feat(custom_op): use dynamic process number (#42)

Co-authored-by: Xuehai Pan <XuehaiPan@pku.edu.cn>
metaopt · Jul 26, 2022 · ab2ebff · ab2ebff
1 parent 809f9c7
commit ab2ebff
Show file tree

Hide file tree

Showing 2 changed files with 11 additions and 7 deletions.
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -9,6 +9,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+### Added
+
+- Use dynamic process number in CPU kernels by [@JieRen98](https://github.com/JieRen98) in [#42](https://github.com/metaopt/TorchOpt/pull/42).
+
 ------
 
 ## [0.4.2] - 2022-07-26

diff --git a/src/adam_op/adam_op_impl.cpp b/src/adam_op/adam_op_impl.cpp
@@ -31,7 +31,7 @@ void adamForwardInplaceCPUKernel(
     const other_t inv_one_minus_pow_b2, const other_t eps,
     const other_t eps_root, const size_t n, scalar_t* __restrict__ updates_ptr,
     scalar_t* __restrict__ mu_ptr, scalar_t* __restrict__ nu_ptr) {
-#pragma omp parallel for num_threads(32)
+#pragma omp parallel for num_threads(omp_get_num_procs())
   for (size_t tid = 0; tid < n; ++tid) {
     const scalar_t updates = updates_ptr[tid];
     const scalar_t mu = mu_ptr[tid];
@@ -76,7 +76,7 @@ void adamForwardMuCPUKernel(const scalar_t* __restrict__ updates_ptr,
                             const scalar_t* __restrict__ mu_ptr,
                             const other_t b1, const size_t n,
                             scalar_t* __restrict__ mu_out_ptr) {
-#pragma omp parallel for num_threads(32)
+#pragma omp parallel for num_threads(omp_get_num_procs())
   for (size_t tid = 0; tid < n; ++tid) {
     const scalar_t updates = updates_ptr[tid];
     const scalar_t mu = mu_ptr[tid];
@@ -108,7 +108,7 @@ void adamForwardNuCPUKernel(const scalar_t* __restrict__ updates_ptr,
                             const scalar_t* __restrict__ nu_ptr,
                             const other_t b2, const size_t n,
                             scalar_t* __restrict__ nu_out_ptr) {
-#pragma omp parallel for num_threads(32)
+#pragma omp parallel for num_threads(omp_get_num_procs())
   for (size_t tid = 0; tid < n; ++tid) {
     const scalar_t updates = updates_ptr[tid];
     const scalar_t nu = nu_ptr[tid];
@@ -144,7 +144,7 @@ void adamForwardUpdatesCPUKernel(const scalar_t* __restrict__ new_mu_ptr,
                                  const other_t eps, const other_t eps_root,
                                  const size_t n,
                                  scalar_t* __restrict__ updates_out_ptr) {
-#pragma omp parallel for num_threads(32)
+#pragma omp parallel for num_threads(omp_get_num_procs())
   for (size_t tid = 0; tid < n; ++tid) {
     const scalar_t new_mu = new_mu_ptr[tid];
     const scalar_t new_nu = new_nu_ptr[tid];
@@ -185,7 +185,7 @@ void adamBackwardMuCPUKernel(const scalar_t* __restrict__ dmu_ptr,
                              const other_t b1, const size_t n,
                              scalar_t* __restrict__ dupdates_out_ptr,
                              scalar_t* __restrict__ dmu_out_ptr) {
-#pragma omp parallel for num_threads(32)
+#pragma omp parallel for num_threads(omp_get_num_procs())
   for (size_t tid = 0; tid < n; ++tid) {
     const scalar_t dmu = dmu_ptr[tid];
 
@@ -220,7 +220,7 @@ void adamBackwardNuCPUKernel(const scalar_t* __restrict__ dnu_ptr,
                              const other_t b2, const size_t n,
                              scalar_t* __restrict__ dupdates_out_ptr,
                              scalar_t* __restrict__ dnu_out_ptr) {
-#pragma omp parallel for num_threads(32)
+#pragma omp parallel for num_threads(omp_get_num_procs())
   for (size_t tid = 0; tid < n; ++tid) {
     const scalar_t dnu = dnu_ptr[tid];
     const scalar_t updates = updates_ptr[tid];
@@ -259,7 +259,7 @@ void adamBackwardUpdatesCPUKernel(const scalar_t* __restrict__ dupdates_ptr,
                                   const size_t n,
                                   scalar_t* __restrict__ dnew_mu_out_ptr,
                                   scalar_t* __restrict__ dnew_nu_out_ptr) {
-#pragma omp parallel for num_threads(32)
+#pragma omp parallel for num_threads(omp_get_num_procs())
   for (size_t tid = 0; tid < n; ++tid) {
     const scalar_t dupdates = dupdates_ptr[tid];
     const scalar_t updates = updates_ptr[tid];