengths,
DstInMemOp,
SrcData,
DstData,
SrcDesc,
DstDesc,
SrcDimAccessOrder,
DstDimAccessOrder,
SrcVectorDim,
DstVectorDim,
SrcScalarPerVector,
DstScalarPerVector,
SrcScalarStrideInVector,
DstScalarStrideInVector,
ThreadTransferSrcResetCoordinateAfterRun,
ThreadTransferDstResetCoordinateAfterRun>;
ThreadwiseTensorSliceTransfer_v3<ThreadSliceLengths,
DstInMemOp,
SrcData,
DstData,
SrcDesc,
DstDesc,
SrcDimAccessOrder,
DstDimAccessOrder,
SrcVectorDim,
DstVectorDim,
SrcScalarPerVector,
DstScalarPerVector,
SrcScalarStrideInVector,
DstScalarStrideInVector,
ThreadTransferSrcResetCoordinateAfterRun,
ThreadTransferDstResetCoordinateAfterRun>;
ThreadwiseTransfer threadwise_transfer_;
};
......
composable_kernel/include/tensor_operation/blockwise_dynamic_tensor_slice_transfer_v2.hpp → composable_kernel/include/tensor_operation/blockwise_tensor_slice_transfer_v2.hpp
View file @ ccc4a1d3
#ifndef CK_BLOCKWISE_DYNAMIC_TENSOR_SLICE_TRANSFER_V2_HPP
#define CK_BLOCKWISE_DYNAMIC_TENSOR_SLICE_TRANSFER_V2_HPP
#ifndef CK_BLOCKWISE_TENSOR_SLICE_TRANSFER_V2_HPP
#define CK_BLOCKWISE_TENSOR_SLICE_TRANSFER_V2_HPP
#include "common_header.hpp"
#include "dynamic_tensor_descriptor.hpp"
#include "dynamic_tensor_descriptor_helper.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "cluster_descriptor.hpp"
#include "threadwise_dynamic_tensor_slice_transfer_v2.hpp"
#include "threadwise_tensor_slice_transfer_v2.hpp"
namespace ck {
// this version does following things to avoid scratch memory issue
// 1. Use StaticallyIndexedArray instead of C array for thread buffer
// 2. ThreadwiseDynamicTensorSliceTransfer_v3 does not keep reference to tensor descriptor
// 3. ThreadwiseDynamicTensorSliceTransfer_v3::Run() does not construct new tensor coordinate
// 2. ThreadwiseTensorSliceTransfer_v3 does not keep reference to tensor descriptor
// 3. ThreadwiseTensorSliceTransfer_v3::Run() does not construct new tensor coordinate
template <index_t BlockSize,
InMemoryDataOperationEnum_t DstInMemOp,
typename BlockSliceLengths,
......@@ -31,17 +31,16 @@ template <index_t BlockSize,
typename DstVectorTensorContiguousDimOrder,
bool ThreadTransferSrcResetCoordinateAfterRun,
bool ThreadTransferDstResetCoordinateAfterRun>
struct BlockwiseDynamicTensorSliceTransfer_v4r1
struct BlockwiseTensorSliceTransfer_v4r1
{
static constexpr index_t nDim = remove_reference_t<SrcDesc>::GetNumOfDimension();
using Index = MultiIndex<nDim>;
__device__ constexpr BlockwiseDynamicTensorSliceTransfer_v4r1(
const SrcDesc& src_desc,
const Index& src_block_slice_origin,
const DstDesc& dst_desc,
const Index& dst_block_slice_origin)
__device__ constexpr BlockwiseTensorSliceTransfer_v4r1(const SrcDesc& src_desc,
const Index& src_block_slice_origin,
const DstDesc& dst_desc,
const Index& dst_block_slice_origin)
: threadwise_transfer_(
src_desc, make_zero_multi_index<nDim>(), dst_desc, make_zero_multi_index<nDim>())
......@@ -76,15 +75,14 @@ struct BlockwiseDynamicTensorSliceTransfer_v4r1
}
}
template <typename SrcBuffer, typename SrcIteratorHacks>
__device__ void RunRead(const SrcDesc& src_desc,
const SrcBuffer& src_buf,
const SrcIteratorHacks& src_iterator_hacks)
template <typename SrcBuffer, typename SrcStepHacks>
__device__ void
RunRead(const SrcDesc& src_desc, const SrcBuffer& src_buf, const SrcStepHacks& src_step_hacks)
{
if(BlockSize == thread_cluster_desc_.GetElementSize() or
get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.RunRead(src_desc, src_buf, src_iterator_hacks);
threadwise_transfer_.RunRead(src_desc, src_buf, src_step_hacks);
}
}
......@@ -107,18 +105,18 @@ struct BlockwiseDynamicTensorSliceTransfer_v4r1
}
}
// SrcMoveSliceWindowIteratorHack to control index calculation move slice window
template <typename SrcMoveSliceWindowIteratorHack>
// SrcMoveSliceWindowStepHack to control index calculation move slice window
template <typename SrcMoveSliceWindowStepHack>
__device__ void
MoveSrcSliceWindow(const SrcDesc& src_desc,
const Index& step,
const SrcMoveSliceWindowIteratorHack& src_move_slice_window_iterator_hack)
const SrcMoveSliceWindowStepHack& src_move_slice_window_step_hack)
{
if(BlockSize == thread_cluster_desc_.GetElementSize() or
get_thread_local_1d_id() < thread_cluster_desc_.GetElementSize())
{
threadwise_transfer_.MoveSrcSliceWindow(
src_desc, step, src_move_slice_window_iterator_hack);
src_desc, step, src_move_slice_window_step_hack);
}
}
......@@ -136,20 +134,20 @@ struct BlockwiseDynamicTensorSliceTransfer_v4r1
make_cluster_descriptor_v2(ThreadClusterLengths{}, ThreadClusterArrangeOrder{});
using ThreadwiseTransfer =
ThreadwiseDynamicTensorSliceTransfer_v3r1<ThreadSliceLengths,
DstInMemOp,
SrcData,
DstData,
SrcDesc,
DstDesc,
SrcDimAccessOrder,
DstDimAccessOrder,
SrcVectorTensorLengths,
DstVectorTensorLengths,
SrcVectorTensorContiguousDimOrder,
DstVectorTensorContiguousDimOrder,
ThreadTransferSrcResetCoordinateAfterRun,
ThreadTransferDstResetCoordinateAfterRun>;
ThreadwiseTensorSliceTransfer_v3r1<ThreadSliceLengths,
DstInMemOp,
SrcData,
DstData,
SrcDesc,
DstDesc,
SrcDimAccessOrder,
DstDimAccessOrder,
SrcVectorTensorLengths,
DstVectorTensorLengths,
SrcVectorTensorContiguousDimOrder,
DstVectorTensorContiguousDimOrder,
ThreadTransferSrcResetCoordinateAfterRun,
ThreadTransferDstResetCoordinateAfterRun>;
ThreadwiseTransfer threadwise_transfer_;
};
......
