Skip to content

GitLab

"sgl-kernel/vscode:/vscode.git/clone" did not exist on "b188a89a5d09ba634c77c34a2407e95dea5826b8"
Unverified Commit b62bf8c3 authored by Chao Liu's avatar Chao Liu Committed by GitHub
Browse files

Merge pull request #14 from ROCmSoftwarePlatform/miopen_downstream_init_integration 

MIOpen Downstream: Initial integration 2nd PR
parents ccc4a1d3 67ad47e7
Show whitespace changes
Inline Side-by-side
Showing with 63 additions and 69 deletions
composable_kernel/include/tensor_description/cluster_descriptor.hpp
View file @ b62bf8c3
......@@ -8,7 +8,7 @@ namespace ck {
template <typename Lengths,
typename ArrangeOrder = typename arithmetic_sequence_gen<0, Lengths::Size(), 1>::type>
__host__ __device__ constexpr auto make_cluster_descriptor_v2(
__host__ __device__ constexpr auto make_cluster_descriptor(
const Lengths& lengths,
ArrangeOrder order = typename arithmetic_sequence_gen<0, Lengths::Size(), 1>::type{})
{
......
composable_kernel/include/tensor_description/multi_index_transform.hpp
View file @ b62bf8c3
......@@ -481,11 +481,11 @@ struct Merge_v1_carry_check
using LowerIndex = MultiIndex<NDimLow>;
using UpperIndex = MultiIndex<1>;
using LowLengthsScan = decltype(
container_reverse_exclusive_scan(LowLengths{}, math::multiplies_v2{}, Number<1>{}));
using LowLengthsScan =
decltype(container_reverse_exclusive_scan(LowLengths{}, math::multiplies{}, Number<1>{}));
using UpLengths =
decltype(make_tuple(container_reduce(LowLengths{}, math::multiplies_v2{}, Number<1>{})));
decltype(make_tuple(container_reduce(LowLengths{}, math::multiplies{}, Number<1>{})));
LowLengths low_lengths_;
LowLengthsScan low_lengths_scan_;
......@@ -496,8 +496,8 @@ struct Merge_v1_carry_check
__host__ __device__ constexpr Merge_v1_carry_check(const LowLengths& low_lengths)
: low_lengths_{low_lengths},
low_lengths_scan_{
container_reverse_exclusive_scan(low_lengths, math::multiplies_v2{}, Number<1>{})},
up_lengths_{make_tuple(container_reduce(low_lengths, math::multiplies_v2{}, Number<1>{}))}
container_reverse_exclusive_scan(low_lengths, math::multiplies{}, Number<1>{})},
up_lengths_{make_tuple(container_reduce(low_lengths, math::multiplies{}, Number<1>{}))}
{
static_assert(LowerIndex::Size() == NDimLow, "wrong!");
}
......@@ -1037,7 +1037,7 @@ struct Merge_v2_magic_division
using UpperIndex = MultiIndex<1>;
using UpLengths =
decltype(make_tuple(container_reduce(LowLengths{}, math::multiplies_v2{}, Number<1>{})));
decltype(make_tuple(container_reduce(LowLengths{}, math::multiplies{}, Number<1>{})));
using LowLengthsMagicDivisorMultipiler = decltype(
generate_tuple(lambda_merge_generate_MagicDivision_calculate_magic_multiplier<LowLengths>{},
......@@ -1062,7 +1062,7 @@ struct Merge_v2_magic_division
low_lengths_magic_divisor_shift_{generate_tuple(
[&](auto i) { return MagicDivision::CalculateMagicShift(low_lengths[i]); },
Number<NDimLow>{})},
up_lengths_{make_tuple(container_reduce(low_lengths, math::multiplies_v2{}, Number<1>{}))}
up_lengths_{make_tuple(container_reduce(low_lengths, math::multiplies{}, Number<1>{}))}
{
static_assert(LowerIndex::Size() == NDimLow, "wrong!");
