Skip to content

GitLab

Unverified Commit cd167e49 authored by Chao Liu's avatar Chao Liu Committed by GitHub
Browse files

Compile for gfx908 and gfx90a (#130) 

* adding compilation for multiple targets

* fix build

* clean

* update Jekinsfile

* update readme

* update Jenkins

* use ck::half_t instead of ushort for bf16

* rename enum classes

* clean

* rename

* clean
parent ecf337ba
Hide whitespace changes
Inline Side-by-side
Showing with 112 additions and 1091 deletions
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r3.hpp
View file @ cd167e49
......@@ -132,7 +132,7 @@ template <index_t BlockSize,
typename FloatAB,
typename FloatAcc,
typename FloatC,
InMemoryDataOperationEnum_t CGlobalMemoryDataOperation,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
typename AGridDesc_K0_M_K1,
typename BGridDesc_K0_N_K1,
typename CGridDesc_M_N,
......@@ -426,11 +426,11 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
const CElementwiseOperation& c_element_op,
const Block2CTileMap& block_2_ctile_map)
{
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_grid, a_grid_desc_k0_m_k1.GetElementSpaceSize());
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_b_grid, b_grid_desc_k0_n_k1.GetElementSpaceSize());
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_c_grid, c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2.GetElementSpaceSize());
const auto K0 = a_grid_desc_k0_m_k1.GetLength(I0);
......@@ -460,7 +460,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
BlockwiseTensorSliceTransfer_v4r1<BlockSize,
AElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
InMemoryDataOperationEnum::Set,
Sequence<K0PerBlock, MPerBlock, K1>,
ABlockTransferThreadClusterLengths_K0_M_K1,
ABlockTransferThreadClusterArrangeOrder,
......@@ -491,7 +491,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
BlockwiseTensorSliceTransfer_v4r1<BlockSize,
BElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
InMemoryDataOperationEnum::Set,
Sequence<K0PerBlock, NPerBlock, K1>,
BBlockTransferThreadClusterLengths_K0_N_K1,
BBlockTransferThreadClusterArrangeOrder,
......@@ -543,10 +543,10 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
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>(
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<FloatAB*>(p_shared), a_block_desc_k0_m_k1.GetElementSpaceSize());
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
b_block_desc_k0_n_k1.GetElementSpaceSize());
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4.hpp
View file @ cd167e49
......@@ -59,7 +59,7 @@ template <index_t BlockSize,
typename FloatAB,
typename FloatAcc,
typename FloatC,
InMemoryDataOperationEnum_t CGlobalMemoryDataOperation,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
typename ABK0MK1GridDesc,
typename BBK0NK1GridDesc,
typename CMNGridDesc,
......@@ -316,11 +316,11 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4
const CElementwiseOperation& c_element_op,
const CBlockClusterAdaptor& c_block_cluster_adaptor)
{
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_grid, a_b_k0_m_k1_grid_desc.GetElementSpaceSize());
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_b_grid, b_b_k0_n_k1_grid_desc.GetElementSpaceSize());
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_c_grid, c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc.GetElementSpaceSize());
const auto K0 = a_b_k0_m_k1_grid_desc.GetLength(I1);
......@@ -410,7 +410,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4
BlockwiseTensorSliceTransfer_v4r1<BlockSize,
AElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
InMemoryDataOperationEnum::Set,
Sequence<1, K0PerBlock, MPerBlock, K1>,
ABlockTransferThreadClusterLengths_K0_M_K1,
ABlockTransferThreadClusterArrangeOrder,
......@@ -440,7 +440,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4
BlockwiseTensorSliceTransfer_v4r1<BlockSize,
BElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
InMemoryDataOperationEnum::Set,
Sequence<1, K0PerBlock, NPerBlock, K1>,
BBlockTransferThreadClusterLengths_K0_N_K1,
BBlockTransferThreadClusterArrangeOrder,
......@@ -497,9 +497,9 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4
constexpr auto a_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
constexpr auto b_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
p_a_block, a_k0_m_k1_block_desc.GetElementSpaceSize());
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
p_b_block, b_k0_n_k1_block_desc.GetElementSpaceSize());
// preload data into LDS
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v2r4r2.hpp
View file @ cd167e49
......@@ -61,7 +61,7 @@ template <index_t BlockSize,
typename FloatAB,
typename FloatAcc,
typename FloatC,
InMemoryDataOperationEnum_t CGlobalMemoryDataOperation,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
typename AGridDesc_B_K0_M_K1,
typename BGridDesc_B_K0_N_K1,
typename CMNGridDesc,
......@@ -305,11 +305,11 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
const CElementwiseOperation& c_element_op,
const CBlockClusterAdaptor& c_block_cluster_adaptor)
{
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_grid, a_b_k0_m_k1_grid_desc.GetElementSpaceSize());
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_b_grid, b_b_k0_n_k1_grid_desc.GetElementSpaceSize());
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_c_grid, c_grid_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
const auto K0 = a_b_k0_m_k1_grid_desc.GetLength(I1);
......@@ -399,7 +399,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
BlockwiseTensorSliceTransfer_v4r1<BlockSize,
AElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
InMemoryDataOperationEnum::Set,
Sequence<1, K0PerBlock, MPerBlock, K1>,
ABlockTransferThreadClusterLengths_K0_M_K1,
ABlockTransferThreadClusterArrangeOrder,
......@@ -429,7 +429,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
BlockwiseTensorSliceTransfer_v4r1<BlockSize,
BElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
InMemoryDataOperationEnum::Set,
Sequence<1, K0PerBlock, NPerBlock, K1>,
BBlockTransferThreadClusterLengths_K0_N_K1,
BBlockTransferThreadClusterArrangeOrder,
......@@ -486,9 +486,9 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
constexpr auto a_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
constexpr auto b_block_slice_copy_step = make_multi_index(0, K0PerBlock, 0, 0);
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
p_a_block, a_k0_m_k1_block_desc.GetElementSpaceSize());
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
p_b_block, b_k0_n_k1_block_desc.GetElementSpaceSize());
// preload data into LDS
......@@ -560,7 +560,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
constexpr auto c_block_desc_mblock_mperblock_nblock_nperblock =
GetCBlockDescriptor_MBlock_MPerBlock_NBlock_NPerBlock();
auto c_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
auto c_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<FloatC*>(p_shared_block),
c_block_desc_mblock_mperblock_nblock_nperblock.GetElementSpaceSize());
......@@ -632,7 +632,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4r2
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
7,
1,
InMemoryDataOperationEnum_t::Set,
InMemoryDataOperationEnum::Set,
1,
true>{
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r1.hpp
View file @ cd167e49
......@@ -64,7 +64,7 @@ template <
typename FloatAcc,
typename FloatCShuffle,
typename FloatC,
InMemoryDataOperationEnum_t CGlobalMemoryDataOperation,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
typename AGridDesc_AK0_M_AK1,
typename BGridDesc_BK0_N_BK1,
typename CGridDesc_M_N,
......@@ -369,11 +369,11 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
const CElementwiseOperation& c_element_op,
const Block2CTileMap& block_2_ctile_map)
{
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_grid, a_grid_desc_ak0_m_ak1.GetElementSpaceSize());
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_b_grid, b_grid_desc_bk0_n_bk1.GetElementSpaceSize());
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_c_grid,
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
.GetElementSpaceSize());
......@@ -403,7 +403,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
BlockwiseTensorSliceTransfer_v4r1<BlockSize,
AElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
InMemoryDataOperationEnum::Set,
Sequence<AK0, MPerBlock, AK1>,
ABlockTransferThreadClusterLengths_AK0_M_AK1,
ABlockTransferThreadClusterArrangeOrder,
......@@ -434,7 +434,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
BlockwiseTensorSliceTransfer_v4r1<BlockSize,
BElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
InMemoryDataOperationEnum::Set,
Sequence<BK0, NPerBlock, BK1>,
BBlockTransferThreadClusterLengths_BK0_N_BK1,
BBlockTransferThreadClusterArrangeOrder,
......@@ -488,10 +488,10 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
constexpr auto a_block_space_size_aligned = math::integer_least_multiple(
a_block_desc_ak0_m_ak1.GetElementSpaceSize(), max_lds_align);
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<FloatAB*>(p_shared), a_block_desc_ak0_m_ak1.GetElementSpaceSize());
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
b_block_desc_bk0_n_bk1.GetElementSpaceSize());
......@@ -567,7 +567,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
constexpr auto c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl =
