Skip to content

GitLab

Commit a79b95eb authored by Chao Liu's avatar Chao Liu
Browse files

Merge remote-tracking branch 'origin/develop' into int8_qunatization_gemm_xdl

parents 487d9868 7a9b93f4
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 1341 additions and 57 deletions
composable_kernel/include/tensor_operation/element_wise_operation.hpp
View file @ a79b95eb
......@@ -157,7 +157,8 @@ struct RequantReluRequant
float gemm_requant = scaleGemm_ * static_cast<float>(x);
float relu = gemm_requant > 0 ? gemm_requant : 0;
float relu_requant = scaleRelu_ * relu;
y = static_cast<int8_t>(relu_requant > 127 ? 127 : relu_requant < -128 ? -128 : relu_requant);
y = static_cast<int8_t>(relu_requant > 127 ? 127
: relu_requant < -128 ? -128 : relu_requant);
}
// for reference_gemm
......@@ -166,7 +167,8 @@ struct RequantReluRequant
float gemm_requant = scaleGemm_ * x;
float relu = gemm_requant > 0 ? gemm_requant : 0;
float relu_requant = scaleRelu_ * relu;
y = static_cast<float>(relu_requant > 127 ? 127 : relu_requant < -128 ? -128 : relu_requant);
y = static_cast<float>(relu_requant > 127 ? 127
: relu_requant < -128 ? -128 : relu_requant);
}
float scaleGemm_;
......
composable_kernel/include/tensor_operation/gridwise_gemm_xdlops_v3r1.hpp
View file @ a79b95eb
......@@ -61,7 +61,7 @@ template <
index_t BlockSize,
typename FloatAB,
typename FloatAcc,
typename ShuffleType,
typename FloatCShuffle,
typename FloatC,
InMemoryDataOperationEnum_t CGlobalMemoryDataOperation,
typename AGridDesc_AK0_M_AK1,
......@@ -203,7 +203,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
return math::max((a_block_space_size_aligned + b_block_space_size_aligned) *
sizeof(FloatAB),
c_block_size * sizeof(ShuffleType));
c_block_size * sizeof(FloatCShuffle));
}
// block_id to matrix tile idx (m0, n0) mapping are controlled by {M01, N01}
......@@ -566,8 +566,8 @@ 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 shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
static_cast<ShuffleType*>(p_shared),
auto c_shuffle_block_buf = make_dynamic_buffer<AddressSpaceEnum_t::Lds>(
static_cast<FloatCShuffle*>(p_shared),
c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl
.GetElementSpaceSize());
......@@ -595,9 +595,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
Sequence<2, 4, 5, 6>{},
Sequence<>{},
Sequence<1>{},
Sequence<3, 7>{})
);
Sequence<3, 7>{}));
// calculate origin of thread output tensor on global memory
// blockwise GEMM c matrix starting index
......@@ -630,7 +628,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
// VGPR to LDS
auto c_thread_copy_vgpr_to_lds =
ThreadwiseTensorSliceTransfer_v1r3<FloatAcc,
ShuffleType,
FloatCShuffle,
decltype(c_thread_desc_m0_n0_m1_n1_m2_m3_m4_n2),
decltype(c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2),
ck::tensor_operation::element_wise::PassThrough,
......@@ -671,7 +669,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
NWave * NPerXdl>, // BlockSliceLengths,
CBlockTransferClusterLengths_MBlock_MXdlPerWave_MWaveMPerXdl_NBlock_NXdlPerWave_NWaveNPerXdl,
Sequence<0, 1, 2, 3, 4, 5>, // typename ThreadClusterArrangeOrder,
ShuffleType, // typename SrcData,
FloatCShuffle, // typename SrcData,
FloatC, // typename DstData,
decltype(
c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl),
......@@ -720,7 +718,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
make_tuple(mxdlperwave, nxdlperwave, I0, I0, I0, I0, I0, I0),
c_thread_buf,
c_block_desc_m0_n0_m1_n1_m2_m3_m4_n2,
shuffle_block_buf);
c_shuffle_block_buf);
// make sure it's safe to do ds_read
block_sync_lds();
......@@ -728,7 +726,7 @@ struct GridwiseGemm_k0mk1_k0nk1_mn_xdlops_v3r1
// LDS to global
c_block_copy_lds_to_global.Run(
c_block_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
shuffle_block_buf,
c_shuffle_block_buf,
c_grid_desc_mblock_mxdlperwave_mwavemperxdl_nblock_nxdlperwave_nwavenperxdl,
c_grid_buf);
......
