Skip to content

GitLab

Commit 8767acb2 authored by Chao Liu's avatar Chao Liu
Browse files

clean up

parent 35978330
Hide whitespace changes
Inline Side-by-side
Showing with 221 additions and 192 deletions
composable_kernel/include/tensor_operation/gridwise_gemm_xdlops_v3r2.hpp
View file @ 8767acb2
......@@ -17,8 +17,8 @@ template <typename GridwiseGemm,
typename FloatC,
typename AGridDesc_K0_M_K1,
typename BGridDesc_K0_N_K1,
typename CGridDescriptor_MBlock_MRepeat_MWaveMPerXdl_NBlock_NRepeat_NWaveNPerXdl,
typename C0GridDescriptor_MBlock_MRepeat_MWaveMPerXdl_NBlock_NRepeat_NWaveNPerXdl,
typename CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
typename C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
......@@ -35,10 +35,10 @@ __global__ void
const FloatC* __restrict__ p_c0_grid,
const AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1,
const BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1,
const CGridDescriptor_MBlock_MRepeat_MWaveMPerXdl_NBlock_NRepeat_NWaveNPerXdl
c_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl,
const C0GridDescriptor_MBlock_MRepeat_MWaveMPerXdl_NBlock_NRepeat_NWaveNPerXdl
c0_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl,
const CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
const C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const CElementwiseOperation c_element_op,
......@@ -54,8 +54,8 @@ __global__ void
p_shared,
a_grid_desc_k0_m_k1,
b_grid_desc_k0_n_k1,
c_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl,
c0_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl,
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
a_element_op,
b_element_op,
c_element_op,
......@@ -81,8 +81,8 @@ template <
index_t MPerXdl,
index_t NPerXdl,
index_t K1Value,
index_t MRepeat,
index_t NRepeat,
index_t MXdlPerWave,
index_t NXdlPerWave,
typename ABlockTransferThreadClusterLengths_K0_M_K1,
typename ABlockTransferThreadClusterArrangeOrder,
typename ABlockTransferSrcAccessOrder,
......@@ -99,9 +99,9 @@ template <
index_t BBlockTransferDstScalarPerVector_K1,
bool BThreadTransferSrcResetCoordinateAfterRun,
bool BBlockLdsExtraN,
index_t CShuffleMRepeatPerShuffle,
index_t CShuffleNRepeatPerShuffle,
typename CBlockTransferClusterLengths_MBlock_MRepeat_MWaveMPerXdl_NBlock_NRepeat_NWaveNPerXdl,
index_t CShuffleMXdlPerWavePerShuffle,
index_t CShuffleNXdlPerWavePerShuffle,
typename CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
index_t CBlockTransferScalarPerVector_NWaveNPerXdl>
struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
{
......@@ -117,8 +117,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
// K1 should be Number<...>
static constexpr auto K1 = Number<K1Value>{};
// TODO: need to calculate LDS usage for C shuffle
__host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
__host__ __device__ static constexpr auto GetABlockDescriptor_K0PerBlock_MPerBlock_K1()
{
constexpr auto max_lds_align = K1;
......@@ -137,6 +136,13 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
}
}();
return a_block_desc_k0_m_k1;
}
__host__ __device__ static constexpr auto GetBBlockDescriptor_K0PerBlock_NPerBlock_K1()
{
constexpr auto max_lds_align = K1;
// B matrix in LDS memory, dst of blockwise copy
constexpr auto b_block_desc_k0_n_k1 = [&]() {
if constexpr(BBlockLdsExtraN)
......@@ -152,14 +158,55 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
}
}();
return b_block_desc_k0_n_k1;
}
__host__ __device__ static constexpr auto
GetCBlockDescriptor_MBlock_NXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl()
{
constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
constexpr auto
c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl =
make_naive_tensor_descriptor_packed(
make_tuple(I1,
Number<CShuffleMXdlPerWavePerShuffle>{},
Number<MWave * MPerXdl>{},
I1,
Number<CShuffleNXdlPerWavePerShuffle>{},
Number<NWave * NPerXdl>{}));
return c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl;
}
// TODO: need to calculate LDS usage for C shuffle
__host__ __device__ static constexpr index_t GetSharedMemoryNumberOfByte()
{