GetCBlockDescriptor_MBlock_NXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl();
auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<FloatCShuffle*>(p_shared),
c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
.GetElementSpaceSize());
......@@ -644,7 +644,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
7,
1,
InMemoryDataOperationEnum_t::Set,
InMemoryDataOperationEnum::Set,
1,
true>{
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r2.hpp
View file @ cd167e49
......@@ -68,7 +68,7 @@ template <
typename FloatAB,
typename FloatAcc,
typename FloatC,
InMemoryDataOperationEnum_t CGlobalMemoryDataOperation,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
typename AGridDesc_K0_M_K1,
typename BGridDesc_K0_N_K1,
typename CGridDesc_M_N,
......@@ -382,15 +382,15 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
const CElementwiseOperation& c_element_op,
const Block2CTileMap& block_2_ctile_map)
{
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_grid, a_grid_desc_k0_m_k1.GetElementSpaceSize());
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_b_grid, b_grid_desc_k0_n_k1.GetElementSpaceSize());
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_c_grid,
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
.GetElementSpaceSize());
auto c0_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
auto c0_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_c0_grid,
c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
.GetElementSpaceSize());
......@@ -422,7 +422,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
BlockwiseTensorSliceTransfer_v4r1<BlockSize,
AElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
InMemoryDataOperationEnum::Set,
Sequence<K0PerBlock, MPerBlock, K1>,
ABlockTransferThreadClusterLengths_K0_M_K1,
ABlockTransferThreadClusterArrangeOrder,
......@@ -453,7 +453,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
BlockwiseTensorSliceTransfer_v4r1<BlockSize,
BElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
InMemoryDataOperationEnum::Set,
Sequence<K0PerBlock, NPerBlock, K1>,
BBlockTransferThreadClusterLengths_K0_N_K1,
BBlockTransferThreadClusterArrangeOrder,
......@@ -505,10 +505,10 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
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>(
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<FloatAB*>(p_shared), a_block_desc_k0_m_k1.GetElementSpaceSize());
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
b_block_desc_k0_n_k1.GetElementSpaceSize());
......@@ -582,7 +582,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
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>(
auto c_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<FloatC*>(p_shared),
c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
.GetElementSpaceSize());
......@@ -661,7 +661,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r2
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
7,
1,
InMemoryDataOperationEnum_t::Set,
InMemoryDataOperationEnum::Set,
1,
true>{
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
......
include/ck/tensor_operation/gpu/grid/gridwise_gemm_xdlops_v3r3.hpp
View file @ cd167e49
......@@ -74,7 +74,7 @@ template <
typename FloatAB,
typename FloatAcc,
typename FloatC,
InMemoryDataOperationEnum_t CGlobalMemoryDataOperation,
InMemoryDataOperationEnum CGlobalMemoryDataOperation,
typename AGridDesc_K0_M_K1,
typename BGridDesc_K0_N_K1,
typename CGridDesc_M_N,
......@@ -397,19 +397,19 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3
const CElementwiseOperation& c_element_op,
const Block2CTileMap& block_2_ctile_map)
{
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
const auto a_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_a_grid, a_grid_desc_k0_m_k1.GetElementSpaceSize());
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
const auto b_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_b_grid, b_grid_desc_k0_n_k1.GetElementSpaceSize());
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
auto c_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_c_grid,
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
.GetElementSpaceSize());
auto c0_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
auto c0_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_c0_grid,
c0_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
.GetElementSpaceSize());
auto c1_grid_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
auto c1_grid_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_c1_grid,
c1_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
.GetElementSpaceSize());
......@@ -441,7 +441,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3
BlockwiseTensorSliceTransfer_v4r1<BlockSize,
AElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
InMemoryDataOperationEnum::Set,
Sequence<K0PerBlock, MPerBlock, K1>,
ABlockTransferThreadClusterLengths_K0_M_K1,
ABlockTransferThreadClusterArrangeOrder,
......@@ -471,7 +471,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3
BlockwiseTensorSliceTransfer_v4r1<BlockSize,
BElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
InMemoryDataOperationEnum::Set,
Sequence<K0PerBlock, NPerBlock, K1>,
BBlockTransferThreadClusterLengths_K0_N_K1,
BBlockTransferThreadClusterArrangeOrder,
......@@ -522,10 +522,10 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3
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>(
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<FloatAB*>(p_shared), a_block_desc_k0_m_k1.GetElementSpaceSize());
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<FloatAB*>(p_shared) + a_block_space_size_aligned,
b_block_desc_k0_n_k1.GetElementSpaceSize());
......@@ -599,7 +599,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3
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>(
auto c_block_buf = make_dynamic_buffer<AddressSpaceEnum::Lds>(
static_cast<FloatC*>(p_shared),
c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
.GetElementSpaceSize());
......@@ -678,7 +678,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r3
Sequence<0, 1, 2, 3, 4, 5, 6, 7>,
7,
1,
InMemoryDataOperationEnum_t::Set,
InMemoryDataOperationEnum::Set,
1,
true>{
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
......
include/ck/tensor_operation/gpu/grid/gridwise_set_buffer_value.hpp
View file @ cd167e49
......@@ -45,13 +45,13 @@ __global__ void kernel_buffer_set_value(const Grid1dBufferDescType grid_1d_buffe
const index_t thread_global_id = block_global_id * BlockSize + thread_local_id;
StaticBuffer<AddressSpaceEnum_t::Vgpr, DataType, 1, true> value_buf;
StaticBuffer<AddressSpaceEnum::Vgpr, DataType, 1, true> value_buf;
value_buf(I0) = value;
constexpr auto val_buff_desc = make_naive_tensor_descriptor_packed(make_tuple(Number<1>{}));
auto global_buf = make_dynamic_buffer<AddressSpaceEnum_t::Global>(
auto global_buf = make_dynamic_buffer<AddressSpaceEnum::Global>(
p_global, grid_1d_buffer_desc.GetElementSpaceSize());
if(thread_global_id < grid_1d_buffer_desc.GetElementSize())
......@@ -65,7 +65,7 @@ __global__ void kernel_buffer_set_value(const Grid1dBufferDescType grid_1d_buffe
Sequence<0>,
0,
1,
InMemoryDataOperationEnum_t::Set,
InMemoryDataOperationEnum::Set,
1,
true>(
grid_1d_buffer_desc, make_multi_index(thread_global_id), PassThroughOp{});
......
include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer.hpp
View file @ cd167e49
......@@ -56,7 +56,7 @@ template <typename SrcData,
typename DimAccessOrder,
index_t DstVectorDim,
index_t DstScalarPerVector,
InMemoryDataOperationEnum_t DstInMemOp,
InMemoryDataOperationEnum DstInMemOp,
index_t DstScalarStrideInVector,
bool DstResetCoordinateAfterRun,
typename enable_if<SrcDesc::IsKnownAtCompileTime(), bool>::type = false>
......@@ -407,7 +407,7 @@ struct ThreadwiseTensorSliceTransfer_v2
// 3. src_slice_origin and dst_slice_origin are not known at compile-time,
// 4. Use thread buffer
template <typename SliceLengths,
InMemoryDataOperationEnum_t DstInMemOp,
InMemoryDataOperationEnum DstInMemOp,
typename SrcData,
typename DstData,
typename SrcDesc,
......@@ -464,8 +464,8 @@ struct ThreadwiseTensorSliceTransfer_v3
__device__ void
RunRead(const SrcDesc& src_desc, const SrcBuffer& src_buf, const SrcStepHacks& src_step_hacks)
{
static_assert(SrcBuffer::GetAddressSpace() == AddressSpaceEnum_t::Global or
SrcBuffer::GetAddressSpace() == AddressSpaceEnum_t::Lds,
static_assert(SrcBuffer::GetAddressSpace() == AddressSpaceEnum::Global or
SrcBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
"wrong!");
static_assert(
......@@ -621,8 +621,8 @@ struct ThreadwiseTensorSliceTransfer_v3
__device__ void
RunWrite(const DstDesc& dst_desc, DstBuffer& dst_buf, const DstStepHacks& dst_step_hacks)
{
static_assert(DstBuffer::GetAddressSpace() == AddressSpaceEnum_t::Global or
DstBuffer::GetAddressSpace() == AddressSpaceEnum_t::Lds,
static_assert(DstBuffer::GetAddressSpace() == AddressSpaceEnum::Global or
DstBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
"wrong!");
static_assert(
......@@ -979,7 +979,7 @@ struct ThreadwiseTensorSliceTransfer_v3
static constexpr auto buffer_size_ = buffer_desc_.GetElementSpaceSize();
StaticBuffer<AddressSpaceEnum_t::Vgpr, SrcData, buffer_size_, true> buffer_;
StaticBuffer<AddressSpaceEnum::Vgpr, SrcData, buffer_size_, true> buffer_;
SrcCoord src_coord_;
DstCoord dst_coord_;
......
