Skip to content

GitLab

Commit 491437a7 authored by Chao Liu's avatar Chao Liu
Browse files

remove coordinate step hack from all gridwise gemm xdl

parent 619661f8
Hide whitespace changes
Inline Side-by-side
Showing with 140 additions and 322 deletions
composable_kernel/include/tensor_operation/gridwise_gemm_xdlops_v2r3.hpp
View file @ 491437a7
......@@ -156,11 +156,6 @@ template <index_t BlockSize,
typename CThreadTransferSrcDstAccessOrder,
index_t CThreadTransferSrcDstVectorDim,
index_t CThreadTransferDstScalarPerVector,
typename AGridStepHacks,
typename BGridStepHacks,
typename CGridStepHacks,
typename AGridMoveSliceWindowStepHacks,
typename BGridMoveSliceWindowStepHacks,
bool CAccessOrderMRepeatNRepeat,
bool ABlockLdsExtraM,
bool BBlockLdsExtraN>
......@@ -442,6 +437,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
auto a_blockwise_copy =
BlockwiseTensorSliceTransfer_v4<BlockSize,
AElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
Sequence<K0PerBlock, MPerBlock, K1>,
ABlockTransferThreadSliceLengths_K0_M_K1,
......@@ -460,16 +456,19 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
1,
1,
AThreadTransferSrcResetCoordinateAfterRun,
true>(a_grid_desc_k0_m_k1,
make_multi_index(0, m_block_data_idx_on_grid, 0),
a_block_desc_k0_m_k1,
make_multi_index(0, 0, 0),
a_element_op);
true>(
a_grid_desc_k0_m_k1,
make_multi_index(0, m_block_data_idx_on_grid, 0),
a_element_op,
a_block_desc_k0_m_k1,
make_multi_index(0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
// B matrix blockwise copy
auto b_blockwise_copy =
BlockwiseTensorSliceTransfer_v4<BlockSize,
BElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
Sequence<K0PerBlock, NPerBlock, K1>,
BBlockTransferThreadSliceLengths_K0_N_K1,
......@@ -488,11 +487,13 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
1,
1,
BThreadTransferSrcResetCoordinateAfterRun,
true>(b_grid_desc_k0_n_k1,
make_multi_index(0, n_block_data_idx_on_grid, 0),
b_block_desc_k0_n_k1,
make_multi_index(0, 0, 0),
b_element_op);
true>(
b_grid_desc_k0_n_k1,
make_multi_index(0, n_block_data_idx_on_grid, 0),
b_element_op,
b_block_desc_k0_n_k1,
make_multi_index(0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
// GEMM definition
// c_mtx += transpose(a_mtx) * b_mtx
......@@ -526,15 +527,6 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
constexpr auto a_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
constexpr auto b_block_slice_copy_step = make_multi_index(K0PerBlock, 0, 0);
// hack to control index calculation when iterating over A and B matrix for threadwise copy
constexpr auto a_k0_m_k1_grid_step_hacks = AGridStepHacks{};
constexpr auto b_k0_n_k1_grid_step_hacks = BGridStepHacks{};
// hack to control index calculation when move slice window for A and B matrix for
// threadwise copy
constexpr auto a_k0_m_k1_grid_move_slice_window_step_hack = AGridMoveSliceWindowStepHacks{};
constexpr auto b_k0_n_k1_grid_move_slice_window_step_hack = BGridMoveSliceWindowStepHacks{};
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
p_a_block, a_block_desc_k0_m_k1.GetElementSpaceSize());
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
......@@ -542,36 +534,31 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
// preload data into LDS
{
a_blockwise_copy.RunRead(a_grid_desc_k0_m_k1, a_grid_buf, a_k0_m_k1_grid_step_hacks);
b_blockwise_copy.RunRead(b_grid_desc_k0_n_k1, b_grid_buf, b_k0_n_k1_grid_step_hacks);
a_blockwise_copy.RunRead(a_grid_desc_k0_m_k1, a_grid_buf);
b_blockwise_copy.RunRead(b_grid_desc_k0_n_k1, b_grid_buf);
a_blockwise_copy.RunWrite(a_block_desc_k0_m_k1, a_block_buf);
b_blockwise_copy.RunWrite(b_block_desc_k0_n_k1, b_block_buf);
}
// main body
index_t k0_block_data_begin = 0;
// Initialize C
c_thread_buf.Clear();
// main body
if constexpr(HasMainKBlockLoop)
{
index_t k0_block_data_begin = 0;
do
{
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_k0_m_k1,
a_block_slice_copy_step,
a_k0_m_k1_grid_move_slice_window_step_hack);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_k0_n_k1,