}
......@@ -1188,11 +1188,11 @@ struct Merge_v2r2_magic_division
using LowerIndex = MultiIndex<NDimLow>;
using UpperIndex = MultiIndex<1>;
using LowLengthsScan = decltype(
container_reverse_exclusive_scan(LowLengths{}, math::multiplies_v2{}, Number<1>{}));
using LowLengthsScan =
decltype(container_reverse_exclusive_scan(LowLengths{}, math::multiplies{}, Number<1>{}));
using UpLengths =
decltype(make_tuple(container_reduce(LowLengths{}, math::multiplies_v2{}, Number<1>{})));
decltype(make_tuple(container_reduce(LowLengths{}, math::multiplies{}, Number<1>{})));
using LowLengthsScanMagicDivisorMultipiler = decltype(generate_tuple(
lambda_merge_generate_MagicDivision_calculate_magic_multiplier<LowLengthsScan>{},
......@@ -1213,14 +1213,14 @@ struct Merge_v2r2_magic_division
__host__ __device__ constexpr Merge_v2r2_magic_division(const LowLengths& low_lengths)
: low_lengths_{low_lengths},
low_lengths_scan_{
container_reverse_exclusive_scan(low_lengths, math::multiplies_v2{}, Number<1>{})},
container_reverse_exclusive_scan(low_lengths, math::multiplies{}, Number<1>{})},
low_lengths_scan_magic_divisor_multiplier_{generate_tuple(
[&](auto i) { return MagicDivision::CalculateMagicMultiplier(low_lengths_scan_[i]); },
Number<NDimLow>{})},
low_lengths_scan_magic_divisor_shift_{generate_tuple(
[&](auto i) { return MagicDivision::CalculateMagicShift(low_lengths_scan_[i]); },
Number<NDimLow>{})},
up_lengths_{make_tuple(container_reduce(low_lengths, math::multiplies_v2{}, Number<1>{}))}
up_lengths_{make_tuple(container_reduce(low_lengths, math::multiplies{}, Number<1>{}))}
{
static_assert(LowerIndex::Size() == NDimLow, "wrong!");
}
......@@ -1336,7 +1336,7 @@ struct UnMerge
using UpperIndex = MultiIndex<NDimUp>;
using UpLengthsScan =
decltype(container_reverse_exclusive_scan(UpLengths{}, math::multiplies_v2{}, Number<1>{}));
decltype(container_reverse_exclusive_scan(UpLengths{}, math::multiplies{}, Number<1>{}));
UpLengths up_lengths_;
UpLengthsScan up_lengths_scan_;
......@@ -1346,7 +1346,7 @@ struct UnMerge
__host__ __device__ constexpr UnMerge(const UpLengths& up_lengths)
: up_lengths_{up_lengths},
up_lengths_scan_{
container_reverse_exclusive_scan(up_lengths, math::multiplies_v2{}, Number<1>{})}
container_reverse_exclusive_scan(up_lengths, math::multiplies{}, Number<1>{})}
{
}
......
composable_kernel/include/tensor_description/tensor_adaptor.hpp
View file @ b62bf8c3
......@@ -64,7 +64,7 @@ struct TensorAdaptor
Number<ndim_top_>{});
// TODO: make container_reduce support tuple of Number and index_t
return container_reduce(lengths, math::multiplies_v2{}, Number<1>{});
return container_reduce(lengths, math::multiplies{}, Number<1>{});
}
template <index_t IDim>
......
composable_kernel/include/tensor_description/tensor_descriptor.hpp
View file @ b62bf8c3
......@@ -69,7 +69,7 @@ struct TensorDescriptor
Number<ndim_visible_>{});
// TODO: make container_reduce support tuple of Number and index_t
return container_reduce(lengths, math::multiplies_v2{}, Number<1>{});
return container_reduce(lengths, math::multiplies{}, Number<1>{});
}
template <index_t IDim>
......