// LDS allocation for A and B: be careful of alignment
constexpr auto a_block_space_size =
constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
constexpr auto b_block_desc_k0_n_k1 = GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
constexpr auto max_lds_align = K1;
constexpr auto a_block_space_size_aligned =
math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
constexpr auto b_block_space_size =
constexpr auto b_block_space_size_aligned =
math::integer_least_multiple(b_block_desc_k0_n_k1.GetElementSpaceSize(), max_lds_align);
return (a_block_space_size + b_block_space_size) * sizeof(FloatAB);
// LDS allocation for C shuffle in LDS
constexpr auto c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl =
GetCBlockDescriptor_MBlock_NXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl();
constexpr auto c_block_size =
c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
.GetElementSpaceSize();
return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
sizeof(FloatAB),
c_block_size * sizeof(FloatC));
}
// block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
......@@ -173,8 +220,8 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
static_assert(is_known_at_compile_time<remove_cv_t<decltype(K1)>>::value,
"wrong! K1 need to be known at compile-time");
static_assert((MPerBlock % (MPerXdl * MRepeat) == 0) &&
(NPerBlock % (NRepeat * NPerXdl)) == 0,
static_assert((MPerBlock % (MPerXdl * MXdlPerWave) == 0) &&
(NPerBlock % (NXdlPerWave * NPerXdl)) == 0,
"Invalid tuning param!");
const auto M = a_grid_desc_k0_m_k1.GetLength(I1);
......@@ -223,7 +270,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
template <typename CGridDesc_M_N_>
__host__ __device__ static constexpr auto
MakeCGridDescriptor_MBlock_MRepeat_MWaveMPerXdl_NBlock_NRepeat_NWaveNPerXdl(
MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
const CGridDesc_M_N_& c_grid_desc_m_n)
{
const auto M = c_grid_desc_m_n.GetLength(I0);
......@@ -232,20 +279,20 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
const auto MBlock = M / MPerBlock;
const auto NBlock = N / NPerBlock;
constexpr index_t MWave = MPerBlock / (MRepeat * MPerXdl);
constexpr index_t NWave = NPerBlock / (NRepeat * NPerXdl);
constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
const auto c_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl =
const auto c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl =
transform_tensor_descriptor(
c_grid_desc_m_n,
make_tuple(make_unmerge_transform(
make_tuple(MBlock, Number<MRepeat>{}, Number<MWave * MPerXdl>{})),
make_unmerge_transform(
make_tuple(NBlock, Number<NRepeat>{}, Number<NWave * NPerXdl>{}))),
make_tuple(make_unmerge_transform(make_tuple(
MBlock, Number<MXdlPerWave>{}, Number<MWave * MPerXdl>{})),
make_unmerge_transform(make_tuple(
NBlock, Number<NXdlPerWave>{}, Number<NWave * NPerXdl>{}))),
make_tuple(Sequence<0>{}, Sequence<1>{}),
make_tuple(Sequence<0, 1, 2>{}, Sequence<3, 4, 5>{}));
return c_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl;
return c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl;
}
// return block_id to C matrix tile idx (m0, n0) mapping
......@@ -283,14 +330,14 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
return c_blockid_to_m0_n0_block_cluster_adaptor;
}
using CGridDescriptor_MBlock_MRepeat_MWaveMPerXdl_NBlock_NRepeat_NWaveNPerXdl =
using CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl =
remove_cvref_t<decltype(
MakeCGridDescriptor_MBlock_MRepeat_MWaveMPerXdl_NBlock_NRepeat_NWaveNPerXdl(
MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
CGridDesc_M_N{}))>;
using C0GridDescriptor_MBlock_MRepeat_MWaveMPerXdl_NBlock_NRepeat_NWaveNPerXdl =
using C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl =
remove_cvref_t<decltype(
MakeCGridDescriptor_MBlock_MRepeat_MWaveMPerXdl_NBlock_NRepeat_NWaveNPerXdl(
MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
C0GridDesc_M_N{}))>;
using Block2CTileMap = remove_cvref_t<decltype(MakeBlock2CTileMap(CGridDesc_M_N{}, 1, 1))>;