b_block_slice_copy_step,
b_k0_n_k1_grid_move_slice_window_step_hack);
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_k0_m_k1, a_block_slice_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_grid_desc_k0_n_k1, b_block_slice_copy_step);
a_blockwise_copy.RunRead(
a_grid_desc_k0_m_k1, a_grid_buf, a_k0_m_k1_grid_step_hacks);
a_blockwise_copy.RunRead(a_grid_desc_k0_m_k1, a_grid_buf);
block_sync_lds();
b_blockwise_copy.RunRead(
b_grid_desc_k0_n_k1, b_grid_buf, b_k0_n_k1_grid_step_hacks);
b_blockwise_copy.RunRead(b_grid_desc_k0_n_k1, b_grid_buf);
blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
......@@ -619,8 +606,6 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
const index_t n_thread_data_on_grid =
n_block_data_idx_on_grid + c_thread_mtx_on_block[I1];
constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_grid_tensor_step_hacks = CGridStepHacks{};
const auto m_thread_data_on_grid_to_m0_m1_m2_m3_m4_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
......@@ -668,8 +653,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3
make_tuple(I0, I0, I0, I0, I0, I0, I0, I0),
c_thread_buf,
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2,
c_grid_buf,
c_m0_n0_m1_n1_m2_m3_m4_n2_grid_tensor_step_hacks);
c_grid_buf);
}
}
};
......
composable_kernel/include/tensor_operation/gridwise_gemm_xdlops_v2r4.hpp
View file @ 491437a7
......@@ -19,6 +19,9 @@ template <typename GridwiseGemm,
typename ABK0MK1GridDesc,
typename BBK0NK1GridDesc,
typename CM0N0M1N1M2M3M4N2GridDesc,
typename AElementwiseOperation,
typename BElementwiseOperation,
typename CElementwiseOperation,
typename CBlockClusterAdaptor,
bool HasMainKBlockLoop>
__global__ void
......@@ -31,6 +34,9 @@ __global__ void
const ABK0MK1GridDesc a_b_k0_m_k1_grid_desc,
const BBK0NK1GridDesc b_b_k0_n_k1_grid_desc,
const CM0N0M1N1M2M3M4N2GridDesc c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
const AElementwiseOperation a_element_op,
const BElementwiseOperation b_element_op,
const CElementwiseOperation c_element_op,
const CBlockClusterAdaptor c_block_cluster_adaptor)
{
constexpr index_t shared_block_size =
......@@ -45,6 +51,9 @@ __global__ void
a_b_k0_m_k1_grid_desc,
b_b_k0_n_k1_grid_desc,
c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
a_element_op,
b_element_op,
c_element_op,
c_block_cluster_adaptor);
}
#elif CK_EXPERIMENTAL_PASS_TENSOR_DESCRIPTOR_BY_VOID_POINTER
......@@ -129,11 +138,6 @@ template <index_t BlockSize,
typename CThreadTransferSrcDstAccessOrder,
index_t CThreadTransferSrcDstVectorDim,
index_t CThreadTransferDstScalarPerVector,
typename AGridStepHacks,
typename BGridStepHacks,
typename CGridStepHacks,
typename AGridMoveSliceWindowStepHacks,
typename BGridMoveSliceWindowStepHacks,
bool CAccessOrderMRepeatNRepeat,
bool ABlockLdsExtraM,
bool BBlockLdsExtraN>
......@@ -371,6 +375,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4
c_block_cluster_adaptor.CalculateBottomIndex(make_multi_index(get_block_1d_id()));
const index_t k_batch_id = block_work_idx[I0];
// HACK: this force m/n_block_data_idx_on_grid into SGPR
const index_t m_block_data_idx_on_grid =
__builtin_amdgcn_readfirstlane(block_work_idx[I1] * MPerBlock);
......@@ -448,6 +453,8 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4
// A matrix blockwise copy
auto a_blockwise_copy =
BlockwiseTensorSliceTransfer_v4<BlockSize,
AElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
Sequence<1, K0PerBlock, MPerBlock, K1>,
ABlockTransferThreadSliceLengths_K0_M_K1,
......@@ -469,12 +476,16 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4
true>(
a_b_k0_m_k1_grid_desc,
make_multi_index(k_batch_id, 0, m_block_data_idx_on_grid, 0),
a_element_op,
a_b_k0_m_k1_block_desc,
make_multi_index(0, 0, 0, 0));
make_multi_index(0, 0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
// B matrix blockwise copy
auto b_blockwise_copy =
BlockwiseTensorSliceTransfer_v4<BlockSize,
BElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
