-
Notifications
You must be signed in to change notification settings - Fork 80
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Add batched cholesky implementation and tests (#1029)
* add batched cholesky implementation and tests * missing files * fix correctness issues in transpose lower implementation * address PR comments * remove print statements * address more PR comments * test fixes * remove outdated comment * Add missing "throws exception" annotation --------- Co-authored-by: Manolis Papadakis <[email protected]>
- Loading branch information
Showing
17 changed files
with
692 additions
and
110 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
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
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 +1,85 @@ | ||
| /* Copyright 2023 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. | ||
| * | ||
| */ | ||
|
|
||
| #include "cunumeric/matrix/batched_cholesky.h" | ||
| #include "cunumeric/cunumeric.h" | ||
| #include "cunumeric/matrix/batched_cholesky_template.inl" | ||
|
|
||
| #include <cblas.h> | ||
| #include <core/type/type_info.h> | ||
| #include <lapack.h> | ||
|
|
||
| namespace cunumeric { | ||
|
|
||
| using namespace legate; | ||
|
|
||
| template <> | ||
| void CopyBlockImpl<VariantKind::CPU>::operator()(void* dst, const void* src, size_t size) | ||
| { | ||
| ::memcpy(dst, src, size); | ||
| } | ||
|
|
||
| template <Type::Code CODE> | ||
| struct BatchedTransposeImplBody<VariantKind::CPU, CODE> { | ||
| using VAL = legate_type_of<CODE>; | ||
|
|
||
| static constexpr int tile_size = 64; | ||
|
|
||
| void operator()(VAL* out, int n) const | ||
| { | ||
| VAL tile[tile_size][tile_size]; | ||
| int nblocks = (n + tile_size - 1) / tile_size; | ||
|
|
||
| for (int rb = 0; rb < nblocks; ++rb) { | ||
| for (int cb = 0; cb < nblocks; ++cb) { | ||
| int r_start = rb * tile_size; | ||
| int r_stop = std::min(r_start + tile_size, n); | ||
| int c_start = cb * tile_size; | ||
| int c_stop = std::min(c_start + tile_size, n); | ||
| for (int r = r_start, tr = 0; r < r_stop; ++r, ++tr) { | ||
| for (int c = c_start, tc = 0; c < c_stop; ++c, ++tc) { | ||
| if (r <= c) { | ||
| tile[tr][tc] = out[r * n + c]; | ||
| } else { | ||
| tile[tr][tc] = 0; | ||
| } | ||
| } | ||
| } | ||
| for (int r = c_start, tr = 0; r < c_stop; ++r, ++tr) { | ||
| for (int c = r_start, tc = 0; c < r_stop; ++c, ++tc) { out[r * n + c] = tile[tc][tr]; } | ||
| } | ||
| } | ||
| } | ||
| } | ||
| }; | ||
|
|
||
| /*static*/ void BatchedCholeskyTask::cpu_variant(TaskContext& context) | ||
| { | ||
| #ifdef LEGATE_USE_OPENMP | ||
| openblas_set_num_threads(1); // make sure this isn't overzealous | ||
| #endif | ||
| batched_cholesky_task_context_dispatch<VariantKind::CPU>(context); | ||
| } | ||
|
|
||
| namespace // unnamed | ||
| { | ||
| static void __attribute__((constructor)) register_tasks(void) | ||
| { | ||
| BatchedCholeskyTask::register_variants(); | ||
| } | ||
| } // namespace | ||
|
|
||
| } // namespace cunumeric |
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 +1,111 @@ | ||
| /* Copyright 2023 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. | ||
| * | ||
| */ | ||
|
|
||
| #include "cunumeric/matrix/batched_cholesky.h" | ||
| #include "cunumeric/matrix/potrf.h" | ||
| #include "cunumeric/matrix/batched_cholesky_template.inl" | ||
|
|
||
| #include "cunumeric/cuda_help.h" | ||
|
|
||
| namespace cunumeric { | ||
|
|
||
| using namespace legate; | ||
|
|
||
| #define TILE_DIM 32 | ||
| #define BLOCK_ROWS 8 | ||
|
|
||
| template <> | ||
