-
Notifications
You must be signed in to change notification settings - Fork 197
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Add fused cosine 1-NN kernel and unify the fused distance 1-NN kernels
fix doc issue in fused_distance_nn runtime API
- Loading branch information
Showing
23 changed files
with
1,789 additions
and
333 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,89 @@ | ||
| /* | ||
| * Copyright (c) 2021-2024, NVIDIA CORPORATION. | ||
| * | ||
| * Licensed under the Apache License, Version 2.0 (the "License"); | ||
| * you may not use this file except in compliance with the License. | ||
| * You may obtain a copy of the License at | ||
| * | ||
| * http://www.apache.org/licenses/LICENSE-2.0 | ||
| * | ||
| * Unless required by applicable law or agreed to in writing, software | ||
| * distributed under the License is distributed on an "AS IS" BASIS, | ||
| * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | ||
| * See the License for the specific language governing permissions and | ||
| * limitations under the License. | ||
| */ | ||
|
|
||
| #pragma once | ||
|
|
||
| #include <cstddef> // size_t | ||
| #include <limits> // std::numeric_limits | ||
| #include <raft/core/kvp.hpp> // raft::KeyValuePair | ||
| #include <raft/core/operators.hpp> // raft::identity_op | ||
| #include <raft/distance/detail/distance_ops/l2_exp.cuh> // ops::l2_exp_distance_op | ||
| #include <raft/distance/detail/fused_distance_nn/cutlass_base.cuh> | ||
| #include <raft/distance/detail/fused_distance_nn/fused_cosine_nn.cuh> | ||
| #include <raft/distance/detail/fused_distance_nn/helper_structs.cuh> | ||
| #include <raft/distance/detail/fused_distance_nn/simt_kernel.cuh> | ||
| #include <raft/distance/detail/pairwise_distance_base.cuh> // PairwiseDistances | ||
| #include <raft/distance/distance_types.hpp> | ||
| #include <raft/linalg/contractions.cuh> // Policy | ||
| #include <raft/util/arch.cuh> // raft::util::arch::SM_* | ||
| #include <raft/util/cuda_utils.cuh> // raft::ceildiv, raft::shfl | ||
|
|
||
| namespace raft { | ||
| namespace distance { | ||
|
|
||
| namespace detail { | ||
|
|
||
| template <typename DataT, | ||
| typename OutT, | ||
| typename IdxT, | ||
| typename Policy, | ||
| typename ReduceOpT, | ||
| typename KVPReduceOpT> | ||
| void fusedDistanceNNImpl(OutT* min, | ||
| const DataT* x, | ||
| const DataT* y, | ||
| const DataT* xn, | ||
| const DataT* yn, | ||
| IdxT m, | ||
| IdxT n, | ||
| IdxT k, | ||
| int* workspace, | ||
| ReduceOpT redOp, | ||
| KVPReduceOpT pairRedOp, | ||
| bool sqrt, | ||
| bool initOutBuffer, | ||
| bool isRowMajor, | ||
| raft::distance::DistanceType metric, | ||
| float metric_arg, | ||
| cudaStream_t stream) | ||
| { | ||
| // The kernel policy is determined by fusedDistanceNN. | ||
| typedef Policy P; | ||
|
|
||
| dim3 blk(P::Nthreads); | ||
| auto nblks = raft::ceildiv<int>(m, P::Nthreads); | ||
| constexpr auto maxVal = std::numeric_limits<DataT>::max(); | ||
| typedef KeyValuePair<IdxT, DataT> KVPair; | ||
|
|
||
| RAFT_CUDA_TRY(cudaMemsetAsync(workspace, 0, sizeof(int) * m, stream)); | ||
| if (initOutBuffer) { | ||
| initKernel<DataT, OutT, IdxT, ReduceOpT> | ||
| <<<nblks, P::Nthreads, 0, stream>>>(min, m, maxVal, redOp); | ||
| RAFT_CUDA_TRY(cudaGetLastError()); | ||
| } | ||
|
|
||
| switch (metric) { | ||
| case DistanceType::CosineExpanded: | ||
| fusedCosineNN<DataT, OutT, IdxT, P, ReduceOpT, KVPReduceOpT>( | ||
| min, x, y, xn, yn, m, n, k, workspace, redOp, pairRedOp, sqrt, stream); | ||
| break; | ||
| default: assert("only cosine metric is supported with fusedDistanceNN\n"); break; | ||
| } | ||
| } | ||
|
|
||
| } // namespace detail | ||
| } // namespace distance | ||
| } // namespace raft |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
135 changes: 135 additions & 0 deletions
135
cpp/include/raft/distance/detail/fused_distance_nn/fused_cosine_nn.cuh
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,135 @@ | ||
| /* | ||
| * Copyright (c) 2021-2024, NVIDIA CORPORATION. | ||
| * | ||
| * Licensed under the Apache License, Version 2.0 (the "License"); | ||
| * you may not use this file except in compliance with the License. | ||
| * You may obtain a copy of the License at | ||
| * | ||
| * http://www.apache.org/licenses/LICENSE-2.0 | ||