composable_kernel/include/tensor_operation/gridwise_dynamic_contraction_dlops_v1r2.hpp → composable_kernel/include/tensor_operation/gridwise_contraction_dlops_v1r2.hpp
View file @ ccc4a1d3
#ifndef CK_GRIDWISE_DYNAMIC_CONTRACTION_DLOPS_V1R2_HPP
#define CK_GRIDWISE_DYNAMIC_CONTRACTION_DLOPS_V1R2_HPP
#ifndef CK_GRIDWISE_CONTRACTION_DLOPS_V1R2_HPP
#define CK_GRIDWISE_CONTRACTION_DLOPS_V1R2_HPP
#include "common_header.hpp"
#include "dynamic_multi_index_transform_helper.hpp"
#include "dynamic_tensor_descriptor.hpp"
#include "dynamic_tensor_descriptor_helper.hpp"
#include "multi_index_transform_helper.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "blockwise_gemm_dlops_v2r3.hpp"
#include "blockwise_dynamic_tensor_slice_transfer_v2.hpp"
#include "threadwise_dynamic_tensor_slice_transfer.hpp"
#include "threadwise_dynamic_tensor_slice_set.hpp"
#include "blockwise_tensor_slice_transfer_v2.hpp"
#include "threadwise_tensor_slice_transfer.hpp"
#include "threadwise_tensor_slice_set.hpp"
namespace ck {
......@@ -25,7 +25,7 @@ __global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
kernel_dynamic_contraction_dlops_v1r2(
kernel_contraction_dlops_v1r2(
const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid,
......@@ -84,12 +84,12 @@ template <index_t BlockSize,
typename CThreadTransferSrcDstAccessOrder,
index_t CThreadTransferSrcDstVectorDim,
index_t CThreadTransferDstScalarPerVector,
typename AGridIteratorHacks,
typename BGridIteratorHacks,
typename CGridIteratorHacks,
typename AGridMoveSliceWindowIteratorHacks,
typename BGridMoveSliceWindowIteratorHacks>
struct GridwiseDynamicContractionDlops_A_GK0_GM0_GM1_GK1_B_GK0_GN0_GN1_GK1_C_GM0_GM1_GN0_GN1
typename AGridStepHacks,
typename BGridStepHacks,
typename CGridStepHacks,
typename AGridMoveSliceWindowStepHacks,
typename BGridMoveSliceWindowStepHacks>
struct GridwiseContractionDlops_A_GK0_GM0_GM1_GK1_B_GK0_GN0_GN1_GK1_C_GM0_GM1_GN0_GN1
{
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
......@@ -110,17 +110,15 @@ struct GridwiseDynamicContractionDlops_A_GK0_GM0_GM1_GK1_B_GK0_GN0_GN1_GK1_C_GM0
// A matrix in LDS memory, dst of blockwise copy
// be careful of LDS alignment
constexpr auto a_block_desc_gk0_gm0_gm10_gm11_gk1 =
make_dynamic_naive_tensor_descriptor_aligned_v2(
make_tuple(Number<GK0PerBlock>{}, GM0, I1, Number<GM1PerBlockGM11>{}, GK1),
max_lds_align);
constexpr auto a_block_desc_gk0_gm0_gm10_gm11_gk1 = make_naive_tensor_descriptor_aligned_v2(
make_tuple(Number<GK0PerBlock>{}, GM0, I1, Number<GM1PerBlockGM11>{}, GK1),
max_lds_align);
// B matrix in LDS memory, dst of blockwise copy
// be careful of LDS alignment
constexpr auto b_block_desc_gk0_gn0_gn10_gn11_gk1 =
make_dynamic_naive_tensor_descriptor_aligned_v2(
make_tuple(Number<GK0PerBlock>{}, GN0, I1, Number<GN1PerBlockGN11>{}, GK1),
max_lds_align);
constexpr auto b_block_desc_gk0_gn0_gn10_gn11_gk1 = make_naive_tensor_descriptor_aligned_v2(
make_tuple(Number<GK0PerBlock>{}, GN0, I1, Number<GN1PerBlockGN11>{}, GK1),
max_lds_align);
// LDS allocation for A and B: be careful of alignment
constexpr auto a_block_aligned_space_size = math::integer_least_multiple(
......@@ -201,7 +199,7 @@ struct GridwiseDynamicContractionDlops_A_GK0_GM0_GM1_GK1_B_GK0_GN0_GN1_GK1_C_GM0
const auto GM11 = Number<GM1PerBlockGM11>{};
const auto GM10 = GM1 / GM11;
const auto a_grid_desc_gk0_gm0_gm10_gm11_gk1 = transform_dynamic_tensor_descriptor(
const auto a_grid_desc_gk0_gm0_gm10_gm11_gk1 = transform_tensor_descriptor(
a_grid_desc_gk0_gm0_gm1_gk1,
make_tuple(make_pass_through_transform(GK0),
make_pass_through_transform(GM0),
......@@ -222,7 +220,7 @@ struct GridwiseDynamicContractionDlops_A_GK0_GM0_GM1_GK1_B_GK0_GN0_GN1_GK1_C_GM0
const auto GN11 = Number<GN1PerBlockGN11>{};
const auto GN10 = GN1 / GN11;
const auto b_grid_desc_gk0_gn0_gn10_gn11_gk1 = transform_dynamic_tensor_descriptor(
const auto b_grid_desc_gk0_gn0_gn10_gn11_gk1 = transform_tensor_descriptor(
b_grid_desc_gk0_gn0_gn1_gk1,
make_tuple(make_pass_through_transform(GK0),
make_pass_through_transform(GN0),
......@@ -259,7 +257,7 @@ struct GridwiseDynamicContractionDlops_A_GK0_GM0_GM1_GK1_B_GK0_GN0_GN1_GK1_C_GM0
constexpr auto BM0 = BM / BM1;
constexpr auto BN0 = BN / BN1;
const auto c_gm0_gm10_gm11_gn0_gn10_gn11_grid_desc = transform_dynamic_tensor_descriptor(
const auto c_gm0_gm10_gm11_gn0_gn10_gn11_grid_desc = transform_tensor_descriptor(
c_grid_desc_gm0_gm1_gn0_gn1,
make_tuple(make_pass_through_transform(GM0),
make_unmerge_transform(make_tuple(GM10, GM11)),
......@@ -268,7 +266,7 @@ struct GridwiseDynamicContractionDlops_A_GK0_GM0_GM1_GK1_B_GK0_GN0_GN1_GK1_C_GM0
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}, Sequence<4, 5>{}));
const auto c_gm10_bm_gn10_bn_grid_desc = transform_dynamic_tensor_descriptor(
const auto c_gm10_bm_gn10_bn_grid_desc = transform_tensor_descriptor(
c_gm0_gm10_gm11_gn0_gn10_gn11_grid_desc,
make_tuple(make_pass_through_transform(GM10),
make_merge_transform(make_tuple(GM0, GM11)),
......@@ -277,7 +275,7 @@ struct GridwiseDynamicContractionDlops_A_GK0_GM0_GM1_GK1_B_GK0_GN0_GN1_GK1_C_GM0
make_tuple(Sequence<1>{}, Sequence<0, 2>{}, Sequence<4>{}, Sequence<3, 5>{}),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}, Sequence<3>{}));