composable_kernel/include/tensor_operation/threadwise_tensor_slice_transfer.hpp
View file @ a79b95eb
......@@ -312,7 +312,7 @@ struct ThreadwiseTensorSliceTransfer_v2
constexpr auto forward_step = SpaceFillingCurve::GetForwardStep(idx_1d);
move_tensor_coordinate(
src_desc, src_coord_, make_tensor_coordinate_step(dst_desc, forward_step));
src_desc, src_coord_, make_tensor_coordinate_step(src_desc, forward_step));
}
});
......
device_operation/include/device_conv_backward_weight.hpp 0 → 100644
View file @ a79b95eb
#ifndef DEVICE_CONV_WRW_HPP
#define DEVICE_CONV_WRW_HPP
#include <iostream>
#include "device_base.hpp"
namespace ck {
namespace tensor_operation {
namespace device {
template <typename InElementwiseOperation,
typename WeiElementwiseOperation,
typename OutElementwiseOperation>
struct DeviceConvWrw : public BaseOperator
{
virtual std::unique_ptr<BaseArgument>
MakeArgumentPointer(const void* p_in,
void* p_wei,
const void* p_out,
ck::index_t N,
ck::index_t K,
ck::index_t C,
std::vector<ck::index_t> input_spatial_lengths,
std::vector<ck::index_t> filter_spatial_lengths,
std::vector<ck::index_t> output_spatial_lengths,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op,
ck::index_t split_k) = 0;
virtual std::unique_ptr<BaseInvoker> MakeInvokerPointer() = 0;
};
template <typename InElementwiseOperation,
typename WeiElementwiseOperation,
typename OutElementwiseOperation>
using DeviceConvWrwPtr = std::unique_ptr<
DeviceConvWrw<InElementwiseOperation, WeiElementwiseOperation, OutElementwiseOperation>>;
} // namespace device
} // namespace tensor_operation
} // namespace ck
#endif
device_operation/include/device_gemm_xdl_c_shuffle.hpp
View file @ a79b95eb
......@@ -20,7 +20,7 @@ template <
typename BDataType,
typename CDataType,
typename AccDataType,
typename ShuffleDataType,
typename CShuffleDataType,
typename ALayout,
typename BLayout,
typename CLayout,
......@@ -136,7 +136,7 @@ struct DeviceGemmXdl_C_Shuffle
BlockSize,
ADataType, // TODO: distinguish A/B datatype
AccDataType,
ShuffleDataType,
CShuffleDataType,
CDataType,
InMemoryDataOperationEnum_t::Set,
AGridDesc_K0_M_K1,
......
device_operation/src/device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instance.cpp
View file @ a79b95eb
......@@ -23,23 +23,23 @@ using PassThrough = ck::tensor_operation::element_wise::PassThrough;
// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
using device_gemm_xdl_c_shuffle_bf16_bf16_bf16_mk_nk_mn_instances = std::tuple<
// clang-format off
//#####################| AData| BData| CData| AccData| Shuffle| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//#####################| Type| Type| Type| Type| Data| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//#####################| | | | | Type| | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, BF16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, BF16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, BF16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, BF16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, BF16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, BF16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, BF16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, BF16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, BF16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, BF16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, BF16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, BF16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, BF16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>
//#####################| AData| BData| CData| AccData| CShuffle| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//#####################| Type| Type| Type| Type| DataType| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//#####################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, BF16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, BF16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, BF16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, BF16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, BF16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, BF16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, BF16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, BF16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, BF16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, BF16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, BF16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, BF16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle< BF16, BF16, BF16, F32, BF16, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>
// clang-format on
>;
......