Sequence<1, K0PerBlock, NPerBlock, K1>,
BBlockTransferThreadSliceLengths_K0_N_K1,
......@@ -496,8 +507,10 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4
true>(
b_b_k0_n_k1_grid_desc,
make_multi_index(k_batch_id, 0, n_block_data_idx_on_grid, 0),
b_element_op,
b_b_k0_n_k1_block_desc,
make_multi_index(0, 0, 0, 0));
make_multi_index(0, 0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
// GEMM definition
// c_mtx += transpose(a_mtx) * b_mtx
......@@ -531,15 +544,6 @@ 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);
// hack to control index calculation when iterating over A and B matrix for threadwise copy
constexpr auto a_k0_m_k1_grid_step_hacks = AGridStepHacks{};
constexpr auto b_k0_n_k1_grid_step_hacks = BGridStepHacks{};
// hack to control index calculation when move slice window for A and B matrix for
// threadwise copy
constexpr auto a_k0_m_k1_grid_move_slice_window_step_hack = AGridMoveSliceWindowStepHacks{};
constexpr auto b_k0_n_k1_grid_move_slice_window_step_hack = BGridMoveSliceWindowStepHacks{};
auto a_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
p_a_block, a_k0_m_k1_block_desc.GetElementSpaceSize());
auto b_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
......@@ -547,33 +551,31 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4
// preload data into LDS
{
a_blockwise_copy.RunRead(a_b_k0_m_k1_grid_desc, a_grid_buf, a_k0_m_k1_grid_step_hacks);
b_blockwise_copy.RunRead(b_b_k0_n_k1_grid_desc, b_grid_buf, b_k0_n_k1_grid_step_hacks);
a_blockwise_copy.RunRead(a_b_k0_m_k1_grid_desc, a_grid_buf);
b_blockwise_copy.RunRead(b_b_k0_n_k1_grid_desc, b_grid_buf);
a_blockwise_copy.RunWrite(a_b_k0_m_k1_block_desc, a_block_buf);
b_blockwise_copy.RunWrite(b_b_k0_n_k1_block_desc, b_block_buf);
}
// Initialize C
c_thread_buf.Clear();
// main body
index_t k_block_data_begin = 0;
if constexpr(HasMainKBlockLoop)
{
index_t k0_block_data_begin = 0;
do
{
a_blockwise_copy.MoveSrcSliceWindow(a_b_k0_m_k1_grid_desc,
a_block_slice_copy_step,
a_k0_m_k1_grid_move_slice_window_step_hack);
b_blockwise_copy.MoveSrcSliceWindow(b_b_k0_n_k1_grid_desc,
b_block_slice_copy_step,
b_k0_n_k1_grid_move_slice_window_step_hack);
a_blockwise_copy.MoveSrcSliceWindow(a_b_k0_m_k1_grid_desc, a_block_slice_copy_step);
b_blockwise_copy.MoveSrcSliceWindow(b_b_k0_n_k1_grid_desc, b_block_slice_copy_step);
a_blockwise_copy.RunRead(
a_b_k0_m_k1_grid_desc, a_grid_buf, a_k0_m_k1_grid_step_hacks);
a_blockwise_copy.RunRead(a_b_k0_m_k1_grid_desc, a_grid_buf);
block_sync_lds();
b_blockwise_copy.RunRead(
b_b_k0_n_k1_grid_desc, b_grid_buf, b_k0_n_k1_grid_step_hacks);
b_blockwise_copy.RunRead(b_b_k0_n_k1_grid_desc, b_grid_buf);
blockwise_gemm.Run(a_block_buf, b_block_buf, c_thread_buf);
......@@ -622,8 +624,6 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4
const index_t n_thread_data_on_grid =
n_block_data_idx_on_grid + c_thread_mtx_on_block[I1];
constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_grid_tensor_step_hacks = CGridStepHacks{};
const auto m_thread_data_on_grid_to_m0_m1_m2_m3_m4_adaptor =
make_single_stage_tensor_adaptor(
make_tuple(make_merge_transform(make_tuple(M0, M1, M2, M3, M4))),
......@@ -648,6 +648,7 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4
FloatC,
decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc),
decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc),
CElementwiseOperation,
Sequence<M0, N0, I1, I1, M2, I1, M4, I1>,
CThreadTransferSrcDstAccessOrder,
CThreadTransferSrcDstVectorDim,
......@@ -664,14 +665,14 @@ struct GridwiseGemm_bk0mk1_bk0nk1_mn_xdlops_v2r4
m_thread_data_on_grid_idx[I2],
m_thread_data_on_grid_idx[I3],
m_thread_data_on_grid_idx[I4],
n_thread_data_on_grid_idx[I2])};
n_thread_data_on_grid_idx[I2]),
c_element_op};
c_thread_copy.Run(c_m0_n0_m1_n1_m2_m3_m4_n2_thread_desc,
make_tuple(I0, I0, I0, I0, I0, I0, I0, I0),
c_thread_buf,
c_m0_n0_m1_n1_m2_m3_m4_n2_grid_desc,
c_grid_buf,
c_m0_n0_m1_n1_m2_m3_m4_n2_grid_tensor_step_hacks);
c_grid_buf);
}
}
}; // namespace ck
......