const auto c_grid_desc_gm10_bm0_bm1_gn10_bn0_bn1 = transform_dynamic_tensor_descriptor(
const auto c_grid_desc_gm10_bm0_bm1_gn10_bn0_bn1 = transform_tensor_descriptor(
c_gm10_bm_gn10_bn_grid_desc,
make_tuple(make_pass_through_transform(GM10),
make_unmerge_transform(make_tuple(BM0, BM1)),
......@@ -356,26 +354,24 @@ struct GridwiseDynamicContractionDlops_A_GK0_GM0_GM1_GK1_B_GK0_GN0_GN1_GK1_C_GM0
// A matrix in LDS memory, dst of blockwise copy
// be careful of LDS alignment
constexpr auto a_block_desc_gk0_gm0_gm10_gm11_gk1 =
make_dynamic_naive_tensor_descriptor_aligned_v2(
make_tuple(Number<GK0PerBlock>{}, GM0, I1, Number<GM1PerBlockGM11>{}, GK1),
max_lds_align);
constexpr auto a_block_desc_gk0_gm0_gm10_gm11_gk1 = make_naive_tensor_descriptor_aligned_v2(
make_tuple(Number<GK0PerBlock>{}, GM0, I1, Number<GM1PerBlockGM11>{}, GK1),
max_lds_align);
// B matrix in LDS memory, dst of blockwise copy
// be careful of LDS alignment
constexpr auto b_block_desc_gk0_gn0_gn10_gn11_gk1 =
make_dynamic_naive_tensor_descriptor_aligned_v2(
make_tuple(Number<GK0PerBlock>{}, GN0, I1, Number<GN1PerBlockGN11>{}, GK1),
max_lds_align);
constexpr auto b_block_desc_gk0_gn0_gn10_gn11_gk1 = make_naive_tensor_descriptor_aligned_v2(
make_tuple(Number<GK0PerBlock>{}, GN0, I1, Number<GN1PerBlockGN11>{}, GK1),
max_lds_align);
// A matrix in LDS memory for blockwise GEMM
// be careful of LDS alignment
constexpr auto a_block_desc_gk0_bm_gk1 = make_dynamic_naive_tensor_descriptor_aligned_v2(
constexpr auto a_block_desc_gk0_bm_gk1 = make_naive_tensor_descriptor_aligned_v2(
make_tuple(Number<GK0PerBlock>{}, GM0 * Number<GM1PerBlockGM11>{}, GK1), max_lds_align);
// B matrix in LDS memory for blockwise GEMM
// be careful of LDS alignment
constexpr auto b_block_desc_gk0_bn_gk1 = make_dynamic_naive_tensor_descriptor_aligned_v2(
constexpr auto b_block_desc_gk0_bn_gk1 = make_naive_tensor_descriptor_aligned_v2(
make_tuple(Number<GK0PerBlock>{}, GN0 * Number<GN1PerBlockGN11>{}, GK1), max_lds_align);
static_assert(a_block_desc_gk0_gm0_gm10_gm11_gk1.GetElementSpaceSize() ==
......@@ -385,7 +381,7 @@ struct GridwiseDynamicContractionDlops_A_GK0_GM0_GM1_GK1_B_GK0_GN0_GN1_GK1_C_GM0
"wrong!");
// A matrix blockwise copy
auto a_blockwise_copy = BlockwiseDynamicTensorSliceTransfer_v4r1<
auto a_blockwise_copy = BlockwiseTensorSliceTransfer_v4r1<
BlockSize,
InMemoryDataOperationEnum_t::Set,
Sequence<GK0PerBlock, GM0, 1, GM1PerBlockGM11, GK1.value>,
......@@ -409,7 +405,7 @@ struct GridwiseDynamicContractionDlops_A_GK0_GM0_GM1_GK1_B_GK0_GN0_GN1_GK1_C_GM0
make_multi_index(0, 0, 0, 0, 0));
// B matrix blockwise copy
auto b_blockwise_copy = BlockwiseDynamicTensorSliceTransfer_v4r1<
auto b_blockwise_copy = BlockwiseTensorSliceTransfer_v4r1<
BlockSize,
InMemoryDataOperationEnum_t::Set,
Sequence<GK0PerBlock, GN0, 1, GN1PerBlockGN11, GK1.value>,
......@@ -457,9 +453,8 @@ struct GridwiseDynamicContractionDlops_A_GK0_GM0_GM1_GK1_B_GK0_GN0_GN1_GK1_C_GM0
constexpr auto c_thread_tensor_lengths_bm0_bm1_bn0_bn1 =
decltype(blockwise_gemm)::GetCThreadTensorLengths_BM0_BM1_BN0_BN1();
constexpr auto c_thread_desc_bm0_bm1_bn0_bn1 =
make_dynamic_naive_tensor_descriptor_packed_v2(
sequence_to_tuple_of_number(c_thread_tensor_lengths_bm0_bm1_bn0_bn1));
constexpr auto c_thread_desc_bm0_bm1_bn0_bn1 = make_naive_tensor_descriptor_packed(
sequence_to_tuple_of_number(c_thread_tensor_lengths_bm0_bm1_bn0_bn1));
// LDS allocation for A and B: be careful of alignment
constexpr auto a_block_aligned_space_size = math::integer_least_multiple(
......@@ -475,9 +470,9 @@ struct GridwiseDynamicContractionDlops_A_GK0_GM0_GM1_GK1_B_GK0_GN0_GN1_GK1_C_GM0
auto c_thread_buf = make_static_buffer<AddressSpaceEnum_t::Vgpr, FloatAcc>(
c_thread_desc_bm0_bm1_bn0_bn1.GetElementSpaceSize());
ThreadwiseDynamicTensorSliceSet_v1<FloatAcc,
decltype(c_thread_desc_bm0_bm1_bn0_bn1),
decltype(c_thread_tensor_lengths_bm0_bm1_bn0_bn1)>{}
ThreadwiseTensorSliceSet_v1<FloatAcc,
decltype(c_thread_desc_bm0_bm1_bn0_bn1),
decltype(c_thread_tensor_lengths_bm0_bm1_bn0_bn1)>{}
.Run(c_thread_desc_bm0_bm1_bn0_bn1,
make_tuple(I0, I0, I0, I0),
c_thread_buf,
......@@ -501,9 +496,9 @@ struct GridwiseDynamicContractionDlops_A_GK0_GM0_GM1_GK1_B_GK0_GN0_GN1_GK1_C_GM0
// LDS double buffer: preload data into LDS
{
a_blockwise_copy.RunRead(
a_grid_desc_gk0_gm0_gm10_gm11_gk1, a_global_buf, AGridIteratorHacks{});
a_grid_desc_gk0_gm0_gm10_gm11_gk1, a_global_buf, AGridStepHacks{});
b_blockwise_copy.RunRead(
b_grid_desc_gk0_gn0_gn10_gn11_gk1, b_global_buf, BGridIteratorHacks{});
b_grid_desc_gk0_gn0_gn10_gn11_gk1, b_global_buf, BGridStepHacks{});
a_blockwise_copy.RunWrite(a_block_desc_gk0_gm0_gm10_gm11_gk1, a_block_even_buf);
b_blockwise_copy.RunWrite(b_block_desc_gk0_gn0_gn10_gn11_gk1, b_block_even_buf);
......@@ -520,18 +515,18 @@ struct GridwiseDynamicContractionDlops_A_GK0_GM0_GM1_GK1_B_GK0_GN0_GN1_GK1_C_GM0
// even iteration
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_gk0_gm0_gm10_gm11_gk1,
a_block_slice_copy_step,
AGridMoveSliceWindowIteratorHacks{});
AGridMoveSliceWindowStepHacks{});
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_gk0_gn0_gn10_gn11_gk1,
b_block_slice_copy_step,
BGridMoveSliceWindowIteratorHacks{});
BGridMoveSliceWindowStepHacks{});
__syncthreads();
// LDS doubel buffer: load next data from device mem