include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v1r4.hpp deleted 100644 → 0
View file @ ecf337ba
#ifndef CK_THREADWISE_TENSOR_SLICE_TRANSFER_V1R4_HPP
#define CK_THREADWISE_TENSOR_SLICE_TRANSFER_V1R4_HPP
#include "common_header.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
namespace ck {
// Do following things to avoid "alloca" in LLVM-IR, which would cause scratch memory
// and sometimes useless instructions:
// 1. Don't save a reference to tensor descriptor in class, pass in tensor descriptor as argument
// instead
// 2. Don't construct a new tensor coordinate everytime when using it, update and reuse the same
// tensor coordinate instead
// 3. Don't use a pointer to VGPR buffer, use vector instead
// WARNING!!!!!!: this logic is only correct if DstScalarPerVector=1
// TODO: fix this
// Assume:
// 1. src:
// 1. SrcDesc is known at compile-time
// 2. SrcBuffer is StaticBuffer
// 3. SrcSliceOrginIdx is known at compile-time
// 2. dst:
// 1. DstDesc is not known at compile-time
// 2. DstBuffer is DynamicBuffer
// 3. DstSliceOrginIdx is not known at compile time
template <typename SrcData,
typename DstData,
typename SrcDesc,
typename DstDesc,
typename Dst0Desc, // this is really one of sources, but it has same shape as DstDesc
typename Dst1Desc, // this is really one of sources, but it has same shape as DstDesc
typename DstElementwiseOperation,
typename SliceLengths,
typename DimAccessOrder,
index_t DstVectorDim,
index_t DstScalarPerVector,
InMemoryDataOperationEnum_t DstInMemOp,
index_t DstScalarStrideInVector,
bool DstResetCoordinateAfterRun,
typename enable_if<SrcDesc::IsKnownAtCompileTime(), bool>::type = false>
struct ThreadwiseTensorSliceTransfer_v1r4
{
static constexpr index_t nDim = SliceLengths::Size();
using Index = MultiIndex<nDim>;
using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
using Dst0Coord = decltype(make_tensor_coordinate(Dst0Desc{}, Index{}));
using Dst1Coord = decltype(make_tensor_coordinate(Dst1Desc{}, Index{}));
using DstCoordStep = decltype(make_tensor_coordinate_step(DstDesc{}, Index{}));
using Dst0CoordStep = decltype(make_tensor_coordinate_step(Dst0Desc{}, Index{}));
using Dst1CoordStep = decltype(make_tensor_coordinate_step(Dst1Desc{}, Index{}));
__device__ constexpr ThreadwiseTensorSliceTransfer_v1r4(
const DstDesc& dst_desc,
const Dst0Desc& dst0_desc,
const Dst1Desc& dst1_desc,
const Index& dst_slice_origin_idx,
const DstElementwiseOperation& dst_element_op)
: dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin_idx)),
dst0_coord_(make_tensor_coordinate(dst0_desc, dst_slice_origin_idx)),
dst1_coord_(make_tensor_coordinate(dst1_desc, dst_slice_origin_idx)),
dst_element_op_{dst_element_op}
{
static_assert(SrcDesc::IsKnownAtCompileTime(),
"wrong! SrcDesc need to known at compile-time");
}
__device__ void SetDstSliceOrigin(const DstDesc& dst_desc, const Index& dst_slice_origin_idx)
{
dst_coord_ = make_tensor_coordinate(dst_desc, dst_slice_origin_idx);
}
template <typename SrcSliceOriginIdx,
typename SrcBuffer,
typename DstBuffer,
typename Dst0Buffer,
typename Dst1Buffer,
typename DstStepHacks,
typename Dst0StepHacks,
typename Dst1StepHacks>
__device__ void Run(const SrcDesc&,
const SrcSliceOriginIdx&,
const SrcBuffer& src_buf,
const DstDesc& dst_desc,
DstBuffer& dst_buf,
const DstStepHacks& dst_step_hacks,
const Dst0Desc& dst0_desc,
const Dst0Buffer& dst0_buf,
const Dst0StepHacks& dst0_step_hacks,
const Dst1Desc& dst1_desc,
const Dst1Buffer& dst1_buf,
const Dst1StepHacks& dst1_step_hacks)
{
static_assert(SrcDesc::IsKnownAtCompileTime(),
"wrong! SrcDesc need to known at compile-time");
static_assert(is_known_at_compile_time<remove_cvref_t<SrcSliceOriginIdx>>::value,
"wrong! SrcSliceOrigin need to known at compile-time");
static_assert(SrcBuffer::IsStaticBuffer(), "wrong! SrcBuffer need to be StaticBuffer");
// SrcDesc and src_slice_origin_idx are known at compile-time
constexpr auto src_desc = remove_cvref_t<SrcDesc>{};
constexpr auto src_slice_origin_idx = to_multi_index(SrcSliceOriginIdx{});
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
// scalar per access on each dim
// TODO: don't use lambda_scalar_per_access
constexpr auto dst_scalar_per_access = generate_sequence(
detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
constexpr auto access_lengths = SliceLengths{} / dst_scalar_per_access;
constexpr auto dim_access_order = DimAccessOrder{};
constexpr auto ordered_access_lengths =
container_reorder_given_new2old(access_lengths, dim_access_order);
// make forward steps: dst
const auto dst_forward_steps = generate_tuple(
[&](auto i) {
Index forward_step_idx;
static_for<0, nDim, 1>{}([&](auto j) {
forward_step_idx(j) = (i.value == j.value) ? dst_scalar_per_access[i] : 0;
});
return make_tensor_coordinate_step(
dst_desc, forward_step_idx, dst_step_hacks[I0][i]);
},
Number<nDim>{});
// make forward steps: dst0
// WARNING!!!!!!: this logic is only correct if dst/dst0/dst1 can use the same
// DstScalarPerVector
// TODO: fix this
const auto dst0_forward_steps = generate_tuple(
[&](auto i) {
Index forward_step_idx;
static_for<0, nDim, 1>{}([&](auto j) {
forward_step_idx(j) = (i.value == j.value) ? dst_scalar_per_access[i] : 0;
});
return make_tensor_coordinate_step(
dst0_desc, forward_step_idx, dst0_step_hacks[I0][i]);
},
Number<nDim>{});
// make forward steps: dst1
// WARNING!!!!!!: this logic is only correct if dst/dst0/dst1 can use the same
// DstScalarPerVector
// TODO: fix this
const auto dst1_forward_steps = generate_tuple(
[&](auto i) {
Index forward_step_idx;
static_for<0, nDim, 1>{}([&](auto j) {
forward_step_idx(j) = (i.value == j.value) ? dst_scalar_per_access[i] : 0;
});
return make_tensor_coordinate_step(
dst1_desc, forward_step_idx, dst1_step_hacks[I0][i]);
},
Number<nDim>{});
// make backward steps: dst
const auto dst_backward_steps = generate_tuple(
[&](auto i) {
Index backward_step_idx;
static_for<0, nDim, 1>{}([&](auto j) {
backward_step_idx(j) = (i.value == j.value) ? -dst_scalar_per_access[i] : 0;
});
return make_tensor_coordinate_step(
dst_desc, backward_step_idx, dst_step_hacks[I1][i]);
},
Number<nDim>{});
// make backward steps: dst0
// WARNING!!!!!!: this logic is only correct if dst/dst0/dst1 can use the same
// DstScalarPerVector
// TODO: fix this
const auto dst0_backward_steps = generate_tuple(
[&](auto i) {
Index backward_step_idx;
static_for<0, nDim, 1>{}([&](auto j) {
backward_step_idx(j) = (i.value == j.value) ? -dst_scalar_per_access[i] : 0;
});
return make_tensor_coordinate_step(
dst0_desc, backward_step_idx, dst0_step_hacks[I1][i]);
},
Number<nDim>{});
// make backward steps: dst1
// WARNING!!!!!!: this logic is only correct if dst/dst0/dst1 can use the same
// DstScalarPerVector
// TODO: fix this
const auto dst1_backward_steps = generate_tuple(
[&](auto i) {
Index backward_step_idx;
static_for<0, nDim, 1>{}([&](auto j) {
backward_step_idx(j) = (i.value == j.value) ? -dst_scalar_per_access[i] : 0;
});
return make_tensor_coordinate_step(
dst1_desc, backward_step_idx, dst1_step_hacks[I1][i]);
},
Number<nDim>{});
// loop over tensor and copy
static_ford<decltype(ordered_access_lengths)>{}([&](auto ordered_access_idx) {
// judge move forward or move backward
constexpr auto forward_sweep = [&]() {
StaticallyIndexedArray<bool, nDim> forward_sweep_;
forward_sweep_(I0) = true;
static_for<1, nDim, 1>{}([&](auto i) {
index_t tmp = ordered_access_idx[I0];