......@@ -304,10 +351,10 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
void* __restrict__ p_shared,
const AGridDesc_K0_M_K1& a_grid_desc_k0_m_k1,
const BGridDesc_K0_N_K1& b_grid_desc_k0_n_k1,
const CGridDescriptor_MBlock_MRepeat_MWaveMPerXdl_NBlock_NRepeat_NWaveNPerXdl&
c_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl,
const C0GridDescriptor_MBlock_MRepeat_MWaveMPerXdl_NBlock_NRepeat_NWaveNPerXdl&
c0_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl,
const CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl&
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
const C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl&
c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
const AElementwiseOperation& a_element_op,
const BElementwiseOperation& b_element_op,
const CElementwiseOperation& c_element_op,
......@@ -319,11 +366,11 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
p_b_grid, b_grid_desc_k0_n_k1.GetElementSpaceSize());
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
p_c_grid,
c_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
.GetElementSpaceSize());
auto c0_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
p_c0_grid,
c0_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl
c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
.GetElementSpaceSize());
const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
......@@ -343,34 +390,10 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
constexpr auto max_lds_align = K1;
// A matrix in LDS memory, dst of blockwise copy
constexpr auto a_block_desc_k0_m_k1 = [&]() {
if constexpr(ABlockLdsExtraM)
{
return make_naive_tensor_descriptor(
make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1),
make_tuple(Number<MPerBlock + 1>{} * K1, K1, I1));
}
else
{
return make_naive_tensor_descriptor_aligned(
make_tuple(Number<K0PerBlock>{}, Number<MPerBlock>{}, K1), max_lds_align);
}
}();
constexpr auto a_block_desc_k0_m_k1 = GetABlockDescriptor_K0PerBlock_MPerBlock_K1();
// B matrix in LDS memory, dst of blockwise copy
constexpr auto b_block_desc_k0_n_k1 = [&]() {
if constexpr(BBlockLdsExtraN)
{
return make_naive_tensor_descriptor(
make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1),
make_tuple(Number<NPerBlock + 1>{} * K1, K1, I1));
}
else
{
return make_naive_tensor_descriptor_aligned(
make_tuple(Number<K0PerBlock>{}, Number<NPerBlock>{}, K1), max_lds_align);
}
}();
constexpr auto b_block_desc_k0_n_k1 = GetBBlockDescriptor_K0PerBlock_NPerBlock_K1();
// A matrix blockwise copy
auto a_blockwise_copy =
......@@ -448,21 +471,21 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
decltype(b_block_desc_k0_n_k1),
MPerXdl,
NPerXdl,
MRepeat,
NRepeat,
MXdlPerWave,
NXdlPerWave,
K1>{};
auto c_thread_buf = blockwise_gemm.GetCThreadBuffer();
// LDS allocation for A and B: be careful of alignment
constexpr auto a_block_space_size =
constexpr auto a_block_space_size_aligned =
math::integer_least_multiple(a_block_desc_k0_m_k1.GetElementSpaceSize(), max_lds_align);
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
static_cast<FloatAB*>(p_shared), a_block_desc_k0_m_k1.GetElementSpaceSize());
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
static_cast<FloatAB*>(p_shared) + a_block_space_size,
static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
b_block_desc_k0_n_k1.GetElementSpaceSize());
constexpr auto a_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
......@@ -516,12 +539,12 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
// shuffle C and write out
{
static_assert(MRepeat % CShuffleMRepeatPerShuffle == 0 &&
NRepeat % CShuffleNRepeatPerShuffle == 0,
static_assert(MXdlPerWave % CShuffleMXdlPerWavePerShuffle == 0 &&
NXdlPerWave % CShuffleNXdlPerWavePerShuffle == 0,
"wrong!");
constexpr index_t MWave = MPerBlock / (MRepeat * MPerXdl);
constexpr index_t NWave = NPerBlock / (NRepeat * NPerXdl);
constexpr index_t MWave = MPerBlock / (MXdlPerWave * MPerXdl);
constexpr index_t NWave = NPerBlock / (NXdlPerWave * NPerXdl);
// TODO: hacky, fix it!