a_blockwise_copy.RunRead(
a_grid_desc_gk0_gm0_gm10_gm11_gk1, a_global_buf, AGridIteratorHacks{});
a_grid_desc_gk0_gm0_gm10_gm11_gk1, a_global_buf, AGridStepHacks{});
b_blockwise_copy.RunRead(
b_grid_desc_gk0_gn0_gn10_gn11_gk1, b_global_buf, BGridIteratorHacks{});
b_grid_desc_gk0_gn0_gn10_gn11_gk1, b_global_buf, BGridStepHacks{});
// LDS double buffer: GEMM on current data
blockwise_gemm.Run(c_thread_desc_bm0_bm1_bn0_bn1,
......@@ -546,18 +541,18 @@ struct GridwiseDynamicContractionDlops_A_GK0_GM0_GM1_GK1_B_GK0_GN0_GN1_GK1_C_GM0
// odd iteration
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_gk0_gm0_gm10_gm11_gk1,
a_block_slice_copy_step,
AGridMoveSliceWindowIteratorHacks{});
AGridMoveSliceWindowStepHacks{});
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_gk0_gn0_gn10_gn11_gk1,
b_block_slice_copy_step,
BGridMoveSliceWindowIteratorHacks{});
BGridMoveSliceWindowStepHacks{});
__syncthreads();
// LDS doubel buffer: load next data from device mem
a_blockwise_copy.RunRead(
a_grid_desc_gk0_gm0_gm10_gm11_gk1, a_global_buf, AGridIteratorHacks{});
a_grid_desc_gk0_gm0_gm10_gm11_gk1, a_global_buf, AGridStepHacks{});
b_blockwise_copy.RunRead(
b_grid_desc_gk0_gn0_gn10_gn11_gk1, b_global_buf, BGridIteratorHacks{});
b_grid_desc_gk0_gn0_gn10_gn11_gk1, b_global_buf, BGridStepHacks{});
// LDS double buffer: GEMM on current data
blockwise_gemm.Run(
......@@ -576,18 +571,18 @@ struct GridwiseDynamicContractionDlops_A_GK0_GM0_GM1_GK1_B_GK0_GN0_GN1_GK1_C_GM0
{
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_gk0_gm0_gm10_gm11_gk1,
a_block_slice_copy_step,
AGridMoveSliceWindowIteratorHacks{});
AGridMoveSliceWindowStepHacks{});
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_gk0_gn0_gn10_gn11_gk1,
b_block_slice_copy_step,
BGridMoveSliceWindowIteratorHacks{});
BGridMoveSliceWindowStepHacks{});
__syncthreads();
// LDS double buffer: load last data from device mem
a_blockwise_copy.RunRead(
a_grid_desc_gk0_gm0_gm10_gm11_gk1, a_global_buf, AGridIteratorHacks{});
a_grid_desc_gk0_gm0_gm10_gm11_gk1, a_global_buf, AGridStepHacks{});
b_blockwise_copy.RunRead(
b_grid_desc_gk0_gn0_gn10_gn11_gk1, b_global_buf, BGridIteratorHacks{});
b_grid_desc_gk0_gn0_gn10_gn11_gk1, b_global_buf, BGridStepHacks{});
// LDS double buffer: GEMM on 2nd-last data
blockwise_gemm.Run(
......@@ -615,7 +610,7 @@ struct GridwiseDynamicContractionDlops_A_GK0_GM0_GM1_GK1_B_GK0_GN0_GN1_GK1_C_GM0
// output: register to global memory
{
constexpr auto c_thread_desc_gm10_bm0_bm1_gn10_bn0_bn1 =
make_dynamic_naive_tensor_descriptor_packed_v2(
make_naive_tensor_descriptor_packed(
make_tuple(I1,
Number<c_thread_tensor_lengths_bm0_bm1_bn0_bn1[I0]>{},
Number<c_thread_tensor_lengths_bm0_bm1_bn0_bn1[I1]>{},
......@@ -627,7 +622,7 @@ struct GridwiseDynamicContractionDlops_A_GK0_GM0_GM1_GK1_B_GK0_GN0_GN1_GK1_C_GM0
blockwise_gemm.CalculateCThreadOriginOnBlock_BM0_BM1_BN0_BN1(
get_thread_local_1d_id());
ThreadwiseDynamicTensorSliceTransfer_v1r3<
ThreadwiseTensorSliceTransfer_v1r3<
FloatAcc,
FloatC,
decltype(c_thread_desc_gm10_bm0_bm1_gn10_bn0_bn1),
......@@ -655,7 +650,7 @@ struct GridwiseDynamicContractionDlops_A_GK0_GM0_GM1_GK1_B_GK0_GN0_GN1_GK1_C_GM0
c_thread_buf,
c_grid_desc_gm10_bm0_bm1_gn10_bn0_bn1,
c_grid_buf,
CGridIteratorHacks{});
CGridStepHacks{});
}
}
};
......
composable_kernel/include/tensor_operation/gridwise_dynamic_gemm_dlops_v1r3.hpp → composable_kernel/include/tensor_operation/gridwise_gemm_dlops_v1r3.hpp
View file @ ccc4a1d3
#ifndef CK_GRIDWISE_DYNAMIC_GEMM_V1R3_HPP
#define CK_GRIDWISE_DYNAMIC_GEMM_V1R3_HPP
#ifndef CK_GRIDWISE_GEMM_V1R3_HPP
#define CK_GRIDWISE_GEMM_V1R3_HPP
#include "common_header.hpp"
#include "dynamic_multi_index_transform_helper.hpp"
#include "dynamic_tensor_descriptor.hpp"
#include "dynamic_tensor_descriptor_helper.hpp"
#include "multi_index_transform_helper.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
#include "blockwise_gemm_dlops_v2r3.hpp"
#include "blockwise_dynamic_tensor_slice_transfer_v2.hpp"
#include "threadwise_dynamic_tensor_slice_transfer_v2.hpp"
#include "threadwise_dynamic_tensor_slice_set.hpp"
#include "blockwise_tensor_slice_transfer_v2.hpp"
#include "threadwise_tensor_slice_transfer_v2.hpp"
#include "threadwise_tensor_slice_set.hpp"
namespace ck {
......@@ -26,7 +26,7 @@ __global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
kernel_dynamic_gemm_dlops_v1r3(
kernel_gemm_dlops_v1r3(
const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid,
......@@ -68,28 +68,27 @@ __global__ void
#if CK_USE_LAUNCH_BOUNDS
__launch_bounds__(CK_MAX_THREAD_PER_BLOCK, CK_MIN_BLOCK_PER_CU)
#endif
kernel_dynamic_gemm_dlops_v1r3(
const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid,
const void CONSTANT* p_a_k0_m0_m1_k1_grid_desc,
const void CONSTANT* p_b_k0_n0_n1_k1_grid_desc,
const void CONSTANT* p_c_m0_m10_m11_n0_n10_n11_grid_desc,
const void CONSTANT* p_c_blockid_to_m0_n0_block_cluster_adaptor)
kernel_gemm_dlops_v1r3(const FloatAB* __restrict__ p_a_grid,
const FloatAB* __restrict__ p_b_grid,
FloatC* __restrict__ p_c_grid,
const void CONSTANT* p_a_k0_m0_m1_k1_grid_desc,
const void CONSTANT* p_b_k0_n0_n1_k1_grid_desc,
const void CONSTANT* p_c_m0_m10_m11_n0_n10_n11_grid_desc,
const void CONSTANT* p_c_blockid_to_m0_n0_block_cluster_adaptor)