static_for<1, i, 1>{}([&](auto j) {
tmp = tmp * ordered_access_lengths[j] + ordered_access_idx[j];
});
forward_sweep_(i) = tmp % 2 == 0;
});
return forward_sweep_;
}();
// calculate dst data index
constexpr auto dst_data_idx = [&]() {
Index ordered_idx;
static_for<0, nDim, 1>{}([&](auto i) {
ordered_idx(i) = forward_sweep[i]
? ordered_access_idx[i]
: ordered_access_lengths[i] - 1 - ordered_access_idx[i];
});
return container_reorder_given_old2new(ordered_idx, dim_access_order) *
dst_scalar_per_access;
}();
typename vector_type_maker<DstData, DstScalarPerVector>::type dst_vector;
using dst_vector_t =
typename vector_type_maker<DstData, DstScalarPerVector>::type::type;
// load dst0 and dst1 and apply elementwise operation
{
// WARNING!!!!!!: this logic is only correct if DstScalarPerVector=1
// TODO: fix this
static_assert(DstScalarPerVector == 1, "wrong!");
// copy data from src_buf into dst_vector_src_data
constexpr index_t src_offset =
src_desc.CalculateOffset(src_slice_origin_idx + dst_data_idx);
const SrcData src_v = src_buf[Number<src_offset>{}];
// load dst0 and dst1
const bool is_dst0_valid =
coordinate_has_valid_offset_assuming_visible_index_is_valid(dst0_desc,
dst0_coord_);
const bool is_dst1_valid =
coordinate_has_valid_offset_assuming_visible_index_is_valid(dst1_desc,
dst1_coord_);
const DstData dst0_v =
dst0_buf.template Get<DstData>(dst0_coord_.GetOffset(), is_dst0_valid);
const DstData dst1_v =
dst1_buf.template Get<DstData>(dst1_coord_.GetOffset(), is_dst1_valid);
#if !CK_WORKAROUND_SWDEV_XXXXXX_THREAD_WISE_COPY_V1R4_TYPE_CONVERT_ISSUE
// apply element-wise operation in SrcData type
const SrcData dst_v = dst_element_op_(
src_v, type_convert<SrcData>(dst0_v), type_convert<SrcData>(dst1_v));
// apply type convert
dst_vector.template AsType<DstData>()(Number<0>{}) = type_convert<DstData>(dst_v);
#else
// apply element-wise operation in DstData type
DstData dst_v;
dst_element_op_(dst_v, src_v, dst0_v, dst1_v);
dst_vector.template AsType<DstData>()(Number<0>{}) = dst_v;
#endif
}
const bool is_dst_valid =
coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_desc, dst_coord_);
// copy data from dst_vector into dst_buf
if constexpr(DstInMemOp == InMemoryDataOperationEnum_t::Set)
{
dst_buf.template Set<dst_vector_t>(
dst_coord_.GetOffset(),
is_dst_valid,
dst_vector.template AsType<dst_vector_t>()[Number<0>{}]);
}
else if constexpr(DstInMemOp == InMemoryDataOperationEnum_t::AtomicAdd)
{
dst_buf.template AtomicAdd<dst_vector_t>(
dst_coord_.GetOffset(),
is_dst_valid,
dst_vector.template AsType<dst_vector_t>()[Number<0>{}]);
}
else if constexpr(DstInMemOp == InMemoryDataOperationEnum_t::Add)
{
typename vector_type_maker<DstData, DstScalarPerVector>::type tmp;
tmp.template AsType<dst_vector_t>()(Number<0>{}) =
dst_buf.template Get<dst_vector_t>(dst_coord_.GetOffset(), is_dst_valid);
static_for<0, DstScalarPerVector, 1>{}([&](auto t) {
dst_vector.template AsType<DstData>()(t) += tmp.template AsType<DstData>()[t];
});
dst_buf.template Set<dst_vector_t>(
dst_coord_.GetOffset(),
is_dst_valid,
dst_vector.template AsType<dst_vector_t>()[Number<0>{}]);
}
constexpr auto move_on_dim = [&]() constexpr
{
StaticallyIndexedArray<bool, nDim> move_on_dim_;
static_for<0, nDim, 1>{}([&](auto i) {
move_on_dim_(i) = ordered_access_idx[i] < ordered_access_lengths[i] - 1;
static_for<i + 1, nDim, 1>{}([&](auto j) {
move_on_dim_(i) &= ordered_access_idx[j] == ordered_access_lengths[j] - 1;
});
});
return move_on_dim_;
}
();
// move
static_for<0, nDim, 1>{}([&](auto i) {
if constexpr(move_on_dim[i])
{
if constexpr(forward_sweep[i])
{
move_tensor_coordinate(
dst_desc, dst_coord_, dst_forward_steps[dim_access_order[i]]);
// dst0
move_tensor_coordinate(
dst0_desc, dst0_coord_, dst0_forward_steps[dim_access_order[i]]);
// dst1
move_tensor_coordinate(
dst1_desc, dst1_coord_, dst1_forward_steps[dim_access_order[i]]);
}
else
{
move_tensor_coordinate(
dst_desc, dst_coord_, dst_backward_steps[dim_access_order[i]]);
// dst0
move_tensor_coordinate(
dst0_desc, dst0_coord_, dst0_backward_steps[dim_access_order[i]]);
// dst1
move_tensor_coordinate(
dst1_desc, dst1_coord_, dst1_backward_steps[dim_access_order[i]]);
}
}
});
});
// move dst coordinate back to slice origin (or not)
if constexpr(DstResetCoordinateAfterRun)
{
const auto dst_reset_step =
make_tensor_coordinate_step(dst_desc, GetDstCoordinateResetStep());
move_tensor_coordinate(dst_desc, dst_coord_, dst_reset_step);
}
}
template <typename SrcSliceOriginIdx,
typename SrcBuffer,
typename DstBuffer,
typename Dst0Buffer,
typename Dst1Buffer>
__device__ void Run(const SrcDesc&,
const SrcSliceOriginIdx&,
const SrcBuffer& src_buf,
const DstDesc& dst_desc,
DstBuffer& dst_buf,
const Dst0Desc& dst0_desc,
const Dst0Buffer& dst0_buf,
const Dst1Desc& dst1_desc,
const Dst1Buffer& dst1_buf)
{
auto f_step_hacks = [&](auto desc) {
constexpr index_t ntransform = decltype(desc)::GetNumOfTransform();
constexpr auto zeros = typename uniform_sequence_gen<ntransform, 0>::type{};
constexpr auto step_hacks =
make_tuple(generate_tuple([&](auto) { return zeros; }, Number<nDim>{}),
generate_tuple([&](auto) { return zeros; }, Number<nDim>{}));
return step_hacks;
};
Run(SrcDesc{},
SrcSliceOriginIdx{},
src_buf,
dst_desc,
dst_buf,
f_step_hacks(dst_desc),
dst0_desc,
dst0_buf,
f_step_hacks(dst0_desc),
dst1_desc,
dst1_buf,
f_step_hacks(dst1_desc));
}
__device__ static constexpr auto GetDstCoordinateResetStep()
{
constexpr auto I0 = Number<0>{};
// scalar per access on each dim
// TODO: don't use lambda_scalar_per_access
constexpr auto dst_scalar_per_access = generate_sequence(
detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
constexpr auto access_lengths = SliceLengths{} / dst_scalar_per_access;
constexpr auto dim_access_order = DimAccessOrder{};
constexpr auto ordered_access_lengths =
container_reorder_given_new2old(access_lengths, dim_access_order);
// judge move forward or move backward during the last iteration
constexpr auto forward_sweep = [&]() {
StaticallyIndexedArray<bool, nDim> forward_sweep_;
forward_sweep_(I0) = true;
static_for<1, nDim, 1>{}([&](auto i) {
index_t tmp = ordered_access_lengths[I0] - 1;
static_for<1, i, 1>{}([&](auto j) {
tmp = tmp * ordered_access_lengths[j] + ordered_access_lengths[j] - 1;
});
forward_sweep_(i) = tmp % 2 == 0;
});
return forward_sweep_;
}();
// calculate dst data index after last iteration in Run(), if it has not being reset by
// RunWrite()
constexpr auto dst_data_idx = [&]() {
Index ordered_idx;
static_for<0, nDim, 1>{}([&](auto i) {
ordered_idx(i) = forward_sweep[i] ? ordered_access_lengths[i] - 1 : 0;
});
return container_reorder_given_old2new(ordered_idx, dim_access_order) *
dst_scalar_per_access;
}();
//
constexpr auto reset_dst_data_step = [&]() {
Index reset_dst_data_step_;
static_for<0, nDim, 1>{}([&](auto i) { reset_dst_data_step_(i) = -dst_data_idx[i]; });
return reset_dst_data_step_;
}();
return reset_dst_data_step;
}
// dst_slice_origin_step_idx need to be known at compile-time, for performance reason
__device__ void MoveDstSliceWindow(const DstDesc& dst_desc,
const Index& dst_slice_origin_step_idx)
{
// if dst coord was not reset by Run(), then need to adjust the step here
const auto adjusted_step_idx =
DstResetCoordinateAfterRun ? dst_slice_origin_step_idx
: dst_slice_origin_step_idx + GetDstCoordinateResetStep();
// is it OK to construct a new step every time?