device_operation/src/device_gemm_xdl_c_shuffle_int8_int8_int8_mk_nk_mn_instance.cpp
View file @ a79b95eb
......@@ -22,23 +22,23 @@ using PassThrough = ck::tensor_operation::element_wise::PassThrough;
// Compilation parameters for a[m, k] * b[n, k] = c[m, n]
using device_gemm_xdl_c_shuffle_int8_int8_int8_mk_nk_mn_instances = std::tuple<
// clang-format off
//#####################| AData| BData| CData| AccData| Shuffle| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//#####################| Type| Type| Type| Type| Data| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//#####################| | | | | Type| | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>
//#####################| AData| BData| CData| AccData| CShuffle| ALayout| BLayout| CLayout| A| B| C| Block| MPer| NPer| KPer| AK1| BK1| MPer| NPer| MXdl| NXdl| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockTransfer| ABlockLds| BBlockTransfer| BBlockTransfer| BBlockTransfer| BlockTransfer| BBlockTransfer| BBlockTransfer| BBlockLds| CShuffle| CShuffle| CBlockTransferClusterLengths| CBlockTransfer|
//#####################| Type| Type| Type| Type| DataType| | | | Elementwise| Elementwise| Elementwise| Size| Block| Block| Block| | | XDL| XDL| Per| Per| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraM| ThreadCluster| ThreadCluster| SrcAccessOrder| SrcVectorDim| SrcScalar| DstScalar| AddExtraN| MXdlPerWave| NXdlPerWave| _MBlock_MXdlPerWave_MWaveMPerXdl| ScalarPerVector|
//#####################| | | | | | | | | Operation| Operation| Operation| | | | | | | | | Wave| Wave| Lengths_K0_M_K1| ArrangeOrder| | | PerVector| PerVector_K1| | Lengths_K0_N_K1| ArrangeOrder| | | PerVector| PerVector_K1| | PerShuffle| PerShuffle| _NBlock_NXdlPerWave_NWaveNPerXdl| _NWaveNPerXdl|
//#####################| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 256, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 256, 32, 8, 8, 32, 32, 2, 4, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 128, 32, 8, 8, 32, 32, 4, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 64, 128, 32, 8, 8, 32, 32, 2, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 64, 32, 8, 8, 32, 32, 2, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 128, 64, 32, 8, 8, 32, 32, 2, 1, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 256, 64, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 64, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 128, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 32, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 128, 32, 128, 32, 8, 8, 32, 32, 1, 2, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 32, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 8>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 64, 32, 32, 8, 8, 32, 32, 2, 1, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>,
DeviceGemmXdl_C_Shuffle<int8_t, int8_t, int8_t, int32_t, int32_t, Row, Col, Row, PassThrough, PassThrough, PassThrough, 64, 32, 64, 32, 8, 8, 32, 32, 1, 2, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, S<4, 16, 1>, S<1, 0, 2>, S<1, 0, 2>, 2, 8, 8, true, 1, 1, S<1, 1, 16, 1, 1, 4>, 8>
// clang-format on
>;
......
example/13_conv2d_backward_weight_xdl/README.md 0 → 100644
View file @ a79b95eb
# Instructions for ```conv2d_wrw_xdl``` Example
## Docker script
```bash
docker run \
-it \
--rm \
--privileged \
--group-add sudo \
-w /root/workspace \
-v ${PATH_TO_LOCAL_WORKSPACE}:/root/workspace \
rocm/tensorflow:rocm4.3.1-tf2.6-dev \
/bin/bash
```
## Build ```conv2d_wrw_xdl```
```bash
mkdir build && cd build
```
```bash
# Need to specify target ID, example below is gfx908
cmake \
-D BUILD_DEV=OFF \
-D CMAKE_BUILD_TYPE=Release \
-D CMAKE_CXX_FLAGS="-DCK_AMD_GPU_GFX908 --amdgpu-target=gfx908 -O3 " \
-D CMAKE_CXX_COMPILER=/opt/rocm/bin/hipcc \
-D CMAKE_PREFIX_PATH=/opt/rocm \
..
```
```bash
make -j conv2d_wrw_xdl
```
## Run ```conv2d_wrw_xdl```
```bash
#arg1: verification (0=no, 1=yes)
#arg2: initialization (0=no init, 1=integer value, 2=decimal value)
#arg3: run kernel # of times (>1)
#arg4: is show log (0=no, 1=yes)
#arg5 to 19: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, RightPx, split-k
./example/conv2d_fwd_xdl 0 1 5 0 4
```
Result
```
in_n_c_hi_wi: dim 4, lengths {128, 1024, 14, 14}, strides {200704, 1, 14336, 1024}
wei_k_c_y_x: dim 4, lengths {256, 1024, 3, 3}, strides {9216, 1, 3072, 1024}
out_n_k_ho_wo: dim 4, lengths {128, 256, 6, 6}, strides {9216, 1, 1536, 256}
arg.a_grid_desc_kbatch_k0_m_k1_{4, 144, 256, 8}
arg.b_grid_desc_kbatch_k0_n_k1_{4, 144, 9216, 8}
arg.c_grid_desc_m_n_{ 256, 9216}
launch_and_time_kernel: grid_dim {576, 1, 1}, block_dim {256, 1, 1}
Warm up
Start running 5 times...