composable_kernel/include/tensor_description/tensor_descriptor_helper.hpp
View file @ b62bf8c3
......@@ -38,7 +38,7 @@ __host__ __device__ constexpr auto calculate_element_space_size_impl(const Lengt
template <typename... Lengths,
typename... Strides,
typename enable_if<sizeof...(Lengths) == sizeof...(Strides), bool>::type = false>
__host__ __device__ constexpr auto make_naive_tensor_descriptor_v2(const Tuple<Lengths...>& lengths,
__host__ __device__ constexpr auto make_naive_tensor_descriptor(const Tuple<Lengths...>& lengths,
const Tuple<Strides...>& strides)
{
constexpr index_t N = sizeof...(Lengths);
......@@ -100,7 +100,7 @@ make_naive_tensor_descriptor_packed(const Tuple<Lengths...>& lengths)
constexpr auto visible_dim_hidden_ids = typename arithmetic_sequence_gen<1, N + 1, 1>::type{};
const auto element_space_size = container_reduce(lengths, math::multiplies_v2{}, Number<1>{});
const auto element_space_size = container_reduce(lengths, math::multiplies{}, Number<1>{});
return TensorDescriptor<remove_cv_t<decltype(transforms)>,
remove_cv_t<decltype(low_dim_hidden_idss)>,
......@@ -112,7 +112,7 @@ make_naive_tensor_descriptor_packed(const Tuple<Lengths...>& lengths)
template <typename... Lengths, typename Align>
__host__ __device__ constexpr auto
make_naive_tensor_descriptor_aligned_v2(const Tuple<Lengths...>& lengths, Align align)
make_naive_tensor_descriptor_aligned(const Tuple<Lengths...>& lengths, Align align)
{
constexpr auto I1 = Number<1>{};
......@@ -133,7 +133,7 @@ make_naive_tensor_descriptor_aligned_v2(const Tuple<Lengths...>& lengths, Align
else
{
return container_reduce(lengths,
math::multiplies_v2{},
math::multiplies{},
Number<stride_n_minus_2>{},
i + I1,
Number<N - 1>{},
......@@ -142,7 +142,7 @@ make_naive_tensor_descriptor_aligned_v2(const Tuple<Lengths...>& lengths, Align
},
Number<N>{});
return make_naive_tensor_descriptor_v2(lengths, strides);
return make_naive_tensor_descriptor(lengths, strides);
}
} // namespace ck
......
composable_kernel/include/tensor_operation/blockwise_tensor_slice_transfer.hpp
View file @ b62bf8c3
......@@ -143,7 +143,7 @@ struct BlockwiseTensorSliceTransfer_v4
private:
static constexpr auto thread_cluster_desc_ =
make_cluster_descriptor_v2(ThreadClusterLengths{}, ThreadClusterArrangeOrder{});
make_cluster_descriptor(ThreadClusterLengths{}, ThreadClusterArrangeOrder{});
using ThreadwiseTransfer =
ThreadwiseTensorSliceTransfer_v3<ThreadSliceLengths,
......
composable_kernel/include/tensor_operation/blockwise_tensor_slice_transfer_v2.hpp
View file @ b62bf8c3
......@@ -131,7 +131,7 @@ struct BlockwiseTensorSliceTransfer_v4r1
private:
static constexpr auto thread_cluster_desc_ =
make_cluster_descriptor_v2(ThreadClusterLengths{}, ThreadClusterArrangeOrder{});
make_cluster_descriptor(ThreadClusterLengths{}, ThreadClusterArrangeOrder{});
using ThreadwiseTransfer =
ThreadwiseTensorSliceTransfer_v3r1<ThreadSliceLengths,
......
composable_kernel/include/tensor_operation/gridwise_contraction_dlops_v1r2.hpp
View file @ b62bf8c3
......@@ -110,13 +110,13 @@ struct GridwiseContractionDlops_A_GK0_GM0_GM1_GK1_B_GK0_GN0_GN1_GK1_C_GM0_GM1_GN
// 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_naive_tensor_descriptor_aligned_v2(
constexpr auto a_block_desc_gk0_gm0_gm10_gm11_gk1 = make_naive_tensor_descriptor_aligned(
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_naive_tensor_descriptor_aligned_v2(
constexpr auto b_block_desc_gk0_gn0_gn10_gn11_gk1 = make_naive_tensor_descriptor_aligned(
make_tuple(Number<GK0PerBlock>{}, GN0, I1, Number<GN1PerBlockGN11>{}, GK1),
max_lds_align);
......@@ -248,10 +248,10 @@ struct GridwiseContractionDlops_A_GK0_GM0_GM1_GK1_B_GK0_GN0_GN1_GK1_C_GM0_GM1_GN