const auto adjusted_step = make_tensor_coordinate_step(dst_desc, adjusted_step_idx);
move_tensor_coordinate(dst_desc, dst_coord_, adjusted_step);
}
private:
DstCoord dst_coord_;
Dst0Coord dst0_coord_;
Dst1Coord dst1_coord_;
const DstElementwiseOperation dst_element_op_;
}; // namespace ck
} // namespace ck
#endif
include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v1r5.hpp deleted 100644 → 0
View file @ ecf337ba
#ifndef CK_THREADWISE_TENSOR_SLICE_TRANSFER_V1R5_HPP
#define CK_THREADWISE_TENSOR_SLICE_TRANSFER_V1R5_HPP
#include "common_header.hpp"
#include "tensor_descriptor.hpp"
#include "tensor_descriptor_helper.hpp"
namespace ck {
// Do following things to avoid "alloca" in LLVM-IR, which would cause scratch memory
// and sometimes useless instructions:
// 1. Don't save a reference to tensor descriptor in class, pass in tensor descriptor as argument
// instead
// 2. Don't construct a new tensor coordinate everytime when using it, update and reuse the same
// tensor coordinate instead
// 3. Don't use a pointer to VGPR buffer, use vector instead
// WARNING!!!!!!: this logic is only correct if DstScalarPerVector=1
// TODO: fix this
// Assume:
// 1. src:
// 1. SrcDesc is known at compile-time
// 2. SrcBuffer is StaticBuffer
// 3. SrcSliceOrginIdx is known at compile-time
// 2. dst:
// 1. DstDesc is not known at compile-time
// 2. DstBuffer is DynamicBuffer
// 3. DstSliceOrginIdx is not known at compile time
template <typename SrcData,
typename DstData,
typename SrcDesc,
typename DstDesc,
typename Dst0Desc, // this is really one of sources, but it has same shape as DstDesc
typename DstElementwiseOperation,
typename SliceLengths,
typename DimAccessOrder,
index_t DstVectorDim,
index_t DstScalarPerVector,
InMemoryDataOperationEnum_t DstInMemOp,
index_t DstScalarStrideInVector,
bool DstResetCoordinateAfterRun,
typename enable_if<SrcDesc::IsKnownAtCompileTime(), bool>::type = false>
struct ThreadwiseTensorSliceTransfer_v1r5
{
static constexpr index_t nDim = SliceLengths::Size();
using Index = MultiIndex<nDim>;
using DstCoord = decltype(make_tensor_coordinate(DstDesc{}, Index{}));
using Dst0Coord = decltype(make_tensor_coordinate(Dst0Desc{}, Index{}));
using DstCoordStep = decltype(make_tensor_coordinate_step(DstDesc{}, Index{}));
using Dst0CoordStep = decltype(make_tensor_coordinate_step(Dst0Desc{}, Index{}));
__device__ constexpr ThreadwiseTensorSliceTransfer_v1r5(
const DstDesc& dst_desc,
const Dst0Desc& dst0_desc,
const Index& dst_slice_origin_idx,
const DstElementwiseOperation& dst_element_op)
: dst_coord_(make_tensor_coordinate(dst_desc, dst_slice_origin_idx)),
dst0_coord_(make_tensor_coordinate(dst0_desc, dst_slice_origin_idx)),
dst_element_op_{dst_element_op}
{
static_assert(SrcDesc::IsKnownAtCompileTime(),
"wrong! SrcDesc need to known at compile-time");
}
__device__ void SetDstSliceOrigin(const DstDesc& dst_desc, const Index& dst_slice_origin_idx)
{
dst_coord_ = make_tensor_coordinate(dst_desc, dst_slice_origin_idx);
}
template <typename SrcSliceOriginIdx,
typename SrcBuffer,
typename DstBuffer,
typename Dst0Buffer,
typename DstStepHacks,
typename Dst0StepHacks>
__device__ void Run(const SrcDesc&,
const SrcSliceOriginIdx&,
const SrcBuffer& src_buf,
const DstDesc& dst_desc,
DstBuffer& dst_buf,
const DstStepHacks& dst_step_hacks,
const Dst0Desc& dst0_desc,
const Dst0Buffer& dst0_buf,
const Dst0StepHacks& dst0_step_hacks)
{
static_assert(SrcDesc::IsKnownAtCompileTime(),
"wrong! SrcDesc need to known at compile-time");
static_assert(is_known_at_compile_time<remove_cvref_t<SrcSliceOriginIdx>>::value,
"wrong! SrcSliceOrigin need to known at compile-time");
static_assert(SrcBuffer::IsStaticBuffer(), "wrong! SrcBuffer need to be StaticBuffer");
// SrcDesc and src_slice_origin_idx are known at compile-time
constexpr auto src_desc = remove_cvref_t<SrcDesc>{};
constexpr auto src_slice_origin_idx = to_multi_index(SrcSliceOriginIdx{});
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
// scalar per access on each dim
// TODO: don't use lambda_scalar_per_access
constexpr auto dst_scalar_per_access = generate_sequence(
detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
constexpr auto access_lengths = SliceLengths{} / dst_scalar_per_access;
constexpr auto dim_access_order = DimAccessOrder{};
constexpr auto ordered_access_lengths =
container_reorder_given_new2old(access_lengths, dim_access_order);
// make forward steps: dst
const auto dst_forward_steps = generate_tuple(
[&](auto i) {
Index forward_step_idx;
static_for<0, nDim, 1>{}([&](auto j) {
forward_step_idx(j) = (i.value == j.value) ? dst_scalar_per_access[i] : 0;
});
return make_tensor_coordinate_step(
dst_desc, forward_step_idx, dst_step_hacks[I0][i]);
},
Number<nDim>{});
// make forward steps: dst0
// WARNING!!!!!!: this logic is only correct if DstScalarPerVector=1
// TODO: fix this
const auto dst0_forward_steps = generate_tuple(
[&](auto i) {
Index forward_step_idx;
static_for<0, nDim, 1>{}([&](auto j) {
forward_step_idx(j) = (i.value == j.value) ? dst_scalar_per_access[i] : 0;
});
return make_tensor_coordinate_step(
dst0_desc, forward_step_idx, dst0_step_hacks[I0][i]);
},
Number<nDim>{});
// make backward steps: dst
const auto dst_backward_steps = generate_tuple(
[&](auto i) {
Index backward_step_idx;
static_for<0, nDim, 1>{}([&](auto j) {
backward_step_idx(j) = (i.value == j.value) ? -dst_scalar_per_access[i] : 0;
});
return make_tensor_coordinate_step(
dst_desc, backward_step_idx, dst_step_hacks[I1][i]);
},
Number<nDim>{});
// make backward steps: dst0
// WARNING!!!!!!: this logic is only correct if DstScalarPerVector=1
// TODO: fix this
const auto dst0_backward_steps = generate_tuple(
[&](auto i) {
Index backward_step_idx;
static_for<0, nDim, 1>{}([&](auto j) {
backward_step_idx(j) = (i.value == j.value) ? -dst_scalar_per_access[i] : 0;
});
return make_tensor_coordinate_step(
dst0_desc, backward_step_idx, dst0_step_hacks[I1][i]);
},
Number<nDim>{});
// loop over tensor and copy
static_ford<decltype(ordered_access_lengths)>{}([&](auto ordered_access_idx) {
// judge move forward or move backward
constexpr auto forward_sweep = [&]() {
StaticallyIndexedArray<bool, nDim> forward_sweep_;
forward_sweep_(I0) = true;
static_for<1, nDim, 1>{}([&](auto i) {
index_t tmp = ordered_access_idx[I0];
static_for<1, i, 1>{}([&](auto j) {
tmp = tmp * ordered_access_lengths[j] + ordered_access_idx[j];
});
forward_sweep_(i) = tmp % 2 == 0;
});
return forward_sweep_;
}();
// calculate dst data index
constexpr auto dst_data_idx = [&]() {
Index ordered_idx;