Perf: 0.401084 ms, 54.2112 TFlops, 145.75 GB/s
```
example/13_conv2d_backward_weight_xdl/main.cpp 0 → 100644
View file @ a79b95eb
#include <iostream>
#include <numeric>
#include <initializer_list>
#include <cstdlib>
#include <stdlib.h>
#include <half.hpp>
#include "config.hpp"
#include "print.hpp"
#include "device.hpp"
#include "host_tensor.hpp"
#include "host_tensor_generator.hpp"
#include "device_tensor.hpp"
#include "tensor_layout.hpp"
#include "element_wise_operation.hpp"
#include "device_conv2d_backward_weight_xdl_c_shuffle_nhwc_kyxc_nhwk.hpp"
#include "reference_conv_backward_weight.hpp"
using InDataType = ck::half_t;
using WeiDataType = ck::half_t;
using OutDataType = ck::half_t;
using AccDataType = float;
template <ck::index_t... Is>
using S = ck::Sequence<Is...>;
using InLayout = ck::tensor_layout::convolution::NHWC;
using WeiLayout = ck::tensor_layout::convolution::KYXC;
using OutLayout = ck::tensor_layout::convolution::NHWK;
using InElementOp = ck::tensor_operation::element_wise::PassThrough;
using WeiElementOp = ck::tensor_operation::element_wise::PassThrough;
using OutElementOp = ck::tensor_operation::element_wise::PassThrough;
// clang-format off
using DeviceConvWrWInstance = ck::tensor_operation::device::
DeviceConv2dWrWXdl_C_Shuffle_Input_N_Hi_Wi_C_Weight_K_Y_X_C_Output_N_Ho_Wo_K<
InDataType, // InDataType
WeiDataType, // WeiDataType
OutDataType, // OutDataType
AccDataType, // AccDataType
InElementOp, // InElementwiseOperation
WeiElementOp, // WeiElementwiseOperation
OutElementOp, // OutElementwiseOperation
256, // BlockSize
128, // MPerBlock
128, // NPerBlock
4, // K0PerBlock
8, // K1
32, // MPerXdl
32, // NPerXdl
2, // MXdlPerWave
2, // NXdlPerWave
S<1, 4, 16, 4>, // ABlockTransferThreadClusterLengths_K0_M_K1
S<0, 3, 1, 2>, // ABlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // ABlockTransferSrcAccessOrder
2, // ABlockTransferSrcVectorDim
8, // ABlockTransferSrcScalarPerVector
2, // ABlockTransferDstScalarPerVector_K1
true, // ABlockLdsAddExtraM
S<1, 4, 16, 4>, // BBlockTransferThreadClusterLengths_K0_N_K1
S<0, 3, 1, 2>, // BBlockTransferThreadClusterArrangeOrder
S<0, 2, 1, 3>, // BBlockTransferSrcAccessOrder
2, // BBlockTransferSrcVectorDim
8, // BBlockTransferSrcScalarPerVector
2, // BBlockTransferDstScalarPerVector_K1
true, // BBlockLdsAddExtraN
1, // CShuffleMXdlPerWavePerShuffle
1, // CShuffleNXdlPerWavePerShuffle
S<1, 32, 1, 4>, // CBlockTransferClusterLengths_MBlock_MPerBlock_NBlock_NPerBlock
8>; // CBlockTransferScalarPerVector_NWaveNPerXdl
// clang-format on
using ReferenceConvWrwInstance = ck::tensor_operation::host::
ReferenceConvWrw<InDataType, WeiDataType, OutDataType, InElementOp, WeiElementOp, OutElementOp>;
int main(int argc, char* argv[])
{
bool do_verification = 0;
int init_method = 0;
int nrepeat = 5;
int do_log = 0;
int split_k = 4;
// Conv shape
ck::index_t N = 128;
ck::index_t K = 256;
ck::index_t C = 1024;
ck::index_t Y = 3;
ck::index_t X = 3;
ck::index_t Hi = 14;
ck::index_t Wi = 14;
ck::index_t conv_stride_h = 2;
ck::index_t conv_stride_w = 2;
ck::index_t conv_dilation_h = 1;
ck::index_t conv_dilation_w = 1;
ck::index_t in_left_pad_h = 0;
ck::index_t in_left_pad_w = 0;
ck::index_t in_right_pad_h = 0;
ck::index_t in_right_pad_w = 0;
if(argc == 6)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
do_log = std::stoi(argv[4]);
split_k = std::stoi(argv[5]);
}
else if(argc == 21)
{
do_verification = std::stoi(argv[1]);
init_method = std::stoi(argv[2]);
nrepeat = std::stoi(argv[3]);
do_log = std::stoi(argv[4]);
split_k = std::stoi(argv[5]);
N = std::stoi(argv[6]);