composable_kernel/include/tensor_operation/gridwise_gemm_xdlops_v2r5.hpp
View file @ 491437a7
......@@ -412,6 +412,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r5
auto a_blockwise_copy =
BlockwiseTensorSliceTransfer_v4<BlockSize,
AElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
Sequence<K0PerBlock, MPerBlock, K1>,
ABlockTransferThreadSliceLengths_K0_M_K1,
......@@ -430,16 +431,19 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r5
1,
1,
AThreadTransferSrcResetCoordinateAfterRun,
true>(a_grid_desc_k0_m_k1,
make_multi_index(0, m_block_data_idx_on_grid, 0),
a_block_desc_k0_m_k1,
make_multi_index(0, 0, 0),
a_element_op);
true>(
a_grid_desc_k0_m_k1,
make_multi_index(0, m_block_data_idx_on_grid, 0),
a_element_op,
a_block_desc_k0_m_k1,
make_multi_index(0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
// B matrix blockwise copy
auto b_blockwise_copy =
BlockwiseTensorSliceTransfer_v4<BlockSize,
BElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
Sequence<K0PerBlock, NPerBlock, K1>,
BBlockTransferThreadSliceLengths_K0_N_K1,
......@@ -458,11 +462,13 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r5
1,
1,
BThreadTransferSrcResetCoordinateAfterRun,
true>(b_grid_desc_k0_n_k1,
make_multi_index(0, n_block_data_idx_on_grid, 0),
b_block_desc_k0_n_k1,
make_multi_index(0, 0, 0),
b_element_op);
true>(
b_grid_desc_k0_n_k1,
make_multi_index(0, n_block_data_idx_on_grid, 0),
b_element_op,
b_block_desc_k0_n_k1,
make_multi_index(0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
// GEMM definition
// c_mtx += transpose(a_mtx) * b_mtx
......@@ -519,11 +525,14 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r5
b_blockwise_copy.RunWrite(b_block_desc_k0_n_k1, b_block_buf);
}
// main body
index_t k0_block_data_begin = 0;
// Initialize C
c_thread_buf.Clear();
// main body
if constexpr(HasMainKBlockLoop)
{
index_t k0_block_data_begin = 0;
do
{
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_k0_m_k1,
......
composable_kernel/include/tensor_operation/gridwise_gemm_xdlops_v2r6.hpp
View file @ 491437a7
......@@ -398,6 +398,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r6
auto a_blockwise_copy =
BlockwiseTensorSliceTransfer_v4<BlockSize,
AElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
Sequence<K0PerBlock, MPerBlock, K1>,
ABlockTransferThreadSliceLengths_K0_M_K1,
......@@ -416,16 +417,19 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r6
1,
1,
AThreadTransferSrcResetCoordinateAfterRun,
true>(a_grid_desc_k0_m_k1,
make_multi_index(0, m_block_data_idx_on_grid, 0),
a_block_desc_k0_m_k1,
make_multi_index(0, 0, 0),
a_element_op);
true>(
a_grid_desc_k0_m_k1,
make_multi_index(0, m_block_data_idx_on_grid, 0),
a_element_op,
a_block_desc_k0_m_k1,
make_multi_index(0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
// B matrix blockwise copy
auto b_blockwise_copy =
BlockwiseTensorSliceTransfer_v4<BlockSize,
BElementwiseOperation,
ck::tensor_operation::element_wise::PassThrough,
InMemoryDataOperationEnum_t::Set,
Sequence<K0PerBlock, NPerBlock, K1>,
BBlockTransferThreadSliceLengths_K0_N_K1,
......@@ -444,11 +448,13 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r6
1,
1,
BThreadTransferSrcResetCoordinateAfterRun,
true>(b_grid_desc_k0_n_k1,
make_multi_index(0, n_block_data_idx_on_grid, 0),
b_block_desc_k0_n_k1,
make_multi_index(0, 0, 0),
b_element_op);
true>(
b_grid_desc_k0_n_k1,
make_multi_index(0, n_block_data_idx_on_grid, 0),
b_element_op,
b_block_desc_k0_n_k1,
make_multi_index(0, 0, 0),
ck::tensor_operation::element_wise::PassThrough{});
// GEMM definition
// c_mtx += transpose(a_mtx) * b_mtx
......@@ -505,11 +511,14 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r6
b_blockwise_copy.RunWrite(b_block_desc_k0_n_k1, b_block_buf);
}
// main body
index_t k0_block_data_begin = 0;
// Initialize C
c_thread_buf.Clear();
// main body
if constexpr(HasMainKBlockLoop)
{
index_t k0_block_data_begin = 0;
do
{
a_blockwise_copy.MoveSrcSliceWindow(a_grid_desc_k0_m_k1,
......