constexpr auto c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2 =
......@@ -541,31 +564,27 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
constexpr auto M4 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I6);
constexpr auto N2 = c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2_tmp.GetLength(I7);
constexpr auto c_block_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl =
make_naive_tensor_descriptor_packed(make_tuple(I1,
Number<CShuffleMRepeatPerShuffle>{},
Number<MWave * MPerXdl>{},
I1,
Number<CShuffleNRepeatPerShuffle>{},
Number<NWave * NPerXdl>{}));
constexpr auto c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl =
GetCBlockDescriptor_MBlock_NXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl();
auto c_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
static_cast<FloatC*>(p_shared),
c_block_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl
c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
.GetElementSpaceSize());
constexpr auto c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2 = transform_tensor_descriptor(
c_block_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl,
make_tuple(make_freeze_transform(I0), // freeze mblock
make_pass_through_transform(
Number<CShuffleMRepeatPerShuffle>{}), // M0 (MRepeat) per shuffle
make_unmerge_transform(
make_tuple(M1, M2, M3, M4)), // M1 = MWave, M2 * M3 * M4 = MPerXdl
make_freeze_transform(I0), // freeze nblock
make_pass_through_transform(
Number<CShuffleNRepeatPerShuffle>{}), // N0 (NRepeat) per shuffle
make_unmerge_transform(
make_tuple(N1, N2))), // M1 = MWave, M2 * M3 * M4 = MPerXdl
c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
make_tuple(
make_freeze_transform(I0), // freeze mblock
make_pass_through_transform(
Number<CShuffleMXdlPerWavePerShuffle>{}), // M0 (MXdlPerWave) per shuffle
make_unmerge_transform(
make_tuple(M1, M2, M3, M4)), // M1 = MWave, M2 * M3 * M4 = MPerXdl
make_freeze_transform(I0), // freeze nblock
make_pass_through_transform(
Number<CShuffleNXdlPerWavePerShuffle>{}), // N0 (NXdlPerWave) per shuffle
make_unmerge_transform(
make_tuple(N1, N2))), // M1 = MWave, M2 * M3 * M4 = MPerXdl
make_tuple(Sequence<0>{},
Sequence<1>{},
Sequence<2>{},
......@@ -616,8 +635,8 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
ck::tensor_operation::element_wise::PassThrough,
Sequence<CShuffleMRepeatPerShuffle,
CShuffleNRepeatPerShuffle,
Sequence<CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
I1,
I1,
M2,
......@@ -646,53 +665,58 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
CElementwiseOperation, // ElementwiseOperation,
CGlobalMemoryDataOperation, // DstInMemOp,
Sequence<1,
CShuffleMRepeatPerShuffle,
CShuffleMXdlPerWavePerShuffle,
MWave * MPerXdl,
1,
CShuffleNRepeatPerShuffle,
CShuffleNXdlPerWavePerShuffle,
NWave * NPerXdl>, // BlockSliceLengths,
CBlockTransferClusterLengths_MBlock_MRepeat_MWaveMPerXdl_NBlock_NRepeat_NWaveNPerXdl,
CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
Sequence<0, 1, 2, 3, 4, 5>, // typename ThreadClusterArrangeOrder,
FloatC, // typename Src0Data,
FloatC, // typename Src1Data,
FloatC, // typename DstData,
decltype(c_block_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl),
decltype(c0_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl),
decltype(c_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl),
decltype(
c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
decltype(
c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
decltype(
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
Sequence<0, 1, 2, 3, 4, 5>, // typename DimAccessOrder,
5, // index_t VectorDim,
CBlockTransferScalarPerVector_NWaveNPerXdl, // index_t ScalarPerVector,
true, // bool ThreadTransferSrc0ResetCoordinateAfterRun,
false, // bool ThreadTransferSrc1ResetCoordinateAfterRun,
false> // bool ThreadTransferDstResetCoordinateAfterRun>
{c_block_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl,
{c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
make_multi_index(0, 0, 0, 0, 0, 0),
c0_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl,
c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
make_multi_index(block_work_idx[I0], 0, 0, block_work_idx[I1], 0, 0),