constexpr auto BN = GN0 * GN11;
constexpr auto BM1 =
Number<container_reduce(BM10BN10ThreadClusterBM10Xs{}, math::multiplies_v2{}, I1) *
Number<container_reduce(BM10BN10ThreadClusterBM10Xs{}, math::multiplies{}, I1) *
BM1PerThreadBM11>{};
constexpr auto BN1 =
Number<container_reduce(BM10BN10ThreadClusterBN10Xs{}, math::multiplies_v2{}, I1) *
Number<container_reduce(BM10BN10ThreadClusterBN10Xs{}, math::multiplies{}, I1) *
BN1PerThreadBN11>{};
constexpr auto BM0 = BM / BM1;
......@@ -354,24 +354,24 @@ struct GridwiseContractionDlops_A_GK0_GM0_GM1_GK1_B_GK0_GN0_GN1_GK1_C_GM0_GM1_GN
// 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_naive_tensor_descriptor_aligned_v2(
constexpr auto a_block_desc_gk0_gm0_gm10_gm11_gk1 = make_naive_tensor_descriptor_aligned(
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_naive_tensor_descriptor_aligned_v2(
constexpr auto b_block_desc_gk0_gn0_gn10_gn11_gk1 = make_naive_tensor_descriptor_aligned(
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_naive_tensor_descriptor_aligned_v2(
constexpr auto a_block_desc_gk0_bm_gk1 = make_naive_tensor_descriptor_aligned(
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_naive_tensor_descriptor_aligned_v2(
constexpr auto b_block_desc_gk0_bn_gk1 = make_naive_tensor_descriptor_aligned(
make_tuple(Number<GK0PerBlock>{}, GN0 * Number<GN1PerBlockGN11>{}, GK1), max_lds_align);
static_assert(a_block_desc_gk0_gm0_gm10_gm11_gk1.GetElementSpaceSize() ==
......
composable_kernel/include/tensor_operation/gridwise_gemm_dlops_v1r2.hpp
View file @ b62bf8c3
......@@ -166,12 +166,12 @@ struct GridwiseGemmDlops_km_kn_mn_v1r2
// A matrix in LDS memory, dst of blockwise copy
// be careful of LDS alignment
constexpr auto a_k_m_block_desc = make_naive_tensor_descriptor_aligned_v2(
constexpr auto a_k_m_block_desc = make_naive_tensor_descriptor_aligned(
make_tuple(Number<KPerBlock>{}, Number<MPerBlockM1>{}), max_lds_align);
// B matrix in LDS memory, dst of blockwise copy
// be careful of LDS alignment
constexpr auto b_k_n_block_desc = make_naive_tensor_descriptor_aligned_v2(
constexpr auto b_k_n_block_desc = make_naive_tensor_descriptor_aligned(
make_tuple(Number<KPerBlock>{}, Number<NPerBlockN1>{}), max_lds_align);
// LDS allocation for A and B: be careful of alignment
......@@ -351,22 +351,22 @@ struct GridwiseGemmDlops_km_kn_mn_v1r2
// A matrix in LDS memory, dst of blockwise copy
// be careful of LDS alignment
constexpr auto a_k_m_block_desc = make_naive_tensor_descriptor_aligned_v2(
constexpr auto a_k_m_block_desc = make_naive_tensor_descriptor_aligned(
make_tuple(Number<KPerBlock>{}, Number<MPerBlockM1>{}), max_lds_align);
// B matrix in LDS memory, dst of blockwise copy
// be careful of LDS alignment
constexpr auto b_k_n_block_desc = make_naive_tensor_descriptor_aligned_v2(
constexpr auto b_k_n_block_desc = make_naive_tensor_descriptor_aligned(
make_tuple(Number<KPerBlock>{}, Number<NPerBlockN1>{}), max_lds_align);
// A matrix in LDS memory, dst of blockwise copy
// be careful of LDS alignment
constexpr auto a_k_m0_m1_block_desc = make_naive_tensor_descriptor_aligned_v2(
constexpr auto a_k_m0_m1_block_desc = make_naive_tensor_descriptor_aligned(
make_tuple(Number<KPerBlock>{}, I1, Number<MPerBlockM1>{}), max_lds_align);
// B matrix in LDS memory, dst of blockwise copy
// be careful of LDS alignment
constexpr auto b_k_n0_n1_block_desc = make_naive_tensor_descriptor_aligned_v2(
constexpr auto b_k_n0_n1_block_desc = make_naive_tensor_descriptor_aligned(
make_tuple(Number<KPerBlock>{}, I1, Number<NPerBlockN1>{}), max_lds_align);
// A matrix blockwise copy
......