composable_kernel/include/tensor_operation/gridwise_gemm_xdlops_v3r1.hpp
View file @ 491437a7
......@@ -459,7 +459,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
b_blockwise_copy.RunWrite(b_block_desc_k0_n_k1, b_block_buf);
}
// clear C
// Initialize C
c_thread_buf.Clear();
// main body
......
composable_kernel/include/tensor_operation/threadwise_tensor_slice_transfer.hpp
View file @ 491437a7
......@@ -290,7 +290,7 @@ struct ThreadwiseTensorSliceTransfer_v1r3
const DstDesc& dst_desc,
DstBuffer& dst_buf)
{
constexpr index_t ntransform_dst = DstDesc::GetNumOfTransform();
constexpr index_t ntransform_dst = remove_cvref_t<DstDesc>::GetNumOfTransform();
constexpr auto zeros = typename uniform_sequence_gen<ntransform_dst, 0>::type{};
......
device_operation/include/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk.hpp
View file @ 491437a7
......@@ -209,46 +209,6 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
using BGridDesc_K0_N_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
using CGridDesc_M_N = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
// TODO remove these hacks
static constexpr auto a_k0_m_k1_grid_step_hacks = make_tuple(
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0>{}, // 0+: K0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0>{}, // 1+: M
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0>{}), // 2+: K1
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0>{}, // 0-: K0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0>{}, // 1-: M
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0, 0>{})); // 2-: K1
static constexpr auto b_k0_n_k1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0>{}, // 0+: K0
Sequence<0, 0, 0, 0, 0>{}, // 1+: N
Sequence<0, 0, 0, 0, 0>{}), // 2+: K1
make_tuple(Sequence<0, 0, 0, 0, 0>{}, // 0-: K0
Sequence<0, 0, 0, 0, 0>{}, // 1-: N
Sequence<0, 0, 0, 0, 0>{})); // 2-: K1
static constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1+: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2+: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3+: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4+: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5+: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6+: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 7+: N2
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2-: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3-: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4-: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5-: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6-: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 7-: N2
static constexpr auto a_k0_m_k1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 2, 0, 0, 0, 0, 0>{};
static constexpr auto b_k0_n_k1_grid_move_slice_window_step_hacks = Sequence<0, 0, 0, 0, 0>{};
// GridwiseGemm
using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<
BlockSize,
......@@ -289,19 +249,16 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
Sequence<2, 3, 0, 1, 7, 5, 4, 6>, // CThreadTransferSrcDstAccessOrder,
7, // CThreadTransferSrcDstVectorDim,
CThreadTransferDstScalarPerVector,
decltype(a_k0_m_k1_grid_step_hacks), // AGridStepHacks,
decltype(b_k0_n_k1_grid_step_hacks), // BGridStepHacks,
decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks), // CGridStepHacks,
decltype(a_k0_m_k1_grid_move_slice_window_step_hacks), // AGridMoveSliceWindowStepHacks,
decltype(b_k0_n_k1_grid_move_slice_window_step_hacks), // BGridMoveSliceWindowStepHacks,
false, // CAccessOrderMRepeatNRepeat,
false, // CAccessOrderMRepeatNRepeat,
ABlockLdsAddExtraM,
BBlockLdsAddExtraN>;
#if 0
using CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 =
decltype(GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(CGridDesc_M_N{}));
using Block2CTileMap = decltype(GridwiseGemm::MakeBlock2CTileMap(CGridDesc_M_N{}, 1, 1));
#endif
// Argument
struct Argument : public BaseArgument
......@@ -371,8 +328,9 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
CGridDesc_M_N c_grid_desc_m_n_;
CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_;
Block2CTileMap block_2_ctile_map_;
typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_;
typename GridwiseGemm::Block2CTileMap block_2_ctile_map_;
index_t M01_;
index_t N01_;
InElementwiseOperation in_element_op_;
......@@ -426,11 +384,11 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
CDataType,
remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
remove_reference_t<DeviceOp::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
InElementwiseOperation,
WeiElementwiseOperation,
OutElementwiseOperation,
remove_reference_t<DeviceOp::Block2CTileMap>,
remove_reference_t<typename GridwiseGemm::Block2CTileMap>,
true>;
ave_time = launch_and_time_kernel(kernel,
......@@ -457,11 +415,11 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K
CDataType,
remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
remove_reference_t<DeviceOp::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
InElementwiseOperation,
WeiElementwiseOperation,
OutElementwiseOperation,
remove_reference_t<DeviceOp::Block2CTileMap>,
remove_reference_t<typename GridwiseGemm::Block2CTileMap>,
false>;
ave_time = launch_and_time_kernel(kernel,
......