static_for<0, nDim, 1>{}([&](auto i) {
ordered_idx(i) = forward_sweep[i]
? ordered_access_idx[i]
: ordered_access_lengths[i] - 1 - ordered_access_idx[i];
});
return container_reorder_given_old2new(ordered_idx, dim_access_order) *
dst_scalar_per_access;
}();
typename vector_type_maker<DstData, DstScalarPerVector>::type dst_vector;
using dst_vector_t =
typename vector_type_maker<DstData, DstScalarPerVector>::type::type;
// load dst0 and apply elementwise operation
{
// WARNING!!!!!!: this logic is only correct if DstScalarPerVector=1
// TODO: fix this
static_assert(DstScalarPerVector == 1, "wrong!");
// copy data from src_buf into dst_vector_src_data
constexpr index_t src_offset =
src_desc.CalculateOffset(src_slice_origin_idx + dst_data_idx);
const SrcData src_v = src_buf[Number<src_offset>{}];
// load dst0
const bool is_dst0_valid =
coordinate_has_valid_offset_assuming_visible_index_is_valid(dst0_desc,
dst0_coord_);
const DstData dst0_v =
dst0_buf.template Get<DstData>(dst0_coord_.GetOffset(), is_dst0_valid);
#if !CK_WORKAROUND_SWDEV_XXXXXX_THREAD_WISE_COPY_V1R5_TYPE_CONVERT_ISSUE
// apply element-wise operation in SrcData type
const SrcData dst_v = dst_element_op_(src_v, type_convert<SrcData>(dst0_v));
// apply type convert
dst_vector.template AsType<DstData>()(Number<0>{}) = type_convert<DstData>(dst_v);
#else
// apply element-wise operation in DstData type
const DstData dst_v = dst_element_op_(src_v, dst0_v);
dst_vector.template AsType<DstData>()(Number<0>{}) = dst_v;
#endif
}
const bool is_dst_valid =
coordinate_has_valid_offset_assuming_visible_index_is_valid(dst_desc, dst_coord_);
// copy data from dst_vector into dst_buf
if constexpr(DstInMemOp == InMemoryDataOperationEnum_t::Set)
{
dst_buf.template Set<dst_vector_t>(
dst_coord_.GetOffset(),
is_dst_valid,
dst_vector.template AsType<dst_vector_t>()[Number<0>{}]);
}
else if constexpr(DstInMemOp == InMemoryDataOperationEnum_t::AtomicAdd)
{
dst_buf.template AtomicAdd<dst_vector_t>(
dst_coord_.GetOffset(),
is_dst_valid,
dst_vector.template AsType<dst_vector_t>()[Number<0>{}]);
}
else if constexpr(DstInMemOp == InMemoryDataOperationEnum_t::Add)
{
typename vector_type_maker<DstData, DstScalarPerVector>::type tmp;
tmp.template AsType<dst_vector_t>()(Number<0>{}) =
dst_buf.template Get<dst_vector_t>(dst_coord_.GetOffset(), is_dst_valid);
static_for<0, DstScalarPerVector, 1>{}([&](auto t) {
dst_vector.template AsType<DstData>()(t) += tmp.template AsType<DstData>()[t];
});
dst_buf.template Set<dst_vector_t>(
dst_coord_.GetOffset(),
is_dst_valid,
dst_vector.template AsType<dst_vector_t>()[Number<0>{}]);
}
constexpr auto move_on_dim = [&]() constexpr
{
StaticallyIndexedArray<bool, nDim> move_on_dim_;
static_for<0, nDim, 1>{}([&](auto i) {
move_on_dim_(i) = ordered_access_idx[i] < ordered_access_lengths[i] - 1;
static_for<i + 1, nDim, 1>{}([&](auto j) {
move_on_dim_(i) &= ordered_access_idx[j] == ordered_access_lengths[j] - 1;
});
});
return move_on_dim_;
}
();
// move
static_for<0, nDim, 1>{}([&](auto i) {
if constexpr(move_on_dim[i])
{
if constexpr(forward_sweep[i])
{
move_tensor_coordinate(
dst_desc, dst_coord_, dst_forward_steps[dim_access_order[i]]);
// dst0
move_tensor_coordinate(
dst0_desc, dst0_coord_, dst0_forward_steps[dim_access_order[i]]);
}
else
{
move_tensor_coordinate(
dst_desc, dst_coord_, dst_backward_steps[dim_access_order[i]]);
// dst0
move_tensor_coordinate(
dst0_desc, dst0_coord_, dst0_backward_steps[dim_access_order[i]]);
}
}
});
});
// move dst coordinate back to slice origin (or not)
if constexpr(DstResetCoordinateAfterRun)
{
const auto dst_reset_step =
make_tensor_coordinate_step(dst_desc, GetDstCoordinateResetStep());
move_tensor_coordinate(dst_desc, dst_coord_, dst_reset_step);
}
}
template <typename SrcSliceOriginIdx,
typename SrcBuffer,
typename DstBuffer,
typename Dst0Buffer>
__device__ void Run(const SrcDesc&,
const SrcSliceOriginIdx&,
const SrcBuffer& src_buf,
const DstDesc& dst_desc,
DstBuffer& dst_buf,
const Dst0Desc& dst0_desc,
const Dst0Buffer& dst0_buf)
{
auto f_step_hacks = [&](auto desc) {
constexpr index_t ntransform = decltype(desc)::GetNumOfTransform();
constexpr auto zeros = typename uniform_sequence_gen<ntransform, 0>::type{};
constexpr auto step_hacks =
make_tuple(generate_tuple([&](auto) { return zeros; }, Number<nDim>{}),
generate_tuple([&](auto) { return zeros; }, Number<nDim>{}));
return step_hacks;
};
Run(SrcDesc{},
SrcSliceOriginIdx{},
src_buf,
dst_desc,
dst_buf,
f_step_hacks(dst_desc),
dst0_desc,
dst0_buf,
f_step_hacks(dst0_desc));
}
__device__ static constexpr auto GetDstCoordinateResetStep()
{
constexpr auto I0 = Number<0>{};
// scalar per access on each dim
// TODO: don't use lambda_scalar_per_access
constexpr auto dst_scalar_per_access = generate_sequence(
detail::lambda_scalar_per_access<DstVectorDim, DstScalarPerVector>{}, Number<nDim>{});
constexpr auto access_lengths = SliceLengths{} / dst_scalar_per_access;
constexpr auto dim_access_order = DimAccessOrder{};
constexpr auto ordered_access_lengths =
container_reorder_given_new2old(access_lengths, dim_access_order);
// judge move forward or move backward during the last iteration
constexpr auto forward_sweep = [&]() {
StaticallyIndexedArray<bool, nDim> forward_sweep_;
forward_sweep_(I0) = true;
static_for<1, nDim, 1>{}([&](auto i) {
index_t tmp = ordered_access_lengths[I0] - 1;
static_for<1, i, 1>{}([&](auto j) {
tmp = tmp * ordered_access_lengths[j] + ordered_access_lengths[j] - 1;
});
forward_sweep_(i) = tmp % 2 == 0;
});
return forward_sweep_;
}();
// calculate dst data index after last iteration in Run(), if it has not being reset by
// RunWrite()
constexpr auto dst_data_idx = [&]() {
Index ordered_idx;
static_for<0, nDim, 1>{}([&](auto i) {
ordered_idx(i) = forward_sweep[i] ? ordered_access_lengths[i] - 1 : 0;
});
return container_reorder_given_old2new(ordered_idx, dim_access_order) *
dst_scalar_per_access;
}();
//
constexpr auto reset_dst_data_step = [&]() {
Index reset_dst_data_step_;
static_for<0, nDim, 1>{}([&](auto i) { reset_dst_data_step_(i) = -dst_data_idx[i]; });
return reset_dst_data_step_;
}();
return reset_dst_data_step;
}
// dst_slice_origin_step_idx need to be known at compile-time, for performance reason
__device__ void MoveDstSliceWindow(const DstDesc& dst_desc,
const Index& dst_slice_origin_step_idx)
{
// if dst coord was not reset by Run(), then need to adjust the step here
const auto adjusted_step_idx =
DstResetCoordinateAfterRun ? dst_slice_origin_step_idx
: dst_slice_origin_step_idx + GetDstCoordinateResetStep();
// is it OK to construct a new step every time?