{
// first cast void CONSTANT void* to void*
// second cast void* to Desc*
// the copy constructor of tensor descriptor doesn't take address_space(4)
const auto a_k0_m0_m1_k1_grid_desc =
*reinterpret_cast<const AK0M0M1K1GridDesc*>((const void*)p_a_k0_m0_m1_k1_grid_desc);
const auto b_k0_n0_n1_k1_grid_desc =
*reinterpret_cast<const BK0N0N1K1GridDesc*>((const void*)p_b_k0_n0_n1_k1_grid_desc);
const auto a_k0_m0_m1_k1_grid_desc = *reinterpret_cast<const AK0M0M1K1GridDesc*>(
cast_pointer_to_generic_address_space(p_a_k0_m0_m1_k1_grid_desc));
const auto b_k0_n0_n1_k1_grid_desc = *reinterpret_cast<const BK0N0N1K1GridDesc*>(
cast_pointer_to_generic_address_space(p_b_k0_n0_n1_k1_grid_desc));
const auto c_m0_m10_m11_n0_n10_n11_grid_desc =
*reinterpret_cast<const CM0M10M11N0N10N11GridDesc*>(
(const void*)p_c_m0_m10_m11_n0_n10_n11_grid_desc);
cast_pointer_to_generic_address_space(p_c_m0_m10_m11_n0_n10_n11_grid_desc));
const auto c_blockid_to_m0_n0_block_cluster_adaptor =
*reinterpret_cast<const CBlockIdToM0N0BlockClusterAdaptor*>(
(const void*)p_c_blockid_to_m0_n0_block_cluster_adaptor);
cast_pointer_to_generic_address_space(p_c_blockid_to_m0_n0_block_cluster_adaptor));
constexpr index_t shared_block_size =
GridwiseGemm::GetSharedMemoryNumberOfByte() / sizeof(FloatAB);
......@@ -142,12 +141,12 @@ template <index_t BlockSize,
typename CThreadTransferSrcDstAccessOrder,
index_t CThreadTransferSrcDstVectorDim,
index_t CThreadTransferDstScalarPerVector,
typename AGridIteratorHacks,
typename BGridIteratorHacks,
typename CGridIteratorHacks,
typename AGridMoveSliceWindowIteratorHacks,
typename BGridMoveSliceWindowIteratorHacks>
struct GridwiseDynamicGemmDlops_km_kn_mn_v1r3
typename AGridStepHacks,
typename BGridStepHacks,
typename CGridStepHacks,
typename AGridMoveSliceWindowStepHacks,
typename BGridMoveSliceWindowStepHacks>
struct GridwiseGemmDlops_km_kn_mn_v1r3
{
static constexpr auto I0 = Number<0>{};
static constexpr auto I1 = Number<1>{};
......@@ -164,12 +163,12 @@ struct GridwiseDynamicGemmDlops_km_kn_mn_v1r3
// TODO: check alignment
// A matrix in LDS memory, dst of blockwise copy
constexpr auto a_k_m_block_desc = make_dynamic_naive_tensor_descriptor_aligned_v2(
constexpr auto a_k_m_block_desc = make_naive_tensor_descriptor_aligned_v2(
make_tuple(Number<KPerBlock>{}, Number<MPerBlockM1>{}, K1), max_lds_align);
// TODO: check alignment
// B matrix in LDS memory, dst of blockwise copy
constexpr auto b_k_n_block_desc = make_dynamic_naive_tensor_descriptor_aligned_v2(
constexpr auto b_k_n_block_desc = make_naive_tensor_descriptor_aligned_v2(
make_tuple(Number<KPerBlock>{}, Number<NPerBlockN1>{}, K1), max_lds_align);
// TODO: check alignment
......@@ -191,12 +190,12 @@ struct GridwiseDynamicGemmDlops_km_kn_mn_v1r3
const auto M = a_k0_m_k1_grid_desc.GetLength(I1);
const auto N = b_k0_n_k1_grid_desc.GetLength(I1);
const auto K0 = a_k0_m_k1_grid_desc.GetLength(I0);
const auto K1 = a_k0_m_k1_grid_desc.GetLength(I2);
// TODO: also check validity of all components (blockwise-copy, threadwise-copy, etc)
return (M == c_m_n_grid_desc.GetLength(I0) && N == c_m_n_grid_desc.GetLength(I1) &&
K0 == b_k0_n_k1_grid_desc.GetLength(I0) &&
K1 == a_k0_m_k1_grid_desc.GetLength(I2) &&
K1 == b_k0_n_k1_grid_desc.GetLength(I2)) &&
(M % MPerBlockM1 == 0 && N % NPerBlockN1 == 0 && K0 % KPerBlock == 0);
}
......@@ -231,13 +230,13 @@ struct GridwiseDynamicGemmDlops_km_kn_mn_v1r3
const auto M1 = Number<MPerBlockM1>{};
const auto M0 = M / M1;
const auto a_k0_m0_m1_k1_grid_desc = transform_dynamic_tensor_descriptor(
a_k0_m_k1_grid_desc,
make_tuple(make_pass_through_transform(K0),
make_unmerge_transform(make_tuple(M0, M1)),
make_pass_through_transform(K1)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
const auto a_k0_m0_m1_k1_grid_desc =
transform_tensor_descriptor(a_k0_m_k1_grid_desc,
make_tuple(make_pass_through_transform(K0),
make_unmerge_transform(make_tuple(M0, M1)),
make_pass_through_transform(K1)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
return a_k0_m0_m1_k1_grid_desc;
}
......@@ -251,13 +250,13 @@ struct GridwiseDynamicGemmDlops_km_kn_mn_v1r3
const auto N1 = Number<NPerBlockN1>{};
const auto N0 = N / N1;
const auto b_k0_n0_n1_k1_grid_desc = transform_dynamic_tensor_descriptor(
b_k0_n_k1_grid_desc,
make_tuple(make_pass_through_transform(K0),
make_unmerge_transform(make_tuple(N0, N1)),
make_pass_through_transform(K1)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
const auto b_k0_n0_n1_k1_grid_desc =
transform_tensor_descriptor(b_k0_n_k1_grid_desc,
make_tuple(make_pass_through_transform(K0),
make_unmerge_transform(make_tuple(N0, N1)),
make_pass_through_transform(K1)),
make_tuple(Sequence<0>{}, Sequence<1>{}, Sequence<2>{}),
make_tuple(Sequence<0>{}, Sequence<1, 2>{}, Sequence<3>{}));
return b_k0_n0_n1_k1_grid_desc;
}
......@@ -284,7 +283,7 @@ struct GridwiseDynamicGemmDlops_km_kn_mn_v1r3
constexpr auto M10 = M1 / M11;
constexpr auto N10 = N1 / N11;
const auto c_m0_m10_m11_n0_n10_n11_grid_desc = transform_dynamic_tensor_descriptor(
const auto c_m0_m10_m11_n0_n10_n11_grid_desc = transform_tensor_descriptor(
c_m_n_grid_desc,
make_tuple(make_unmerge_transform(make_tuple(M0, M10, M11)),
make_unmerge_transform(make_tuple(N0, N10, N11))),