composable_kernel/include/tensor_operation/gridwise_gemm_dlops_v1r3.hpp
View file @ b62bf8c3
......@@ -163,12 +163,12 @@ struct GridwiseGemmDlops_km_kn_mn_v1r3
// TODO: check alignment
// A matrix in LDS memory, dst of blockwise copy
constexpr auto a_k_m_block_desc = make_naive_tensor_descriptor_aligned_v2(
constexpr auto a_k_m_block_desc = make_naive_tensor_descriptor_aligned(
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_naive_tensor_descriptor_aligned_v2(
constexpr auto b_k_n_block_desc = make_naive_tensor_descriptor_aligned(
make_tuple(Number<KPerBlock>{}, Number<NPerBlockN1>{}, K1), max_lds_align);
// TODO: check alignment
......@@ -274,10 +274,10 @@ struct GridwiseGemmDlops_km_kn_mn_v1r3
const auto N0 = N / N1;
constexpr auto M11 =
Number<container_reduce(M11N11ThreadClusterM110Xs{}, math::multiplies_v2{}, I1) *
Number<container_reduce(M11N11ThreadClusterM110Xs{}, math::multiplies{}, I1) *
M1PerThreadM111>{};
constexpr auto N11 =
Number<container_reduce(M11N11ThreadClusterN110Xs{}, math::multiplies_v2{}, I1) *
Number<container_reduce(M11N11ThreadClusterN110Xs{}, math::multiplies{}, I1) *
N1PerThreadN111>{};
constexpr auto M10 = M1 / M11;
......@@ -354,23 +354,23 @@ struct GridwiseGemmDlops_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_naive_tensor_descriptor_aligned_v2(
constexpr auto a_k0_m0_m1_k1_block_desc = make_naive_tensor_descriptor_aligned(
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_naive_tensor_descriptor_aligned_v2(
constexpr auto b_k0_n0_n1_k1_block_desc = make_naive_tensor_descriptor_aligned(
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_naive_tensor_descriptor_aligned_v2(
constexpr auto a_k0_m_k1_block_desc = make_naive_tensor_descriptor_aligned(
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_naive_tensor_descriptor_aligned_v2(
constexpr auto b_k0_n_k1_block_desc = make_naive_tensor_descriptor_aligned(
make_tuple(Number<KPerBlock>{}, Number<NPerBlockN1>{}, K1), max_lds_align);
static_assert(a_k0_m0_m1_k1_block_desc.GetElementSpaceSize() ==
......
composable_kernel/include/tensor_operation/gridwise_gemm_dlops_v2.hpp
View file @ b62bf8c3
......@@ -58,7 +58,7 @@ struct GridwiseGemmDlops_km_kn_mn_v3
// A matrix in LDS memory, dst of blockwise copy
// be careful of LDS alignment
constexpr auto a_e_k_desc = make_naive_tensor_descriptor_aligned_v2(
constexpr auto a_e_k_desc = make_naive_tensor_descriptor_aligned(
make_tuple(Number<E>{}, Number<KPerBlock>{}), max_lds_align);
// LDS allocation for A and B: be careful of alignment
......@@ -132,10 +132,10 @@ struct GridwiseGemmDlops_km_kn_mn_v3
// A matrix in LDS memory, dst of blockwise copy
// be careful of LDS alignment
constexpr auto a_e_k_block_desc = make_naive_tensor_descriptor_aligned_v2(
constexpr auto a_e_k_block_desc = make_naive_tensor_descriptor_aligned(
make_tuple(Number<EPerBlock>{}, Number<KPerBlock>{}), max_lds_align);
constexpr auto a_e_k_desc = make_naive_tensor_descriptor_aligned_v2(
constexpr auto a_e_k_desc = make_naive_tensor_descriptor_aligned(
make_tuple(Number<E>{}, Number<KPerBlock>{}), max_lds_align);
// B matrix in LDS memory, dst of blockwise copy
......