K = std::stoi(argv[7]);
C = std::stoi(argv[8]);
Y = std::stoi(argv[9]);
X = std::stoi(argv[10]);
Hi = std::stoi(argv[11]);
Wi = std::stoi(argv[12]);
conv_stride_h = std::stoi(argv[13]);
conv_stride_w = std::stoi(argv[14]);
conv_dilation_h = std::stoi(argv[15]);
conv_dilation_w = std::stoi(argv[16]);
in_left_pad_h = std::stoi(argv[17]);
in_left_pad_w = std::stoi(argv[18]);
in_right_pad_h = std::stoi(argv[19]);
in_right_pad_w = std::stoi(argv[20]);
}
else
{
printf("arg1: verification (0=no, 1=yes)\n");
printf("arg2: initialization (0=no init, 1=integer value, 2=decimal value)\n");
printf("arg3: run kernel # of times (>1)\n");
printf("arg4: is show log (0=no, 1=yes)\n");
printf("arg5: split-k \n");
printf("arg6 to 19: N, K, C, Y, X, Hi, Wi, Sy, Sx, Dy, Dx, LeftPy, LeftPx, RightPy, "
"RightPx\n");
exit(0);
}
const ck::index_t YEff = (Y - 1) * conv_dilation_h + 1;
const ck::index_t XEff = (X - 1) * conv_dilation_w + 1;
const ck::index_t Ho = (Hi + in_left_pad_h + in_right_pad_h - YEff) / conv_stride_h + 1;
const ck::index_t Wo = (Wi + in_left_pad_w + in_right_pad_w - XEff) / conv_stride_w + 1;
const std::vector<ck::index_t> conv_filter_strides{{conv_stride_h, conv_stride_w}};
const std::vector<ck::index_t> conv_filter_dilations{{conv_dilation_h, conv_dilation_w}};
const std::vector<ck::index_t> input_left_pads{{in_left_pad_h, in_left_pad_w}};
const std::vector<ck::index_t> input_right_pads{{in_right_pad_h, in_right_pad_w}};
// tensor layout
auto f_host_tensor_descriptor = [](std::size_t N_,
std::size_t C_,
std::size_t H,
std::size_t W,
auto layout) {
if constexpr(ck::is_same<decltype(layout), ck::tensor_layout::convolution::NCHW>::value ||
ck::is_same<decltype(layout), ck::tensor_layout::convolution::KCYX>::value ||
ck::is_same<decltype(layout), ck::tensor_layout::convolution::NKHW>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
std::vector<std::size_t>({C_ * H * W, H * W, W, 1}));
}
else if constexpr(ck::is_same<decltype(layout),
ck::tensor_layout::convolution::NHWC>::value ||
ck::is_same<decltype(layout),
ck::tensor_layout::convolution::KYXC>::value ||
ck::is_same<decltype(layout),
ck::tensor_layout::convolution::NHWK>::value)
{
return HostTensorDescriptor(std::vector<std::size_t>({N_, C_, H, W}),
std::vector<std::size_t>({C_ * H * W, 1, W * C_, C_}));
}
};
Tensor<InDataType> in_n_c_hi_wi(f_host_tensor_descriptor(N, C, Hi, Wi, InLayout{}));
Tensor<WeiDataType> wei_k_c_y_x_host_result(f_host_tensor_descriptor(K, C, Y, X, WeiLayout{}));
Tensor<WeiDataType> wei_k_c_y_x_device_result(
f_host_tensor_descriptor(K, C, Y, X, WeiLayout{}));
Tensor<OutDataType> out_n_k_ho_wo(f_host_tensor_descriptor(N, K, Ho, Wo, OutLayout{}));
std::cout << "in_n_c_hi_wi: " << in_n_c_hi_wi.mDesc << std::endl;
std::cout << "wei_k_c_y_x: " << wei_k_c_y_x_host_result.mDesc << std::endl;
std::cout << "out_n_k_ho_wo: " << out_n_k_ho_wo.mDesc << std::endl;
switch(init_method)
{
case 0: break;
case 1:
in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_2<InDataType>{-5, 5});
out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_2<OutDataType>{-5, 5});
break;
default:
in_n_c_hi_wi.GenerateTensorValue(GeneratorTensor_1<InDataType>{1});
out_n_k_ho_wo.GenerateTensorValue(GeneratorTensor_1<OutDataType>{1});
}
wei_k_c_y_x_device_result.GenerateTensorValue(GeneratorTensor_1<WeiDataType>{0});
DeviceMem in_device_buf(sizeof(InDataType) * in_n_c_hi_wi.mDesc.GetElementSpace());
DeviceMem wei_device_buf(sizeof(WeiDataType) *
wei_k_c_y_x_device_result.mDesc.GetElementSpace());
DeviceMem out_device_buf(sizeof(OutDataType) * out_n_k_ho_wo.mDesc.GetElementSpace());