......@@ -355,23 +354,23 @@ struct GridwiseDynamicGemmDlops_km_kn_mn_v1r3
// TODO: check alignment
// A matrix in LDS memory, dst of blockwise copy
// be careful of LDS alignment
constexpr auto a_k0_m0_m1_k1_block_desc = make_dynamic_naive_tensor_descriptor_aligned_v2(
constexpr auto a_k0_m0_m1_k1_block_desc = make_naive_tensor_descriptor_aligned_v2(
make_tuple(Number<KPerBlock>{}, I1, Number<MPerBlockM1>{}, K1), max_lds_align);
// TODO: check alignment
// B matrix in LDS memory, dst of blockwise copy
// be careful of LDS alignment
constexpr auto b_k0_n0_n1_k1_block_desc = make_dynamic_naive_tensor_descriptor_aligned_v2(
constexpr auto b_k0_n0_n1_k1_block_desc = make_naive_tensor_descriptor_aligned_v2(
make_tuple(Number<KPerBlock>{}, I1, Number<NPerBlockN1>{}, K1), max_lds_align);
// TODO: check alignment
// A matrix in LDS memory, for blockwise GEMM
constexpr auto a_k0_m_k1_block_desc = make_dynamic_naive_tensor_descriptor_aligned_v2(
constexpr auto a_k0_m_k1_block_desc = make_naive_tensor_descriptor_aligned_v2(
make_tuple(Number<KPerBlock>{}, Number<MPerBlockM1>{}, K1), max_lds_align);
// TODO: check alignment
// B matrix in LDS memory, for blockwise GEMM
constexpr auto b_k0_n_k1_block_desc = make_dynamic_naive_tensor_descriptor_aligned_v2(
constexpr auto b_k0_n_k1_block_desc = make_naive_tensor_descriptor_aligned_v2(
make_tuple(Number<KPerBlock>{}, Number<NPerBlockN1>{}, K1), max_lds_align);
static_assert(a_k0_m0_m1_k1_block_desc.GetElementSpaceSize() ==
......@@ -381,7 +380,7 @@ struct GridwiseDynamicGemmDlops_km_kn_mn_v1r3
"wrong!");
// A matrix blockwise copy
auto a_blockwise_copy = BlockwiseDynamicTensorSliceTransfer_v4r1<
auto a_blockwise_copy = BlockwiseTensorSliceTransfer_v4r1<
BlockSize,
InMemoryDataOperationEnum_t::Set,
Sequence<KPerBlock, 1, MPerBlockM1, K1.value>,
......@@ -405,7 +404,7 @@ struct GridwiseDynamicGemmDlops_km_kn_mn_v1r3
make_multi_index(0, 0, 0, 0));
// B matrix blockwise copy
auto b_blockwise_copy = BlockwiseDynamicTensorSliceTransfer_v4r1<
auto b_blockwise_copy = BlockwiseTensorSliceTransfer_v4r1<
BlockSize,
InMemoryDataOperationEnum_t::Set,
Sequence<KPerBlock, 1, NPerBlockN1, K1.value>,
......@@ -453,9 +452,8 @@ struct GridwiseDynamicGemmDlops_km_kn_mn_v1r3
constexpr auto c_m10_m11_n10_n11_thread_tensor_lengths =
decltype(blockwise_gemm)::GetCThreadTensorLengths_BM0_BM1_BN0_BN1();
constexpr auto c_m10_m11_n10_n11_thread_desc =
make_dynamic_naive_tensor_descriptor_packed_v2(
sequence_to_tuple_of_number(c_m10_m11_n10_n11_thread_tensor_lengths));
constexpr auto c_m10_m11_n10_n11_thread_desc = make_naive_tensor_descriptor_packed(
sequence_to_tuple_of_number(c_m10_m11_n10_n11_thread_tensor_lengths));
// LDS allocation for A and B: be careful of alignment
constexpr auto a_block_aligned_space_size = math::integer_least_multiple(
......@@ -471,9 +469,9 @@ struct GridwiseDynamicGemmDlops_km_kn_mn_v1r3
auto c_thread_buf = make_static_buffer<AddressSpaceEnum_t::Vgpr, FloatAcc>(
c_m10_m11_n10_n11_thread_desc.GetElementSpaceSize());
ThreadwiseDynamicTensorSliceSet_v1<FloatAcc,
decltype(c_m10_m11_n10_n11_thread_desc),
decltype(c_m10_m11_n10_n11_thread_tensor_lengths)>{}
ThreadwiseTensorSliceSet_v1<FloatAcc,
decltype(c_m10_m11_n10_n11_thread_desc),
decltype(c_m10_m11_n10_n11_thread_tensor_lengths)>{}
.Run(c_m10_m11_n10_n11_thread_desc,
make_tuple(I0, I0, I0, I0),
c_thread_buf,
......@@ -496,8 +494,8 @@ struct GridwiseDynamicGemmDlops_km_kn_mn_v1r3
// LDS double buffer: preload data into LDS
{
a_blockwise_copy.RunRead(a_k0_m0_m1_k1_grid_desc, a_global_buf, AGridIteratorHacks{});
b_blockwise_copy.RunRead(b_k0_n0_n1_k1_grid_desc, b_global_buf, BGridIteratorHacks{});
a_blockwise_copy.RunRead(a_k0_m0_m1_k1_grid_desc, a_global_buf, AGridStepHacks{});
b_blockwise_copy.RunRead(b_k0_n0_n1_k1_grid_desc, b_global_buf, BGridStepHacks{});
a_blockwise_copy.RunWrite(a_k0_m0_m1_k1_block_desc, a_block_even_buf);
b_blockwise_copy.RunWrite(b_k0_n0_n1_k1_block_desc, b_block_even_buf);
......@@ -516,18 +514,16 @@ struct GridwiseDynamicGemmDlops_km_kn_mn_v1r3
// even iteration
a_blockwise_copy.MoveSrcSliceWindow(a_k0_m0_m1_k1_grid_desc,
a_block_slice_copy_step,
AGridMoveSliceWindowIteratorHacks{});
AGridMoveSliceWindowStepHacks{});
b_blockwise_copy.MoveSrcSliceWindow(b_k0_n0_n1_k1_grid_desc,
b_block_slice_copy_step,
BGridMoveSliceWindowIteratorHacks{});
BGridMoveSliceWindowStepHacks{});
__syncthreads();
// LDS doubel buffer: load next data from device mem
a_blockwise_copy.RunRead(
a_k0_m0_m1_k1_grid_desc, a_global_buf, AGridIteratorHacks{});
b_blockwise_copy.RunRead(
b_k0_n0_n1_k1_grid_desc, b_global_buf, BGridIteratorHacks{});
a_blockwise_copy.RunRead(a_k0_m0_m1_k1_grid_desc, a_global_buf, AGridStepHacks{});
b_blockwise_copy.RunRead(b_k0_n0_n1_k1_grid_desc, b_global_buf, BGridStepHacks{});
// LDS double buffer: GEMM on current data
blockwise_gemm.Run(c_m10_m11_n10_n11_thread_desc,
......@@ -542,18 +538,16 @@ struct GridwiseDynamicGemmDlops_km_kn_mn_v1r3
// odd iteration
a_blockwise_copy.MoveSrcSliceWindow(a_k0_m0_m1_k1_grid_desc,
a_block_slice_copy_step,
AGridMoveSliceWindowIteratorHacks{});
AGridMoveSliceWindowStepHacks{});
b_blockwise_copy.MoveSrcSliceWindow(b_k0_n0_n1_k1_grid_desc,
b_block_slice_copy_step,
BGridMoveSliceWindowIteratorHacks{});