const auto adjusted_step = make_tensor_coordinate_step(dst_desc, adjusted_step_idx);
move_tensor_coordinate(dst_desc, dst_coord_, adjusted_step);
}
private:
DstCoord dst_coord_;
Dst0Coord dst0_coord_;
const DstElementwiseOperation dst_element_op_;
}; // namespace ck
} // namespace ck
#endif
include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r1.hpp
View file @ cd167e49
......@@ -48,7 +48,7 @@ struct lambda_scalar_per_access_for_src_and_dst
template <typename SliceLengths,
typename SrcElementwiseOperation,
typename DstElementwiseOperation,
InMemoryDataOperationEnum_t DstInMemOp,
InMemoryDataOperationEnum DstInMemOp,
typename SrcData,
typename DstData,
typename SrcDesc,
......@@ -110,8 +110,8 @@ struct ThreadwiseTensorSliceTransfer_v3r1
const SrcBuffer& src_buf,
Number<ThreadScratchId> thread_scratch_id = Number<ThreadScratchId>{})
{
static_assert(SrcBuffer::GetAddressSpace() == AddressSpaceEnum_t::Global or
SrcBuffer::GetAddressSpace() == AddressSpaceEnum_t::Lds,
static_assert(SrcBuffer::GetAddressSpace() == AddressSpaceEnum::Global or
SrcBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
"wrong!");
static_assert(
......@@ -271,7 +271,7 @@ struct ThreadwiseTensorSliceTransfer_v3r1
static_ford<SliceLengths>{}([&](auto idx) {
// convert from SrcData to DstData here
dst_thread_scratch_(idx) =
type_convert<DstData>(src_thread_scratch_tuple[thread_scratch_id][idx]);
type_convert<DstData>(src_thread_scratch_tuple_[thread_scratch_id][idx]);
});
#else
// sub-dword transpose between src_thread_scratch_ and dst_thread_scratch_
......@@ -361,8 +361,8 @@ struct ThreadwiseTensorSliceTransfer_v3r1
// TODO move this elsewhere
TransferDataFromSrcThreadScratchToDstThreadScratch(thread_scratch_id);
static_assert(DstBuffer::GetAddressSpace() == AddressSpaceEnum_t::Global or
DstBuffer::GetAddressSpace() == AddressSpaceEnum_t::Lds,
static_assert(DstBuffer::GetAddressSpace() == AddressSpaceEnum::Global or
DstBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
"wrong!");
static_assert(
......@@ -763,13 +763,13 @@ struct ThreadwiseTensorSliceTransfer_v3r1
static constexpr auto src_thread_scratch_desc_ = decltype(GetSrcThreadScratchDescriptor()){};
static constexpr auto dst_thread_scratch_desc_ = decltype(GetDstThreadScratchDescriptor()){};
using SrcThreadScratch = StaticTensorTupleOfVectorBuffer<AddressSpaceEnum_t::Vgpr,
using SrcThreadScratch = StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
SrcData,
SrcScalarPerVector,
decltype(src_thread_scratch_desc_),
true>;
using DstThreadScratch = StaticTensorTupleOfVectorBuffer<AddressSpaceEnum_t::Vgpr,
using DstThreadScratch = StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
DstData,
DstScalarPerVector,
decltype(dst_thread_scratch_desc_),
......
include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v3r3.hpp
View file @ cd167e49
......@@ -48,7 +48,7 @@ struct lambda_scalar_per_access_for_src_and_dst
template <typename SliceLengths,
typename SrcElementwiseOperation,
typename DstElementwiseOperation,
InMemoryDataOperationEnum_t DstInMemOp,
InMemoryDataOperationEnum DstInMemOp,
typename SrcData,
typename DstData,
typename SrcDesc,
......@@ -120,8 +120,8 @@ struct ThreadwiseTensorSliceTransfer_v3r3
template <typename SrcBuffer>
__device__ void RunRead(const SrcDesc& src_desc, const SrcBuffer& src_buf)
{
static_assert(SrcBuffer::GetAddressSpace() == AddressSpaceEnum_t::Global or
SrcBuffer::GetAddressSpace() == AddressSpaceEnum_t::Lds,
static_assert(SrcBuffer::GetAddressSpace() == AddressSpaceEnum::Global or
SrcBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
"wrong!");
static_assert(
......@@ -369,8 +369,8 @@ struct ThreadwiseTensorSliceTransfer_v3r3
// TODO move this elsewhere
TransferDataFromSrcThreadScratchToDstThreadScratch();
static_assert(DstBuffer::GetAddressSpace() == AddressSpaceEnum_t::Global or
DstBuffer::GetAddressSpace() == AddressSpaceEnum_t::Lds,
static_assert(DstBuffer::GetAddressSpace() == AddressSpaceEnum::Global or
DstBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
"wrong!");
static_assert(
......@@ -859,14 +859,14 @@ struct ThreadwiseTensorSliceTransfer_v3r3
static constexpr auto src_thread_scratch_desc_ = decltype(GetSrcThreadScratchDescriptor()){};
static constexpr auto dst_thread_scratch_desc_ = decltype(GetDstThreadScratchDescriptor()){};
StaticTensorTupleOfVectorBuffer<AddressSpaceEnum_t::Vgpr,
StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
SrcData,
SrcScalarPerVector,
decltype(src_thread_scratch_desc_),
true>
src_thread_scratch_;
StaticTensorTupleOfVectorBuffer<AddressSpaceEnum_t::Vgpr,
StaticTensorTupleOfVectorBuffer<AddressSpaceEnum::Vgpr,
DstData,
DstScalarPerVector,
decltype(dst_thread_scratch_desc_),
......