c_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl,
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
make_multi_index(block_work_idx[I0], 0, 0, block_work_idx[I1], 0, 0),
c_element_op};
constexpr auto mrepeat_forward_step =
make_multi_index(0, CShuffleMRepeatPerShuffle, 0, 0, 0, 0);
constexpr auto nrepeat_forward_step =
make_multi_index(0, 0, 0, 0, CShuffleNRepeatPerShuffle, 0);
constexpr auto nrepeat_backward_step =
make_multi_index(0, 0, 0, 0, -CShuffleNRepeatPerShuffle, 0);
constexpr auto mxdlperwave_forward_step =
make_multi_index(0, CShuffleMXdlPerWavePerShuffle, 0, 0, 0, 0);
constexpr auto nxdlperwave_forward_step =
make_multi_index(0, 0, 0, 0, CShuffleNXdlPerWavePerShuffle, 0);
constexpr auto nxdlperwave_backward_step =
make_multi_index(0, 0, 0, 0, -CShuffleNXdlPerWavePerShuffle, 0);
static_for<0, MRepeat, CShuffleMRepeatPerShuffle>{}([&](auto mrepeat_iter) {
constexpr auto mrepeat = mrepeat_iter;
static_for<0, MXdlPerWave, CShuffleMXdlPerWavePerShuffle>{}([&](auto mxdlperwave_iter) {
constexpr auto mxdlperwave = mxdlperwave_iter;
static_for<0, NRepeat, CShuffleNRepeatPerShuffle>{}([&](auto nrepeat_iter) {
constexpr bool nrepeat_forward_sweep =
(mrepeat % (2 * CShuffleMRepeatPerShuffle) == 0);
static_for<0,
NXdlPerWave,
CShuffleNXdlPerWavePerShuffle>{}([&](auto nxdlperwave_iter) {
constexpr bool nxdlperwave_forward_sweep =
(mxdlperwave % (2 * CShuffleMXdlPerWavePerShuffle) == 0);
constexpr index_t nrepeat_value =
nrepeat_forward_sweep
? nrepeat_iter
: (NRepeat - nrepeat_iter - CShuffleNRepeatPerShuffle);
constexpr index_t nxdlperwave_value =
nxdlperwave_forward_sweep
? nxdlperwave_iter
: (NXdlPerWave - nxdlperwave_iter - CShuffleNXdlPerWavePerShuffle);
constexpr auto nrepeat = Number<nrepeat_value>{};
constexpr auto nxdlperwave = Number<nxdlperwave_value>{};
// make sure it's safe to do ds_write
block_sync_lds();
......@@ -700,7 +724,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
// VGPR to LDS
c_thread_copy_vgpr_to_lds.Run(
c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2,
make_tuple(mrepeat, nrepeat, I0, I0, I0, I0, I0, I0),
make_tuple(mxdlperwave, nxdlperwave, I0, I0, I0, I0, I0, I0),
c_thread_buf,
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
c_block_buf);
......@@ -710,47 +734,47 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
// LDS to global
c_block_copy_lds_to_global.Run(
c_block_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl,
c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
c_block_buf,
c0_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl,
c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
c0_grid_buf,
c_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl,
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
c_grid_buf);
// move on nrepeat dimension
if constexpr(nrepeat_forward_sweep &&
(nrepeat < NRepeat - CShuffleNRepeatPerShuffle))
// move on nxdlperwave dimension
if constexpr(nxdlperwave_forward_sweep &&
(nxdlperwave < NXdlPerWave - CShuffleNXdlPerWavePerShuffle))
{
c_block_copy_lds_to_global.MoveSrc1SliceWindow(
c0_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl,
nrepeat_forward_step);
c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
nxdlperwave_forward_step);
c_block_copy_lds_to_global.MoveDstSliceWindow(
c_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl,
nrepeat_forward_step);
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
nxdlperwave_forward_step);
}
else if constexpr((!nrepeat_forward_sweep) && (nrepeat > 0))
else if constexpr((!nxdlperwave_forward_sweep) && (nxdlperwave > 0))
{
c_block_copy_lds_to_global.MoveSrc1SliceWindow(
c0_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl,
nrepeat_backward_step);
c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
nxdlperwave_backward_step);
c_block_copy_lds_to_global.MoveDstSliceWindow(
c_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl,
nrepeat_backward_step);
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
nxdlperwave_backward_step);
}
});
// move on mrepeat dimension
if constexpr(mrepeat < MRepeat - CShuffleMRepeatPerShuffle)
// move on mxdlperwave dimension