BGridMoveSliceWindowStepHacks{});
__syncthreads();
// LDS doubel buffer: load next data from device mem
a_blockwise_copy.RunRead(
a_k0_m0_m1_k1_grid_desc, a_global_buf, AGridIteratorHacks{});
b_blockwise_copy.RunRead(
b_k0_n0_n1_k1_grid_desc, b_global_buf, BGridIteratorHacks{});
a_blockwise_copy.RunRead(a_k0_m0_m1_k1_grid_desc, a_global_buf, AGridStepHacks{});
b_blockwise_copy.RunRead(b_k0_n0_n1_k1_grid_desc, b_global_buf, BGridStepHacks{});
// LDS double buffer: GEMM on current data
blockwise_gemm.Run(
......@@ -570,18 +564,16 @@ struct GridwiseDynamicGemmDlops_km_kn_mn_v1r3
// LDS double buffer: tail
if constexpr(HasDoubleTailKBlockLoop) // if has 2 iteration left
{
a_blockwise_copy.MoveSrcSliceWindow(a_k0_m0_m1_k1_grid_desc,
a_block_slice_copy_step,
AGridMoveSliceWindowIteratorHacks{});
b_blockwise_copy.MoveSrcSliceWindow(b_k0_n0_n1_k1_grid_desc,
b_block_slice_copy_step,
BGridMoveSliceWindowIteratorHacks{});
a_blockwise_copy.MoveSrcSliceWindow(
a_k0_m0_m1_k1_grid_desc, a_block_slice_copy_step, AGridMoveSliceWindowStepHacks{});
b_blockwise_copy.MoveSrcSliceWindow(
b_k0_n0_n1_k1_grid_desc, b_block_slice_copy_step, BGridMoveSliceWindowStepHacks{});
__syncthreads();
// LDS double buffer: load last data from device mem
a_blockwise_copy.RunRead(a_k0_m0_m1_k1_grid_desc, a_global_buf, AGridIteratorHacks{});
b_blockwise_copy.RunRead(b_k0_n0_n1_k1_grid_desc, b_global_buf, BGridIteratorHacks{});
a_blockwise_copy.RunRead(a_k0_m0_m1_k1_grid_desc, a_global_buf, AGridStepHacks{});
b_blockwise_copy.RunRead(b_k0_n0_n1_k1_grid_desc, b_global_buf, BGridStepHacks{});
// LDS double buffer: GEMM on 2nd-last data
blockwise_gemm.Run(
......@@ -608,21 +600,8 @@ struct GridwiseDynamicGemmDlops_km_kn_mn_v1r3
// output: register to global memory
{
constexpr auto M11 =
Number<container_reduce(M11N11ThreadClusterM110Xs{}, math::multiplies_v2{}, I1) *
M1PerThreadM111>{};
constexpr auto N11 =
Number<container_reduce(M11N11ThreadClusterN110Xs{}, math::multiplies_v2{}, I1) *
N1PerThreadN111>{};
constexpr index_t M10 = MPerBlockM1 / M11;
constexpr index_t N10 = NPerBlockN1 / N11;
constexpr index_t M111 = M1PerThreadM111;
constexpr index_t N111 = N1PerThreadN111;
constexpr auto c_m0_m10_m11_n0_n10_n11_thread_desc =
make_dynamic_naive_tensor_descriptor_packed_v2(
make_naive_tensor_descriptor_packed(
make_tuple(I1,
Number<c_m10_m11_n10_n11_thread_tensor_lengths[I0]>{},
Number<c_m10_m11_n10_n11_thread_tensor_lengths[I1]>{},
......@@ -634,7 +613,7 @@ struct GridwiseDynamicGemmDlops_km_kn_mn_v1r3
blockwise_gemm.CalculateCThreadOriginOnBlock_BM0_BM1_BN0_BN1(
get_thread_local_1d_id());
ThreadwiseDynamicTensorSliceTransfer_v1r3<
ThreadwiseTensorSliceTransfer_v1r3<
FloatAcc,
FloatC,
decltype(c_m0_m10_m11_n0_n10_n11_thread_desc),
......@@ -662,7 +641,7 @@ struct GridwiseDynamicGemmDlops_km_kn_mn_v1r3
c_thread_buf,
c_m0_m10_m11_n0_n10_n11_grid_desc,
c_grid_buf,
CGridIteratorHacks{});
CGridStepHacks{});
}
}
};
......
composable_kernel/include/tensor_operation/threadwise_contraction_dlops.hpp
View file @ ccc4a1d3
......@@ -21,10 +21,10 @@ template <typename FloatA,
typename TKLengths,
typename TMLengths,
typename TNLengths,
typename std::enable_if<AThreadDesc_TK0_TM0_TM1_TK1::IsKnownAtCompileTime() &&
BThreadDesc_TK0_TN0_TN1_TK1::IsKnownAtCompileTime() &&
CThreadDesc_TM0_TM1_TN0_TN1::IsKnownAtCompileTime(),
bool>::type = false>
typename enable_if<AThreadDesc_TK0_TM0_TM1_TK1::IsKnownAtCompileTime() &&
BThreadDesc_TK0_TN0_TN1_TK1::IsKnownAtCompileTime() &&
CThreadDesc_TM0_TM1_TN0_TN1::IsKnownAtCompileTime(),
bool>::type = false>
struct ThreadwiseGemmDlops_km0m1_kn0n1_m0m1n0n1
{
__device__ constexpr ThreadwiseGemmDlops_km0m1_kn0n1_m0m1n0n1()
......@@ -97,10 +97,9 @@ struct ThreadwiseGemmDlops_km0m1_kn0n1_m0m1n0n1
CThreadDesc_TM0_TM1_TN0_TN1{}.CalculateOffset(
c_origin_idx + make_multi_index(tm0, tm1, tn0, tn1));
amd_inner_product_dlop<FloatA, FloatB, FloatC>(
a_buf[Number<a_offset>{}],
b_buf[Number<b_offset>{}],
c_buf(Number<c_offset>{}));
inner_product<FloatA, FloatB, FloatC>(a_buf[Number<a_offset>{}],
b_buf[Number<b_offset>{}],
c_buf(Number<c_offset>{}));
});
});
});
......@@ -124,10 +123,10 @@ template <typename FloatA,
typename TKLengths,
typename TMLengths,
typename TNLengths,
typename std::enable_if<AThreadDesc_TK0_TM0_TM1_TK1::IsKnownAtCompileTime() &&
BThreadDesc_TK0_TN0_TN1_TK1::IsKnownAtCompileTime() &&
CThreadDesc_TM0_TM1_TN0_TN1::IsKnownAtCompileTime(),
bool>::type = false>
typename enable_if<AThreadDesc_TK0_TM0_TM1_TK1::IsKnownAtCompileTime() &&
BThreadDesc_TK0_TN0_TN1_TK1::IsKnownAtCompileTime() &&
CThreadDesc_TM0_TM1_TN0_TN1::IsKnownAtCompileTime(),
bool>::type = false>
struct ThreadwiseContractionDlops_A_TK0_TM0_TM1_TK1_B_TK0_TN0_TN1_TK1_C_TM0_TM1_TN0_TN1
{
__device__ constexpr ThreadwiseContractionDlops_A_TK0_TM0_TM1_TK1_B_TK0_TN0_TN1_TK1_C_TM0_TM1_TN0_TN1()
......@@ -214,7 +213,7 @@ struct ThreadwiseContractionDlops_A_TK0_TM0_TM1_TK1_B_TK0_TN0_TN1_TK1_C_TM0_TM1_
CThreadDesc_TM0_TM1_TN0_TN1{}.CalculateOffset(
c_origin_idx + make_multi_index(tm0, tm1, tn0, tn1));
amd_inner_product_dlop<a_vector_t, b_vector_t, FloatC>(
inner_product<a_vector_t, b_vector_t, FloatC>(
a_vec.template AsType<a_vector_t>()[I0],
b_vec.template AsType<b_vector_t>()[I0],
c_buf(Number<c_offset>{}));
......