device_operation/include/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_p0.hpp
View file @ 491437a7
......@@ -168,45 +168,6 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K_1x1_P0
using BGridDesc_K0_N_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
using CGridDesc_M_N = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
// TODO remove these hacks
static constexpr auto a_k0_m_k1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: K0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1+: M
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 2+: K1
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: K0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: M
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 2-: K1
static constexpr auto b_k0_n_k1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0>{}, // 0+: K0
Sequence<0, 0, 0>{}, // 1+: N
Sequence<0, 0, 0>{}), // 2+: K1
make_tuple(Sequence<0, 0, 0>{}, // 0-: K0
Sequence<0, 0, 0>{}, // 1-: N
Sequence<0, 0, 0>{})); // 2-: K1
static constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1+: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2+: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3+: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4+: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5+: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6+: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 7+: N2
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2-: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3-: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4-: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5-: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6-: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 7-: N2
static constexpr auto a_k0_m_k1_grid_move_slice_window_step_hacks =
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{};
static constexpr auto b_k0_n_k1_grid_move_slice_window_step_hacks = Sequence<0, 0, 0>{};
// GridwiseGemm
using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<
BlockSize,
......@@ -247,20 +208,10 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K_1x1_P0
Sequence<2, 3, 0, 1, 7, 5, 4, 6>, // CThreadTransferSrcDstAccessOrder,
7, // CThreadTransferSrcDstVectorDim,
CThreadTransferDstScalarPerVector,
decltype(a_k0_m_k1_grid_step_hacks), // AGridStepHacks,
decltype(b_k0_n_k1_grid_step_hacks), // BGridStepHacks,
decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks), // CGridStepHacks,
decltype(a_k0_m_k1_grid_move_slice_window_step_hacks), // AGridMoveSliceWindowStepHacks,
decltype(b_k0_n_k1_grid_move_slice_window_step_hacks), // BGridMoveSliceWindowStepHacks,
false, // CAccessOrderMRepeatNRepeat,
false, // CAccessOrderMRepeatNRepeat,
ABlockLdsAddExtraM,
BBlockLdsAddExtraN>;
using CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 =
decltype(GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(CGridDesc_M_N{}));
using Block2CTileMap = decltype(GridwiseGemm::MakeBlock2CTileMap(CGridDesc_M_N{}, 1, 1));
// Argument
struct Argument : public BaseArgument
{
......@@ -332,8 +283,9 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K_1x1_P0
AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
CGridDesc_M_N c_grid_desc_m_n_;
CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_;
Block2CTileMap block_2_ctile_map_;
typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_;
typename GridwiseGemm::Block2CTileMap block_2_ctile_map_;
index_t M01_;
index_t N01_;
InElementwiseOperation in_element_op_;
......@@ -391,11 +343,11 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K_1x1_P0
CDataType,
remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
remove_reference_t<DeviceOp::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
InElementwiseOperation,
WeiElementwiseOperation,
OutElementwiseOperation,
remove_reference_t<DeviceOp::Block2CTileMap>,
remove_reference_t<typename GridwiseGemm::Block2CTileMap>,
true>;
ave_time = launch_and_time_kernel(kernel,
......@@ -422,11 +374,11 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K_1x1_P0
CDataType,
remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
remove_reference_t<DeviceOp::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
InElementwiseOperation,
WeiElementwiseOperation,
OutElementwiseOperation,
remove_reference_t<DeviceOp::Block2CTileMap>,
remove_reference_t<typename GridwiseGemm::Block2CTileMap>,
false>;
ave_time = launch_and_time_kernel(kernel,
......