if constexpr(mxdlperwave < MXdlPerWave - CShuffleMXdlPerWavePerShuffle)
{
c_block_copy_lds_to_global.MoveSrc1SliceWindow(
c0_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl,
mrepeat_forward_step);
c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
mxdlperwave_forward_step);
c_block_copy_lds_to_global.MoveDstSliceWindow(
c_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl,
mrepeat_forward_step);
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
mxdlperwave_forward_step);
}
});
}
......
composable_kernel/include/tensor_operation/threadwise_tensor_slice_transfer_v6r1.hpp
View file @ 8767acb2
......@@ -54,6 +54,8 @@ struct ThreadwiseTensorSliceTransfer_v6r1
dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin)),
element_op_(element_op)
{
static_assert(SliceLengths::At(Number<VectorDim>{}) % ScalarPerVector == 0,
"wrong! cannot evenly divide");
}
__device__ void SetSrcSliceOrigin(const SrcDesc& src_desc, const Index& src_slice_origin_idx)
......
composable_kernel/include/tensor_operation/threadwise_tensor_slice_transfer_v6r2.hpp
View file @ 8767acb2
......@@ -62,6 +62,8 @@ struct ThreadwiseTensorSliceTransfer_v6r2
dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin)),
element_op_(element_op)
{
static_assert(SliceLengths::At(Number<VectorDim>{}) % ScalarPerVector == 0,
"wrong! cannot evenly divide");
}
__device__ void SetSrc0SliceOrigin(const Src0Desc& src0_desc,
......
composable_kernel/include/tensor_operation/threadwise_tensor_slice_transfer_v6r3.hpp
View file @ 8767acb2
......@@ -70,6 +70,8 @@ struct ThreadwiseTensorSliceTransfer_v6r3
dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin)),
element_op_(element_op)
{
static_assert(SliceLengths::At(Number<VectorDim>{}) % ScalarPerVector == 0,
"wrong! cannot evenly divide");
}
__device__ void SetSrc0SliceOrigin(const Src0Desc& src0_desc,
......
device_operation/device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instance.cpp
View file @ 8767acb2
......@@ -21,23 +21,23 @@ static constexpr auto MemorySet = ck::InMemoryDataOperationEnum_t::Set;
using device_conv2d_fwd_xdl_c_shuffle_bias_relu_nhwc_kyxc_nhwk_f16_instances = std::tuple<
// clang-format off
//##########################################################################################| InData| WeiData| OutData| AccData| In| Wei| Out| Out| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//##########################################################################################| Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| GlobalMemory| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeate| NRepeate| _MBlock_MRepeat_MWaveMPerXdl| ScalarPerVector|
//##########################################################################################| | | | | Operation| Operation| Operation| DataOperation| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NRepeat_NWaveNPerXdl| _NWaveNPerXdl|
//##########################################################################################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>
// | InData| WeiData| OutData| AccData| In| Wei| Out| Out| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
// | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| GlobalMemory| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
// | | | | | Operation| Operation| Operation| DataOperation| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
// | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, 256, 256, 128, 4, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, 128, 128, 128, 4, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, 256, 128, 128, 4, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, 128, 128, 64, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, 128, 64, 128, 4, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, 64, 64, 64, 4, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, 256, 128, 64, 4, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, 256, 64, 128, 4, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, 128, 128, 32, 4, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, AddRelu, MemorySet, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>
// clang-format on
>;
......