composable_kernel/include/tensor_operation/threadwise_gemm_dlops_v3.hpp
View file @ ccc4a1d3
......@@ -19,9 +19,9 @@ template <typename FloatA,
typename CDesc,
index_t H,
index_t W,
typename std::enable_if<ADesc::IsKnownAtCompileTime() && BDesc::IsKnownAtCompileTime() &&
CDesc::IsKnownAtCompileTime(),
bool>::type = false>
typename enable_if<ADesc::IsKnownAtCompileTime() && BDesc::IsKnownAtCompileTime() &&
CDesc::IsKnownAtCompileTime(),
bool>::type = false>
struct ThreadwiseGemmDlops_km_kn_mn_v3
{
template <typename ABuffer,
......@@ -57,8 +57,6 @@ struct ThreadwiseGemmDlops_km_kn_mn_v3
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
constexpr auto I2 = Number<2>{};
constexpr auto I3 = Number<3>{};
constexpr auto E = ADesc{}.GetLength(I0);
constexpr auto K = ADesc{}.GetLength(I1);
......
composable_kernel/include/tensor_operation/threadwise_dynamic_tensor_slice_set.hpp → composable_kernel/include/tensor_operation/threadwise_tensor_slice_set.hpp
View file @ ccc4a1d3
#ifndef CK_THREADWISE_DYNAMIC_TENSOR_SET_HPP
#define CK_THREADWISE_DYNAMIC_TENSOR_SET_HPP
#ifndef CK_THREADWISE_TENSOR_SET_HPP
#define CK_THREADWISE_TENSOR_SET_HPP
#include "common_header.hpp"
#include "dynamic_tensor_descriptor.hpp"
#include "dynamic_tensor_descriptor_helper.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
namespace ck {
......@@ -11,12 +11,12 @@ namespace ck {
// 1. Desc is known at compile-time
// 2. Buffer is StaticBuffer
// 3. OriginIdx is known at compile-time
// 4. use #-iterator
// 4. use #-step
template <typename Data,
typename Desc,
typename SliceLengths,
typename std::enable_if<Desc::IsKnownAtCompileTime(), bool>::type = false>
struct ThreadwiseDynamicTensorSliceSet_v1
typename enable_if<Desc::IsKnownAtCompileTime(), bool>::type = false>
struct ThreadwiseTensorSliceSet_v1
{
static constexpr index_t nDim = SliceLengths::Size();
......@@ -40,7 +40,7 @@ struct ThreadwiseDynamicTensorSliceSet_v1
constexpr auto origin_idx = to_multi_index(OriginIdx{});
static_ford<SliceLengths>{}([&](auto access_idx) {
constexpr auto coord = make_dynamic_tensor_coordinate(desc, origin_idx + access_idx);
constexpr auto coord = make_tensor_coordinate(desc, origin_idx + access_idx);
constexpr bool is_valid =
coordinate_has_valid_offset_assuming_visible_index_is_valid(desc, coord);
......
composable_kernel/include/tensor_operation/xdlops_gemm.hpp
View file @ ccc4a1d3
......@@ -350,8 +350,8 @@ struct mfma_info<mfma_instr::mfma_f32_32x32x2bf16>
class FloatC>
__device__ FloatC run(const FloatA* a, const FloatB* b, FloatC reg_c) const
{
const auto p_a = reinterpret_cast<const ushort2_t*>(a);
const auto p_b = reinterpret_cast<const ushort2_t*>(b);
const auto p_a = c_style_pointer_cast<const ushort2_t*>(a);
const auto p_b = c_style_pointer_cast<const ushort2_t*>(b);
return intrin_mfma_f32_32x32x2bf16<MPerXdlops, NPerXdlops, AStride, BStride>::run(
p_a, p_b, reg_c);
......@@ -384,8 +384,8 @@ struct mfma_info<mfma_instr::mfma_f32_32x32x4bf16>
class FloatC>
__device__ FloatC run(const FloatA* a, const FloatB* b, FloatC reg_c) const
{
const auto p_a = reinterpret_cast<const ushort2_t*>(a);
const auto p_b = reinterpret_cast<const ushort2_t*>(b);
const auto p_a = c_style_pointer_cast<const ushort2_t*>(a);
const auto p_b = c_style_pointer_cast<const ushort2_t*>(b);
return intrin_mfma_f32_32x32x4bf16(p_a, p_b, reg_c);
}
......@@ -417,8 +417,8 @@ struct mfma_info<mfma_instr::mfma_f32_16x16x8bf16>
class FloatC>
__device__ FloatC run(const FloatA* a, const FloatB* b, FloatC reg_c) const
{
const auto p_a = reinterpret_cast<const ushort2_t*>(a);
const auto p_b = reinterpret_cast<const ushort2_t*>(b);
const auto p_a = c_style_pointer_cast<const ushort2_t*>(a);
const auto p_b = c_style_pointer_cast<const ushort2_t*>(b);
return intrin_mfma_f32_16x16x8bf16(p_a, p_b, reg_c);
}
......@@ -450,8 +450,8 @@ struct mfma_info<mfma_instr::mfma_f32_16x16x2bf16>
class FloatC>
__device__ FloatC run(const FloatA* a, const FloatB* b, FloatC reg_c) const
{
const auto p_a = reinterpret_cast<const ushort2_t*>(a);
const auto p_b = reinterpret_cast<const ushort2_t*>(b);
const auto p_a = c_style_pointer_cast<const ushort2_t*>(a);
const auto p_b = c_style_pointer_cast<const ushort2_t*>(b);
return intrin_mfma_f32_16x16x2bf16<MPerXdlops, NPerXdlops>(p_a, p_b, reg_c);
}
......@@ -483,8 +483,8 @@ struct mfma_info<mfma_instr::mfma_f32_4x4x2bf16>
class FloatC>
__device__ FloatC run(const FloatA* a, const FloatB* b, FloatC reg_c) const
{
const auto p_a = reinterpret_cast<const ushort2_t*>(a);
const auto p_b = reinterpret_cast<const ushort2_t*>(b);
const auto p_a = c_style_pointer_cast<const ushort2_t*>(a);
const auto p_b = c_style_pointer_cast<const ushort2_t*>(b);
return intrin_mfma_f32_4x4x2bf16<MPerXdlops, NPerXdlops>::run(p_a, p_b, reg_c);
}
......
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