device_operation/include/device_conv2d_fwd_xdl_nhwc_kyxc_nhwk_1x1_s1_p0.hpp
View file @ 491437a7
......@@ -145,44 +145,6 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K_1x1_S1
using BGridDesc_K0_N_K1 = remove_cvref_t<decltype(ABCGridDescs{}[I1])>;
using CGridDesc_M_N = remove_cvref_t<decltype(ABCGridDescs{}[I2])>;
// TODO remove these hacks
static constexpr auto a_k0_m_k1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0>{}, // 0+: K0
Sequence<0, 0, 0>{}, // 1+: M
Sequence<0, 0, 0>{}), // 2+: K1
make_tuple(Sequence<0, 0, 0>{}, // 0-: K0
Sequence<0, 0, 0>{}, // 1-: M
Sequence<0, 0, 0>{})); // 2-: K1
static constexpr auto b_k0_n_k1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0>{}, // 0+: K0
Sequence<0, 0, 0>{}, // 1+: N
Sequence<0, 0, 0>{}), // 2+: K1
make_tuple(Sequence<0, 0, 0>{}, // 0-: K0
Sequence<0, 0, 0>{}, // 1-: N
Sequence<0, 0, 0>{})); // 2-: K1
static constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1+: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2+: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3+: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4+: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5+: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6+: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 7+: N2
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2-: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3-: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4-: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5-: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6-: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 7-: N2
static constexpr auto a_k0_m_k1_grid_move_slice_window_step_hacks = Sequence<0, 0, 0>{};
static constexpr auto b_k0_n_k1_grid_move_slice_window_step_hacks = Sequence<0, 0, 0>{};
// GridwiseGemm
using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<
BlockSize,
......@@ -223,20 +185,10 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K_1x1_S1
Sequence<2, 3, 0, 1, 7, 5, 4, 6>, // CThreadTransferSrcDstAccessOrder,
7, // CThreadTransferSrcDstVectorDim,
CThreadTransferDstScalarPerVector,
decltype(a_k0_m_k1_grid_step_hacks), // AGridStepHacks,
decltype(b_k0_n_k1_grid_step_hacks), // BGridStepHacks,
decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks), // CGridStepHacks,
decltype(a_k0_m_k1_grid_move_slice_window_step_hacks), // AGridMoveSliceWindowStepHacks,
decltype(b_k0_n_k1_grid_move_slice_window_step_hacks), // BGridMoveSliceWindowStepHacks,
false, // CAccessOrderMRepeatNRepeat,
false, // CAccessOrderMRepeatNRepeat,
ABlockLdsAddExtraM,
BBlockLdsAddExtraN>;
using CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 =
decltype(GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(CGridDesc_M_N{}));
using Block2CTileMap = decltype(GridwiseGemm::MakeBlock2CTileMap(CGridDesc_M_N{}, 1, 1));
// Argument
struct Argument : public BaseArgument
{
......@@ -309,8 +261,9 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K_1x1_S1
AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
CGridDesc_M_N c_grid_desc_m_n_;
CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_;
Block2CTileMap block_2_ctile_map_;
typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_;
typename GridwiseGemm::Block2CTileMap block_2_ctile_map_;
index_t M01_;
index_t N01_;
InElementwiseOperation in_element_op_;
......@@ -369,11 +322,11 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K_1x1_S1
CDataType,
remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
remove_reference_t<DeviceOp::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
InElementwiseOperation,
WeiElementwiseOperation,
OutElementwiseOperation,
remove_reference_t<DeviceOp::Block2CTileMap>,
remove_reference_t<typename GridwiseGemm::Block2CTileMap>,
true>;
ave_time = launch_and_time_kernel(kernel,
......@@ -400,11 +353,11 @@ struct DeviceConv2dFwdXdl_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K_1x1_S1
CDataType,
remove_reference_t<DeviceOp::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceOp::BGridDesc_K0_N_K1>,
remove_reference_t<DeviceOp::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
InElementwiseOperation,
WeiElementwiseOperation,
OutElementwiseOperation,
remove_reference_t<DeviceOp::Block2CTileMap>,
remove_reference_t<typename GridwiseGemm::Block2CTileMap>,
false>;
ave_time = launch_and_time_kernel(kernel,
......