device_operation/device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instance.cpp
View file @ 8767acb2
......@@ -17,9 +17,8 @@ using S = ck::Sequence<Is...>;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using PassThrough_v2 = ck::tensor_operation::element_wise::PassThrough;
using device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances =
std::tuple<
// clang-format off
using device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances = std::tuple<
// clang-format off
// | InData| WeiData| OutData| AccData| In| Wei| Out| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
// | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
// | | | | | Operation| Operation| Operation| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
......@@ -37,8 +36,8 @@ using device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances =
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, 128, 32, 128, 4, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, 64, 64, 32, 4, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceConv2dFwdXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K< F16, F16, F16, F32, PassThrough, PassThrough, PassThrough, 64, 32, 64, 4, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>
// clang-format on
>;
// clang-format on
>;
void add_device_conv2d_fwd_xdl_c_shuffle_nhwc_kyxc_nhwk_f16_instances(
std::vector<DeviceConvFwdPtr<PassThrough, PassThrough, PassThrough_v2>>& device_conv_instances)
......
device_operation/include/device_conv2d_fwd_xdl_c_shuffle_bias_activation_nhwc_kyxc_nhwk.hpp
View file @ 8767acb2
......@@ -50,9 +50,9 @@ template <
ck::index_t BBlockTransferSrcScalarPerVector,
ck::index_t BBlockTransferDstScalarPerVector_K1,
bool BBlockLdsAddExtraN,
index_t CShuffleMRepeatPerShuffle,
index_t CShuffleNRepeatPerShuffle,
typename CBlockTransferClusterLengths_MBlock_MRepeat_MWaveMPerXdl_NBlock_NRepeat_NWaveNPerXdl,
index_t CShuffleMXdlPerWavePerShuffle,
index_t CShuffleNXdlPerWavePerShuffle,
typename CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
index_t CBlockTransferScalarPerVector_NWaveNPerXdl>
struct DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
: public DeviceConvFwdBiasActivation<InElementwiseOperation,
......@@ -262,9 +262,9 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X
BBlockTransferDstScalarPerVector_K1,
false, // BThreadTransferSrcResetCoordinateAfterRun,
BBlockLdsAddExtraN,
CShuffleMRepeatPerShuffle,
CShuffleNRepeatPerShuffle,
CBlockTransferClusterLengths_MBlock_MRepeat_MWaveMPerXdl_NBlock_NRepeat_NWaveNPerXdl,
CShuffleMXdlPerWavePerShuffle,
CShuffleNXdlPerWavePerShuffle,
CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
CBlockTransferScalarPerVector_NWaveNPerXdl>;
// Argument
......@@ -297,8 +297,8 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X
b_grid_desc_k0_n_k1_{},
c_grid_desc_m_n_{},
c0_grid_desc_m_n_{},
c_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl_{},
c0_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl_{},
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_{},
c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_{},
block_2_ctile_map_{},
M01_{M01},
N01_{N01},
......@@ -326,14 +326,14 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X
if(GridwiseGemm::CheckValidity(
a_grid_desc_k0_m_k1_, b_grid_desc_k0_n_k1_, c_grid_desc_m_n_, M01_, N01_))
{
c_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl_ =
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_ =
GridwiseGemm::
MakeCGridDescriptor_MBlock_MRepeat_MWaveMPerXdl_NBlock_NRepeat_NWaveNPerXdl(
MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
c_grid_desc_m_n_);
c0_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl_ =
c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_ =
GridwiseGemm::
MakeCGridDescriptor_MBlock_MRepeat_MWaveMPerXdl_NBlock_NRepeat_NWaveNPerXdl(
MakeCGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl(
c0_grid_desc_m_n_);
block_2_ctile_map_ = GridwiseGemm::MakeBlock2CTileMap(c_grid_desc_m_n_, M01, N01);
......@@ -350,11 +350,11 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X
CGridDesc_M_N c_grid_desc_m_n_;
C0GridDesc_M_N c0_grid_desc_m_n_;
typename GridwiseGemm::
CGridDescriptor_MBlock_MRepeat_MWaveMPerXdl_NBlock_NRepeat_NWaveNPerXdl
c_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl_;
CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_;
typename GridwiseGemm::
C0GridDescriptor_MBlock_MRepeat_MWaveMPerXdl_NBlock_NRepeat_NWaveNPerXdl
c0_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl_;
C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl
c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_;
typename GridwiseGemm::Block2CTileMap block_2_ctile_map_;
index_t M01_;
index_t N01_;
......@@ -414,10 +414,10 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X
remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
remove_reference_t<
typename GridwiseGemm::
CGridDescriptor_MBlock_MRepeat_MWaveMPerXdl_NBlock_NRepeat_NWaveNPerXdl>,
CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl>,
remove_reference_t<
typename GridwiseGemm::
C0GridDescriptor_MBlock_MRepeat_MWaveMPerXdl_NBlock_NRepeat_NWaveNPerXdl>,
C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl>,
InElementwiseOperation,
WeiElementwiseOperation,
OutElementwiseOperation,
......@@ -436,8 +436,8 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X
arg.p_c0_grid_,
arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl_,
arg.c0_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl_,
arg.c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_,
arg.c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_,
arg.in_element_op_,
arg.wei_element_op_,
arg.out_element_op_,
......@@ -453,10 +453,10 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X
remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
remove_reference_t<
typename GridwiseGemm::
CGridDescriptor_MBlock_MRepeat_MWaveMPerXdl_NBlock_NRepeat_NWaveNPerXdl>,
CGridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl>,
remove_reference_t<
typename GridwiseGemm::
C0GridDescriptor_MBlock_MRepeat_MWaveMPerXdl_NBlock_NRepeat_NWaveNPerXdl>,
C0GridDescriptor_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl>,
InElementwiseOperation,
WeiElementwiseOperation,
OutElementwiseOperation,
......@@ -475,8 +475,8 @@ struct DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X
arg.p_c0_grid_,
arg.a_grid_desc_k0_m_k1_,
arg.b_grid_desc_k0_n_k1_,
arg.c_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl_,
arg.c0_grid_desc_mblock_mrepeat_mwavemperxdl_nblock_nrepeat_nwavenperxdl_,
arg.c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_,
arg.c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl_,
arg.in_element_op_,
arg.wei_element_op_,
arg.out_element_op_,
......
example/5_conv2d_fwd_xdl_c_shuffle_bias_relu/conv2d_fwd_xdl_c_shuffle_bias_relu.cpp
View file @ 8767acb2
......@@ -36,11 +36,11 @@ static constexpr auto MemorySet = ck::InMemoryDataOperationEnum_t::Set;
using DeviceConvFwdInstance = ck::tensor_operation::device::
DeviceConv2dFwdXdl_C_Shuffle_Bias_Activation_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
// clang-format off
// | InData| WeiData| OutData| AccData| In| Wei| Out| Out| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
// | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| GlobalMemory| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MRepeate| NRepeate| _MBlock_MRepeat_MWaveMPerXdl| ScalarPerVector|
// | | | | | Operation| Operation| Operation| DataOperation| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NRepeat_NWaveNPerXdl| _NWaveNPerXdl|
// | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
<InDataType, WeiDataType, OutDataType, AccDataType, InElementOp, WeiElementOp, OutElementOp, MemorySet, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>;
// | InData| WeiData| OutData| AccData| In| Wei| Out| Out| Block| MPer| NPer| K0Per| K1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
// | Type| Type| Type| Type| Elementwise| Elementwise| Elementwise| GlobalMemory| Size| Block| Block| Block| | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
// | | | | | Operation| Operation| Operation| DataOperation| | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
// | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
<InDataType, WeiDataType, OutDataType, AccDataType, InElementOp, WeiElementOp, OutElementOp, MemorySet, 256, 128, 256, 4, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>;
// clang-format on
template <typename TIn,
......
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