in_device_buf.ToDevice(in_n_c_hi_wi.mData.data());
out_device_buf.ToDevice(out_n_k_ho_wo.mData.data());
wei_device_buf.ToDevice(wei_k_c_y_x_device_result.mData.data());
// do GEMM
auto conv = DeviceConvWrWInstance{};
auto invoker = conv.MakeInvoker();
auto argument = conv.MakeArgument(static_cast<InDataType*>(in_device_buf.GetDeviceBuffer()),
static_cast<WeiDataType*>(wei_device_buf.GetDeviceBuffer()),
static_cast<OutDataType*>(out_device_buf.GetDeviceBuffer()),
N,
K,
C,
std::vector<ck::index_t>{{Hi, Wi}},
std::vector<ck::index_t>{{Y, X}},
std::vector<ck::index_t>{{Ho, Wo}},
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
InElementOp{},
WeiElementOp{},
OutElementOp{},
split_k);
if(!conv.IsSupportedArgument(argument))
{
std::cout << "wrong! device_conv with the specified compilation parameters does "
"not support this Conv problem"
<< std::endl;
return 1;
}
float ave_time = invoker.Run(argument, nrepeat);
std::size_t flop = std::size_t(2) * N * K * Ho * Wo * C * Y * X;
std::size_t num_btype = sizeof(InDataType) * (N * C * Hi * Wi) +
sizeof(WeiDataType) * (K * C * Y * X) +
sizeof(OutDataType) * (N * K * Ho * Wo);
float tflops = static_cast<float>(flop) / 1.E9 / ave_time;
float gb_per_sec = num_btype / 1.E6 / ave_time;
std::cout << "Perf: " << ave_time << " ms, " << tflops << " TFlops, " << gb_per_sec << " GB/s"
<< std::endl;
if(do_verification)
{
auto ref_conv = ReferenceConvWrwInstance{};
auto ref_invoker = ref_conv.MakeInvoker();
auto ref_argument = ref_conv.MakeArgument(in_n_c_hi_wi,
wei_k_c_y_x_host_result,
out_n_k_ho_wo,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
InElementOp{},
WeiElementOp{},
OutElementOp{});
ref_invoker.Run(ref_argument);
wei_device_buf.FromDevice(wei_k_c_y_x_device_result.mData.data());
if(do_log)
{
LogRangeAsType<float>(std::cout << "out: ", out_n_k_ho_wo.mData, ",") << std::endl;
LogRangeAsType<float>(std::cout << "in : ", in_n_c_hi_wi.mData, ",") << std::endl;
LogRangeAsType<float>(
std::cout << "wei_device(after): ", wei_k_c_y_x_device_result.mData, ",")
<< std::endl;
LogRangeAsType<float>(std::cout << "wei_host : ", wei_k_c_y_x_host_result.mData, ",")
<< std::endl;
}
check_error(wei_k_c_y_x_host_result, wei_k_c_y_x_device_result);
}
}
example/1_gemm_xdl/gemm_xdl_bf16.cpp
View file @ a79b95eb
......@@ -43,7 +43,7 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl_C_Shuffle
BDataType, // BDataType
CDataType, // CDataType
AccDataType, // AccDataType
CDataType, // ShuffleDataType
CDataType, // CShuffleDataType
ALayout, // ALayout
BLayout, // BLayout
CLayout, // CLayout
......
example/1_gemm_xdl/gemm_xdl_int8.cpp
View file @ a79b95eb
......@@ -28,11 +28,11 @@ using Col = ck::tensor_layout::gemm::ColumnMajor;
using PassThrough = ck::tensor_operation::element_wise::PassThrough;
using RequantReluRequant = ck::tensor_operation::element_wise::RequantReluRequant;
using ADataType = int8_t;
using BDataType = int8_t;
using CDataType = int8_t;
using AccDataType = int32_t;
using ShuffleDataType = int32_t;
using ADataType = int8_t;
using BDataType = int8_t;
using CDataType = int8_t;
using AccDataType = int32_t;
using CShuffleDataType = int32_t;
using ALayout = ck::tensor_layout::gemm::RowMajor;
using BLayout = ck::tensor_layout::gemm::ColumnMajor;
......@@ -44,7 +44,7 @@ using DeviceGemmInstance = ck::tensor_operation::device::DeviceGemmXdl_C_Shuffle
BDataType, // BDataType
CDataType, // CDataType
AccDataType, // AccDataType
ShuffleDataType, // ShuffleDataType
CShuffleDataType, // CShuffleDataType
ALayout, // ALayout
BLayout, // BLayout
CLayout, // CLayout
......