include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v5r1.hpp
View file @ cd167e49
......@@ -13,7 +13,7 @@ namespace ck {
// 3. src_slice_origin and dst_slice_origin are not known at compile-time,
// 4. Use thread buffer
template <typename SliceLengths,
InMemoryDataOperationEnum_t DstInMemOp,
InMemoryDataOperationEnum DstInMemOp,
typename SrcData,
typename DstData,
typename SrcDesc,
......@@ -76,8 +76,8 @@ struct ThreadwiseTensorSliceTransfer_v5r1
__device__ void
RunRead(const SrcDesc& src_desc, const SrcBuffer& src_buf, const SrcStepHacks& src_step_hacks)
{
static_assert(SrcBuffer::GetAddressSpace() == AddressSpaceEnum_t::Global or
SrcBuffer::GetAddressSpace() == AddressSpaceEnum_t::Lds,
static_assert(SrcBuffer::GetAddressSpace() == AddressSpaceEnum::Global or
SrcBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
"wrong!");
static_assert(
......@@ -244,8 +244,8 @@ struct ThreadwiseTensorSliceTransfer_v5r1
__device__ void
RunWrite(const DstDesc& dst_desc, DstBuffer& dst_buf, const DstStepHacks& dst_step_hacks)
{
static_assert(DstBuffer::GetAddressSpace() == AddressSpaceEnum_t::Global or
DstBuffer::GetAddressSpace() == AddressSpaceEnum_t::Lds,
static_assert(DstBuffer::GetAddressSpace() == AddressSpaceEnum::Global or
DstBuffer::GetAddressSpace() == AddressSpaceEnum::Lds,
"wrong!");
static_assert(
......@@ -602,7 +602,7 @@ struct ThreadwiseTensorSliceTransfer_v5r1
static constexpr auto buffer_size_ = buffer_desc_.GetElementSpaceSize();
StaticBuffer<AddressSpaceEnum_t::Vgpr, SrcData, buffer_size_, true> buffer_;
StaticBuffer<AddressSpaceEnum::Vgpr, SrcData, buffer_size_, true> buffer_;
SrcCoord src_coord_;
DstCoord dst_coord_;
......
include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r1.hpp
View file @ cd167e49
......@@ -29,7 +29,7 @@ template <typename SrcData,
typename DimAccessOrder,
index_t VectorDim,
index_t ScalarPerVector,
InMemoryDataOperationEnum_t DstInMemOp,
InMemoryDataOperationEnum DstInMemOp,
bool SrcResetCoordinateAfterRun,
bool DstResetCoordinateAfterRun>
struct ThreadwiseTensorSliceTransfer_v6r1
......
include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r2.hpp
View file @ cd167e49
......@@ -31,7 +31,7 @@ template <typename Src0Data,
typename DimAccessOrder,
index_t VectorDim,
index_t ScalarPerVector,
InMemoryDataOperationEnum_t DstInMemOp,
InMemoryDataOperationEnum DstInMemOp,
bool Src0ResetCoordinateAfterRun,
bool Src1ResetCoordinateAfterRun,
bool DstResetCoordinateAfterRun>
......
include/ck/tensor_operation/gpu/thread/threadwise_tensor_slice_transfer_v6r3.hpp
View file @ cd167e49
......@@ -33,7 +33,7 @@ template <typename Src0Data,
typename DimAccessOrder,
index_t VectorDim,
index_t ScalarPerVector,
InMemoryDataOperationEnum_t DstInMemOp,
InMemoryDataOperationEnum DstInMemOp,
bool Src0ResetCoordinateAfterRun,
bool Src1ResetCoordinateAfterRun,
bool Src2ResetCoordinateAfterRun,
......
include/ck/tensor_operation/gpu/warp/xdlops_gemm.hpp
View file @ cd167e49
......@@ -476,7 +476,7 @@ struct MfmaSelector
template <>
static constexpr auto GetMfma<bhalf_t, 32, 32>()
{
#if defined(CK_AMD_GPU_GFX90A)
#if defined(CK_USE_AMD_MFMA_BF16_1K_OP)
return MfmaInstr::mfma_f32_32x32x8bf16_1k;
#else
return MfmaInstr::mfma_f32_32x32x4bf16;
......@@ -486,7 +486,7 @@ struct MfmaSelector
template <>
static constexpr auto GetMfma<bhalf_t, 16, 16>()
{
#if defined(CK_AMD_GPU_GFX90A)
#if defined(CK_USE_AMD_MFMA_BF16_1K_OP)
return MfmaInstr::mfma_f32_16x16x16bf16_1k;
#else
return MfmaInstr::mfma_f32_16x16x8bf16;
......
include/ck/utility/amd_address_space.hpp
View file @ cd167e49
......@@ -9,7 +9,7 @@
namespace ck {
enum struct AddressSpaceEnum_t
enum struct AddressSpaceEnum
{
Generic,
Global,
......@@ -19,7 +19,7 @@ enum struct AddressSpaceEnum_t
};
template <typename T>
__device__ T* cast_pointer_to_generic_address_space(T CONSTANT* p)
__device__ T* cast_pointer_to_generic_address_space(T CK_CONSTANT_ADDRESS_SPACE* p)
{
// cast a pointer in "Constant" address space (4) to "Generic" address space (0)
// only c-style pointer cast seems be able to be compiled
......@@ -30,13 +30,13 @@ __device__ T* cast_pointer_to_generic_address_space(T CONSTANT* p)
}
template <typename T>
__host__ __device__ T CONSTANT* cast_pointer_to_constant_address_space(T* p)
__host__ __device__ T CK_CONSTANT_ADDRESS_SPACE* cast_pointer_to_constant_address_space(T* p)
{
// cast a pointer in "Generic" address space (0) to "Constant" address space (4)
// only c-style pointer cast seems be able to be compiled
#pragma clang diagnostic push
#pragma clang diagnostic ignored "-Wold-style-cast"
return (T CONSTANT*)p; // NOLINT(old-style-cast)
return (T CK_CONSTANT_ADDRESS_SPACE*)p; // NOLINT(old-style-cast)
#pragma clang diagnostic pop
}
......
include/ck/utility/amd_buffer_addressing.hpp
View file @ cd167e49
#ifndef CK_AMD_BUFFER_ADDRESSING_HPP
#define CK_AMD_BUFFER_ADDRESSING_HPP
#pragma once
#include "data_type.hpp"
namespace ck {
......@@ -87,6 +85,7 @@ llvm_amdgcn_raw_buffer_load_i32x4(int32x4_t srsrc,
index_t voffset,
index_t soffset,
index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.load.v4i32");
// buffer load fp16
__device__ half_t
llvm_amdgcn_raw_buffer_load_fp16(int32x4_t srsrc,
......@@ -212,6 +211,7 @@ llvm_amdgcn_raw_buffer_store_fp16x4(half4_t vdata,
index_t voffset,
index_t soffset,
index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v4f16");
// buffer store fp32
__device__ void
llvm_amdgcn_raw_buffer_store_fp32(float vdata,
......@@ -233,6 +233,7 @@ llvm_amdgcn_raw_buffer_store_fp32x4(float4_t vdata,
index_t voffset,
index_t soffset,
index_t glc_slc) __asm("llvm.amdgcn.raw.buffer.store.v4f32");
// buffer atomic-add fp16
__device__ half2_t llvm_amdgcn_raw_buffer_atomic_add_fp16x2(
half2_t vdata,
......@@ -1046,4 +1047,3 @@ amd_buffer_atomic_add(const typename vector_type_maker<T, N>::type::type src_thr
}
} // namespace ck
#endif
include/ck/utility/common_header.hpp
View file @ cd167e49
#ifndef CK_COMMON_HEADER_HPP
#define CK_COMMON_HEADER_HPP
#pragma once
#include "config.hpp"
#include "array.hpp"
#include "container_helper.hpp"
......@@ -20,30 +18,29 @@
#include "number.hpp"
#include "sequence.hpp"
#include "sequence_helper.hpp"
#include "synchronization.hpp"
#include "tuple.hpp"
#include "tuple_helper.hpp"
#include "type.hpp"
#include "magic_division.hpp"
#include "utility.hpp"
#include "c_style_pointer_cast.hpp"
#include "amd_address_space.hpp"
#include "amd_buffer_addressing.hpp"
#include "static_buffer.hpp"
#include "dynamic_buffer.hpp"
#include "is_known_at_compile_time.hpp"
#include "transpose_vectors.hpp"
#include "inner_product.hpp"
#include "element_wise_operation.hpp"
#include "debug.hpp"
#include "amd_buffer_addressing.hpp"
#include "get_id.hpp"
#include "synchronization.hpp"
#include "amd_address_space.hpp"
#include "static_buffer.hpp"
#include "dynamic_buffer.hpp"
// TODO: remove this
#if CK_USE_AMD_INLINE_ASM
#include "amd_inline_asm.hpp"
#endif
#if CK_USE_AMD_XDLOPS
#ifdef CK_USE_AMD_MFMA
#include "amd_xdlops.hpp"
#endif
#endif
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