composable_kernel/include/tensor_operation/gridwise_gemm_xdlops_v2r3.hpp
View file @ b62bf8c3
......@@ -148,12 +148,12 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
// A matrix in LDS memory, dst of blockwise copy
// be careful of LDS alignment
constexpr auto a_k0_m_k1_block_desc = make_naive_tensor_descriptor_aligned_v2(
constexpr auto a_k0_m_k1_block_desc = make_naive_tensor_descriptor_aligned(
make_tuple(Number<KPerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
// B matrix in LDS memory, dst of blockwise copy
// be careful of LDS alignment
constexpr auto b_k0_n_k1_block_desc = make_naive_tensor_descriptor_aligned_v2(
constexpr auto b_k0_n_k1_block_desc = make_naive_tensor_descriptor_aligned(
make_tuple(Number<KPerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
// LDS allocation for A and B: be careful of alignment
......@@ -290,12 +290,12 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
// A matrix in LDS memory, dst of blockwise copy
// be careful of LDS alignment
constexpr auto a_k0_m_k1_block_desc = make_naive_tensor_descriptor_aligned_v2(
constexpr auto a_k0_m_k1_block_desc = make_naive_tensor_descriptor_aligned(
make_tuple(Number<KPerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
// B matrix in LDS memory, dst of blockwise copy
// be careful of LDS alignment
constexpr auto b_k0_n_k1_block_desc = make_naive_tensor_descriptor_aligned_v2(
constexpr auto b_k0_n_k1_block_desc = make_naive_tensor_descriptor_aligned(
make_tuple(Number<KPerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
// A matrix blockwise copy
......
composable_kernel/include/tensor_operation/threadwise_tensor_slice_transfer_v2.hpp
View file @ b62bf8c3
......@@ -91,12 +91,12 @@ struct ThreadwiseTensorSliceTransfer_v3r1
container_reverse_exclusive_scan(
container_reorder_given_new2old(src_vector_tensor_lengths,
SrcVectorTensorContiguousDimOrder{}),
math::multiplies_v2{},
math::multiplies{},
I1),
SrcVectorTensorContiguousDimOrder{});
constexpr auto src_vector_desc =
make_naive_tensor_descriptor_v2(sequence_to_tuple_of_number(src_vector_tensor_lengths),
make_naive_tensor_descriptor(sequence_to_tuple_of_number(src_vector_tensor_lengths),
sequence_to_tuple_of_number(src_vector_tensor_strides));
// access order and lengths
......@@ -259,12 +259,12 @@ struct ThreadwiseTensorSliceTransfer_v3r1
container_reverse_exclusive_scan(
container_reorder_given_new2old(dst_vector_tensor_lengths,
DstVectorTensorContiguousDimOrder{}),
math::multiplies_v2{},
math::multiplies{},
I1),
DstVectorTensorContiguousDimOrder{});
constexpr auto dst_vector_desc =
make_naive_tensor_descriptor_v2(sequence_to_tuple_of_number(dst_vector_tensor_lengths),
make_naive_tensor_descriptor(sequence_to_tuple_of_number(dst_vector_tensor_lengths),
sequence_to_tuple_of_number(dst_vector_tensor_strides));
// dst access order and lengths
......@@ -699,12 +699,12 @@ struct ThreadwiseTensorSliceTransfer_v4r1
container_reverse_exclusive_scan(
container_reorder_given_new2old(src_vector_tensor_lengths,
SrcVectorTensorContiguousDimOrder{}),
math::multiplies_v2{},
math::multiplies{},
I1),
SrcVectorTensorContiguousDimOrder{});
constexpr auto src_vector_desc =
make_naive_tensor_descriptor_v2(sequence_to_tuple_of_number(src_vector_tensor_lengths),
make_naive_tensor_descriptor(sequence_to_tuple_of_number(src_vector_tensor_lengths),
sequence_to_tuple_of_number(src_vector_tensor_strides));
// access order and lengths
......
composable_kernel/include/utility/math.hpp
View file @ b62bf8c3
......@@ -28,13 +28,7 @@ struct minus
__host__ __device__ constexpr T operator()(T a, T b) const { return a - b; }
};
template <typename T>
struct multiplies
{
__host__ __device__ constexpr T operator()(T a, T b) const { return a * b; }
};
struct multiplies_v2
{
template <typename A, typename B>
__host__ __device__ constexpr auto operator()(const A& a, const B& b) const
......
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