| void CopyBlockImpl<VariantKind::GPU>::operator()(void* dst, const void* src, size_t size) | ||
| { | ||
| cudaMemcpyAsync(dst, src, size, cudaMemcpyDeviceToDevice, get_cached_stream()); | ||
| } | ||
|
|
||
| template <typename VAL> | ||
| __global__ static void __launch_bounds__((TILE_DIM * BLOCK_ROWS), MIN_CTAS_PER_SM) | ||
| transpose_2d_lower(VAL* out, int n) | ||
| { | ||
| __shared__ VAL tile[TILE_DIM][TILE_DIM + 1 /*avoid bank conflicts*/]; | ||
|
|
||
| // The y dim is fast-moving index for coalescing | ||
| auto r_block = blockIdx.x * TILE_DIM; | ||
| auto c_block = blockIdx.y * TILE_DIM; | ||
| auto r = blockIdx.x * TILE_DIM + threadIdx.x; | ||
| auto c = blockIdx.y * TILE_DIM + threadIdx.y; | ||
| auto stride = BLOCK_ROWS; | ||
| // The tile coordinates | ||
| auto tr = threadIdx.x; | ||
| auto tc = threadIdx.y; | ||
| auto offset = r * n + c; | ||
|
|
||
| // only execute across the upper diagonal | ||
| // a single thread block will store the upper diagonal block into | ||
| // a temp shared memory then set the block to zeros | ||
| if (c_block >= r_block) { | ||
| #pragma unroll | ||
| for (int i = 0; i < TILE_DIM; i += BLOCK_ROWS, offset += stride) { | ||
| if (r < n && (c + i) < n) { | ||
| if (r <= (c + i)) { | ||
| tile[tr][tc + i] = out[offset]; | ||
| // clear the upper diagonal entry | ||
| out[offset] = 0; | ||
| } else { | ||
| tile[tr][tc + i] = 0; | ||
| } | ||
| } | ||
| } | ||
|
|
||
| // Make sure all the data is in shared memory | ||
| __syncthreads(); | ||
|
|
||
| // Transpose the global coordinates, keep y the fast-moving index | ||
| r = blockIdx.y * TILE_DIM + threadIdx.x; | ||
| c = blockIdx.x * TILE_DIM + threadIdx.y; | ||
| offset = r * n + c; | ||
|
|
||
| #pragma unroll | ||
| for (int i = 0; i < TILE_DIM; i += BLOCK_ROWS, offset += stride) { | ||
| if (r < n && (c + i) < n) { | ||
| if (r >= (c + i)) { out[offset] = tile[tc + i][tr]; } | ||
| } | ||
| } | ||
| } | ||
| } | ||
|
|
||
| template <Type::Code CODE> | ||
| struct BatchedTransposeImplBody<VariantKind::GPU, CODE> { | ||
| using VAL = legate_type_of<CODE>; | ||
|
|
||
| void operator()(VAL* out, int n) const | ||
| { | ||
| const dim3 blocks((n + TILE_DIM - 1) / TILE_DIM, (n + TILE_DIM - 1) / TILE_DIM, 1); | ||
| const dim3 threads(TILE_DIM, BLOCK_ROWS, 1); | ||
|
|
||
| auto stream = get_cached_stream(); | ||
|
|
||
| // CUDA Potrf produces the full matrix, we only want | ||
| // the lower diagonal | ||
| transpose_2d_lower<VAL><<<blocks, threads, 0, stream>>>(out, n); | ||
|
|
||
| CHECK_CUDA_STREAM(stream); | ||
| } | ||
| }; | ||
|
|
||
| /*static*/ void BatchedCholeskyTask::gpu_variant(TaskContext& context) | ||
| { | ||
| batched_cholesky_task_context_dispatch<VariantKind::GPU>(context); | ||
| } | ||
|
|
||
| } // namespace cunumeric |
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 +1,38 @@ | ||
| /* Copyright 2021-2022 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 "cunumeric/cunumeric.h" | ||
| #include "cunumeric/cunumeric_c.h" | ||
|
|
||
| namespace cunumeric { | ||
|
|
||
| class BatchedCholeskyTask : public CuNumericTask<BatchedCholeskyTask> { | ||
| public: | ||
| static const int TASK_ID = CUNUMERIC_BATCHED_CHOLESKY; | ||
|
|
||
| public: | ||
| static void cpu_variant(legate::TaskContext& context); | ||
| #ifdef LEGATE_USE_OPENMP | ||
| static void omp_variant(legate::TaskContext& context); | ||
| #endif | ||
| #ifdef LEGATE_USE_CUDA | ||
| static void gpu_variant(legate::TaskContext& context); | ||
| #endif | ||
| }; | ||
|
|
||
| } // namespace cunumeric |
Oops, something went wrong.