device_operation/include/device_gemm_xdl.hpp
View file @ 491437a7
......@@ -132,45 +132,6 @@ struct DeviceGemmXdl
using BGridDesc_K0_N_K1 = decltype(MakeBGridDescriptor_K0_N_K1(1, 1, 1));
using CGridDesc_M_N = decltype(MakeCGridDescriptor_M_N(1, 1, 1));
// TODO remove these hacks
static constexpr auto a_k0_m_k1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0>{}, // 0+: K0
Sequence<0, 0, 0>{}, // 1+: M
Sequence<0, 0, 0>{}), // 2+: K1
make_tuple(Sequence<0, 0, 0>{}, // 0-: K0
Sequence<0, 0, 0>{}, // 1-: M
Sequence<0, 0, 0>{})); // 2-: K1
static constexpr auto b_k0_n_k1_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0>{}, // 0+: K0
Sequence<0, 0, 0>{}, // 1+: N
Sequence<0, 0, 0>{}), // 2+: K1
make_tuple(Sequence<0, 0, 0>{}, // 0-: K0
Sequence<0, 0, 0>{}, // 1-: N
Sequence<0, 0, 0>{})); // 2-: K1
static constexpr auto c_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks =
make_tuple(make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0+: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1+: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2+: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3+: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4+: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5+: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6+: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}), // 7+: N2
make_tuple(Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 0-: M0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 1-: N0
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 2-: M1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 3-: N1
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 4-: M2
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 5-: M3
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{}, // 6-: M4
Sequence<0, 0, 0, 0, 0, 0, 0, 0, 0>{})); // 7-: N2
static constexpr auto a_k0_m_k1_grid_move_slice_window_step_hacks = Sequence<0, 0, 0>{};
static constexpr auto b_k0_n_k1_grid_move_slice_window_step_hacks = Sequence<0, 0, 0>{};
// GridwiseGemm
using GridwiseGemm = GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v2r3<
BlockSize,
......@@ -211,20 +172,10 @@ struct DeviceGemmXdl
Sequence<0, 2, 4, 5, 6, 1, 3, 7>, // CThreadTransferSrcDstAccessOrder,
CThreadTransferSrcDstVectorDim,
CThreadTransferDstScalarPerVector,
decltype(a_k0_m_k1_grid_step_hacks), // AGridStepHacks,
decltype(b_k0_n_k1_grid_step_hacks), // BGridStepHacks,
decltype(c_m0_n0_m1_n1_m2_m3_m4_n2_grid_step_hacks), // CGridStepHacks,
decltype(a_k0_m_k1_grid_move_slice_window_step_hacks), // AGridMoveSliceWindowStepHacks,
decltype(b_k0_n_k1_grid_move_slice_window_step_hacks), // BGridMoveSliceWindowStepHacks,
false, // CAccessOrderMRepeatNRepeat,
false, // CAccessOrderMRepeatNRepeat,
ABlockLdsAddExtraM,
BBlockLdsAddExtraN>;
using CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 =
decltype(GridwiseGemm::MakeCGridDescriptor_M0_N0_M1_N1_M2_M3_M4_N2(CGridDesc_M_N{}));
using Block2CTileMap = decltype(GridwiseGemm::MakeBlock2CTileMap(CGridDesc_M_N{}, 1, 1));
// Argument
struct Argument : public BaseArgument
{
......@@ -277,8 +228,9 @@ struct DeviceGemmXdl
AGridDesc_K0_M_K1 a_grid_desc_k0_m_k1_;
BGridDesc_K0_N_K1 b_grid_desc_k0_n_k1_;
CGridDesc_M_N c_grid_desc_m_n_;
CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2 c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_;
Block2CTileMap block_2_ctile_map_;
typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2
c_grid_desc_m0_n0_m1_n1_m2_m3_m4_n2_;
typename GridwiseGemm::Block2CTileMap block_2_ctile_map_;
index_t M01_;
index_t N01_;
AElementwiseOperation a_element_op_;
......@@ -332,11 +284,11 @@ struct DeviceGemmXdl
CDataType,
remove_reference_t<DeviceGemmXdl::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceGemmXdl::BGridDesc_K0_N_K1>,
remove_reference_t<DeviceGemmXdl::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
remove_reference_t<DeviceGemmXdl::Block2CTileMap>,
remove_reference_t<typename GridwiseGemm::Block2CTileMap>,
true>;
ave_time = launch_and_time_kernel(kernel,
......@@ -363,11 +315,11 @@ struct DeviceGemmXdl
CDataType,
remove_reference_t<DeviceGemmXdl::AGridDesc_K0_M_K1>,
remove_reference_t<DeviceGemmXdl::BGridDesc_K0_N_K1>,
remove_reference_t<DeviceGemmXdl::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
remove_reference_t<typename GridwiseGemm::CGridDesc_M0_N0_M1_N1_M2_M3_M4_N2>,
AElementwiseOperation,
BElementwiseOperation,
CElementwiseOperation,
remove_reference_t<DeviceGemmXdl::Block2CTileMap>,
remove_reference_t<typename GridwiseGemm::Block2CTileMap>,
false>;
ave_time = launch_and_time_kernel(kernel,
......
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