example/CMakeLists.txt
View file @ a79b95eb
......@@ -24,6 +24,7 @@ set(CONV2D_FWD_XDL_BIAS_RELU_ADD_SOURCE 6_conv2d_fwd_xdl_bias_relu_add/conv2d_fw
set(CONV2D_FWD_XDL_BIAS_RELU_ATOMIC_ADD_SOURCE 7_conv2d_fwd_xdl_bias_relu_atomic_add/conv2d_fwd_xdl_bias_relu_atomic_add.cpp)
set(GEMM_XDL_ALPHA_BETA_SOURCE 8_gemm_xdl_alpha_beta/gemm_xdl_alpha_beta.cpp)
set(CONV2D_FWD_XDL_INT8_SOURCE 9_conv2d_fwd_xdl_int8/conv2d_fwd_xdl_int8.cpp)
set(CONV2D_WRW_XDL_SOURCE 13_conv2d_backward_weight_xdl/main.cpp)
set(CONV3D_FWD_XDL_SOURCE 10_conv3d_fwd_xdl/conv3d_fwd_xdl.cpp)
set(CONVND_FWD_XDL_SOURCE 11_convnd_fwd_xdl/convnd_fwd_xdl.cpp)
set(CONV2D_BWD_DATA_XDL_SOURCE 12_conv2d_bwd_data_xdl/conv2d_bwd_data_xdl.cpp)
......@@ -39,6 +40,7 @@ add_executable(conv2d_fwd_xdl_bias_relu_add ${CONV2D_FWD_XDL_BIAS_RELU_ADD_SOURC
add_executable(conv2d_fwd_xdl_bias_relu_atomic_add ${CONV2D_FWD_XDL_BIAS_RELU_ATOMIC_ADD_SOURCE})
add_executable(gemm_xdl_alpha_beta ${GEMM_XDL_ALPHA_BETA_SOURCE})
add_executable(conv2d_fwd_xdl_int8 ${CONV2D_FWD_XDL_INT8_SOURCE})
add_executable(conv2d_wrw_xdl ${CONV2D_WRW_XDL_SOURCE})
add_executable(conv3d_fwd_xdl ${CONV3D_FWD_XDL_SOURCE})
add_executable(convnd_fwd_xdl ${CONVND_FWD_XDL_SOURCE})
add_executable(conv2d_bwd_data_xdl ${CONV2D_BWD_DATA_XDL_SOURCE})
......@@ -54,6 +56,7 @@ target_link_libraries(conv2d_fwd_xdl_bias_relu_add PRIVATE host_tensor)
target_link_libraries(conv2d_fwd_xdl_bias_relu_atomic_add PRIVATE host_tensor)
target_link_libraries(gemm_xdl_alpha_beta PRIVATE host_tensor)
target_link_libraries(conv2d_fwd_xdl_int8 PRIVATE host_tensor)
target_link_libraries(conv2d_wrw_xdl PRIVATE host_tensor)
target_link_libraries(conv3d_fwd_xdl PRIVATE host_tensor)
target_link_libraries(convnd_fwd_xdl PRIVATE host_tensor)
target_link_libraries(conv2d_bwd_data_xdl PRIVATE host_tensor)
......