| * | ||
| * Unless required by applicable law or agreed to in writing, software | ||
| * distributed under the License is distributed on an "AS IS" BASIS, | ||
| * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. | ||
| * See the License for the specific language governing permissions and | ||
| * limitations under the License. | ||
| */ | ||
|
|
||
| #pragma once | ||
|
|
||
| #include <cstddef> // size_t | ||
| #include <limits> // std::numeric_limits | ||
| #include <raft/core/kvp.hpp> // raft::KeyValuePair | ||
| #include <raft/core/operators.hpp> // raft::identity_op | ||
| #include <raft/distance/detail/distance_ops/cosine.cuh> // ops::l2_exp_distance_op | ||
| #include <raft/distance/detail/fused_distance_nn/cutlass_base.cuh> | ||
| #include <raft/distance/detail/fused_distance_nn/helper_structs.cuh> | ||
| #include <raft/distance/detail/fused_distance_nn/simt_kernel.cuh> | ||
| #include <raft/distance/detail/pairwise_distance_base.cuh> // PairwiseDistances | ||
| #include <raft/linalg/contractions.cuh> // Policy | ||
| #include <raft/util/arch.cuh> // raft::util::arch::SM_* | ||
| #include <raft/util/cuda_utils.cuh> // raft::ceildiv, raft::shfl | ||
|
|
||
| namespace raft { | ||
| namespace distance { | ||
|
|
||
| namespace detail { | ||
|
|
||
| template <typename DataT, | ||
| typename OutT, | ||
| typename IdxT, | ||
| typename Policy, | ||
| typename ReduceOpT, | ||
| typename KVPReduceOpT> | ||
| void fusedCosineNN(OutT* min, | ||
| const DataT* x, | ||
| const DataT* y, | ||
| const DataT* xn, | ||
| const DataT* yn, | ||
| IdxT m, | ||
| IdxT n, | ||
| IdxT k, | ||
| int* workspace, | ||
| ReduceOpT redOp, | ||
| KVPReduceOpT pairRedOp, | ||
| bool sqrt, | ||
| cudaStream_t stream) | ||
| { | ||
| // The kernel policy is determined by fusedL2NN. | ||
| typedef Policy P; | ||
|
|
||
| dim3 blk(P::Nthreads); | ||
| constexpr auto maxVal = std::numeric_limits<DataT>::max(); | ||
| typedef KeyValuePair<IdxT, DataT> KVPair; | ||
|
|
||
| namespace arch = raft::util::arch; | ||
| using AccT = DataT; | ||
| ops::cosine_distance_op<DataT, AccT, IdxT> distance_op{}; | ||
|
|
||
| raft::identity_op fin_op{}; | ||
|
|
||
| auto kernel = fusedDistanceNNkernel<DataT, | ||
| OutT, | ||
| IdxT, | ||
| P, | ||
| ReduceOpT, | ||
| KVPReduceOpT, | ||
| decltype(distance_op), | ||
| decltype(fin_op)>; | ||
|
|
||
| // Get pointer to fp32 SIMT kernel to determine the runtime architecture of the | ||
| // current system. Other methods to determine the architecture (that do not | ||
| // require a pointer) can be error prone. See: | ||
| // https://github.com/NVIDIA/cub/issues/545 | ||
| void* kernel_ptr = reinterpret_cast<void*>(kernel); | ||
| auto runtime_arch = arch::kernel_virtual_arch(kernel_ptr); | ||
| auto cutlass_range = arch::SM_range(arch::SM_80(), arch::SM_future()); | ||
|
|
||
| if (cutlass_range.contains(runtime_arch)) { | ||
| // If device is SM_80 or later, use CUTLASS-based kernel. | ||
| using cosineOp = raft::distance::detail::ops::cosine_cutlass_op<DataT, DataT>; | ||
| using kvp_cg_min_reduce_op_ = kvp_cg_min_reduce_op<DataT, IdxT, OutT>; | ||
| kvp_cg_min_reduce_op_ cg_reduce_op; | ||
| cosineOp cosine_dist_op; | ||
|
|
||
| IdxT lda, ldb, ldd; | ||
| lda = k, ldb = k, ldd = n; | ||
|
|
||
| cutlassFusedDistanceNN<DataT, | ||
| DataT, | ||
| OutT, | ||
| IdxT, | ||
| P::Veclen, | ||
| decltype(cg_reduce_op), | ||
| decltype(cosine_dist_op), | ||
| ReduceOpT, | ||
| KVPReduceOpT>(x, | ||
| y, | ||
| xn, | ||
| yn, | ||
| m, | ||
| n, | ||
| k, | ||
| lda, | ||
| ldb, | ||
| ldd, | ||
| min, | ||
| workspace, | ||
| cg_reduce_op, | ||
| cosine_dist_op, | ||
| redOp, | ||
| pairRedOp, | ||
| stream); | ||
| } else { | ||
| // If device less than SM_80, use fp32 SIMT kernel. | ||
| constexpr size_t shmemSize = P::SmemSize + ((P::Mblk + P::Nblk) * sizeof(DataT)); | ||
| dim3 grid = launchConfigGenerator<P>(m, n, shmemSize, kernel); | ||
|
|
||
| kernel<<<grid, blk, shmemSize, stream>>>( | ||
| min, x, y, xn, yn, m, n, k, maxVal, workspace, redOp, pairRedOp, distance_op, fin_op); | ||
| RAFT_CUDA_TRY(cudaGetLastError()); | ||
| } | ||
| } | ||
|
|
||
| } // namespace detail | ||
| } // namespace distance | ||
| } // namespace raft |
Oops, something went wrong.