reference_operation/include/reference_conv_backward_weight.hpp 0 → 100644
View file @ a79b95eb
#ifndef REFERENCE_CONV_WRW_HPP
#define REFERENCE_CONV_WRW_HPP
#include <iostream>
#include <sstream>
#include "device_base.hpp"
#include "host_tensor.hpp"
namespace ck {
namespace tensor_operation {
namespace host {
// out[N, K, Ho, Wo] = in[N, C, Hi, Wi] * wei[K, C, Y, X]
template <typename InDataType,
typename WeiDataType,
typename OutDataType,
typename InElementwiseOperation,
typename WeiElementwiseOperation,
typename OutElementwiseOperation>
struct ReferenceConvWrw : public device::BaseOperator
{
// Argument
struct Argument : public device::BaseArgument
{
Argument(const Tensor<InDataType>& in_n_c_hi_wi,
Tensor<WeiDataType>& wei_k_c_y_x,
const Tensor<OutDataType>& out_n_k_ho_wo,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op)
: in_n_c_hi_wi_{in_n_c_hi_wi},
wei_k_c_y_x_{wei_k_c_y_x},
out_n_k_ho_wo_{out_n_k_ho_wo},
conv_strides_{conv_filter_strides},
conv_dilations_{conv_filter_dilations},
in_left_pads_{input_left_pads},
in_right_pads_{input_right_pads},
in_element_op_{in_element_op},
wei_element_op_{wei_element_op},
out_element_op_{out_element_op}
{
}
const Tensor<InDataType>& in_n_c_hi_wi_;
Tensor<WeiDataType>& wei_k_c_y_x_;
const Tensor<OutDataType>& out_n_k_ho_wo_;
std::vector<index_t> conv_strides_;
std::vector<index_t> conv_dilations_;
std::vector<index_t> in_left_pads_;
std::vector<index_t> in_right_pads_;
InElementwiseOperation in_element_op_;
WeiElementwiseOperation wei_element_op_;
OutElementwiseOperation out_element_op_;
};
// Invoker
struct Invoker : public device::BaseInvoker
{
using Argument = ReferenceConvWrw::Argument;
float Run(const Argument& arg)
{
constexpr auto I0 = Number<0>{};
constexpr auto I1 = Number<1>{};
auto f_kcyx = [&](auto k, auto c, auto y, auto x) {
float v_acc = 0;
for(int n = 0; n < arg.out_n_k_ho_wo_.mDesc.GetLengths()[0]; ++n)
{
for(int ho = 0; ho < arg.out_n_k_ho_wo_.mDesc.GetLengths()[2]; ++ho)
{
int hi = ho * arg.conv_strides_[I0] + y * arg.conv_dilations_[I0] -
arg.in_left_pads_[I0];
for(int wo = 0; wo < arg.out_n_k_ho_wo_.mDesc.GetLengths()[3]; ++wo)
{
int wi = wo * arg.conv_strides_[I1] + x * arg.conv_dilations_[I1] -
arg.in_left_pads_[I1];
if(hi >= 0 && hi < arg.in_n_c_hi_wi_.mDesc.GetLengths()[2] && wi >= 0 &&
wi < arg.in_n_c_hi_wi_.mDesc.GetLengths()[3])
{
float v_out;
float v_in;
arg.out_element_op_(
v_out,
ck::type_convert<float>(arg.out_n_k_ho_wo_(n, k, ho, wo)));
arg.in_element_op_(
v_in, ck::type_convert<float>(arg.in_n_c_hi_wi_(n, c, hi, wi)));
v_acc += v_out * v_in;
}
}
}
}
float v_wei;
arg.wei_element_op_(v_wei, v_acc);
arg.wei_k_c_y_x_(k, c, y, x) = ck::type_convert<OutDataType>(v_wei);
};
make_ParallelTensorFunctor(f_kcyx,
arg.wei_k_c_y_x_.mDesc.GetLengths()[0],
arg.wei_k_c_y_x_.mDesc.GetLengths()[1],
arg.wei_k_c_y_x_.mDesc.GetLengths()[2],
arg.wei_k_c_y_x_.mDesc.GetLengths()[3])(
std::thread::hardware_concurrency());
return 0;
}
float Run(const device::BaseArgument* p_arg, int) override
{
return Run(*dynamic_cast<const Argument*>(p_arg));
}
};
static constexpr bool IsValidCompilationParameter()
{
// TODO: properly implement this check
return true;
}
bool IsSupportedArgument(const device::BaseArgument*) override { return true; }
static auto MakeArgument(const Tensor<InDataType>& in_n_c_hi_wi,
Tensor<WeiDataType>& wei_k_c_y_x,
const Tensor<OutDataType>& out_n_k_ho_wo,
std::vector<ck::index_t> conv_filter_strides,
std::vector<ck::index_t> conv_filter_dilations,
std::vector<ck::index_t> input_left_pads,
std::vector<ck::index_t> input_right_pads,
InElementwiseOperation in_element_op,
WeiElementwiseOperation wei_element_op,
OutElementwiseOperation out_element_op)
{
return Argument{in_n_c_hi_wi,
wei_k_c_y_x,
out_n_k_ho_wo,
conv_filter_strides,
conv_filter_dilations,
input_left_pads,
input_right_pads,
in_element_op,
wei_element_op,
out_element_op};
}
static auto MakeInvoker() { return Invoker{}; }
virtual std::unique_ptr<device::BaseInvoker> MakeInvokerPointer()
{
return std::make_unique<Invoker>(Invoker{});
}
std::string GetTypeString() const override
{
auto str = std::stringstream();
// clang-format off
str << "ReferenceConvFwd"
<< std::endl;
// clang-format on
return str.str();
}
};
} // namespace host
} // namespace tensor_operation
} // namespace ck
#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