Merge branch 'master' into tbb-arena

.ci/pipeline/ci.yml

@@ -122,7 +122,6 @@ jobs:
 # displayName: 'Uploading on fail'
 # condition: failed()
 # continueOnError: true
-
 # - job: 'LinuxMakeDPCPP'
 # timeoutInMinutes: 0
 # variables:
@@ -395,8 +394,8 @@ jobs:
  .ci/scripts/build.sh --compiler clang --target daal --optimizations "sse2 avx2" --conda-env ci-env
  displayName: 'make daal'
  - script: |
- .ci/scripts/build.sh --compiler clang --target oneapi_c --optimizations "sse2 avx2"
- displayName: 'make oneapi_c'
+ .ci/scripts/build.sh --compiler clang --target onedal_c --optimizations "sse2 avx2"
+ displayName: 'make onedal_c'
  - task: PublishPipelineArtifact@1
  inputs:
  artifactName: '$(platform.type) build'
@@ -471,4 +470,4 @@ jobs:
 # targetPath: '$(Build.Repository.LocalPath)/$(release.dir)'
 # displayName: 'Uploading on fail'
 # condition: failed()
-# continueOnError: true
+# continueOnError: true

cpp/daal/include/algorithms/algorithm_container_base_batch.h

@@ -139,7 +139,7 @@ class AlgorithmContainerImpl<batch> : public AlgorithmContainer<batch>
  *
  * \tparam mode Computation mode of the algorithm, \ref ComputeMode
  * \tparam sse2Container Implementation for Intel(R) Streaming SIMD Extensions 2 (Intel(R) SSE2)
- * \tparam sse42Container Implementation for Intel(R) Streaming SIMD Extensions 42 (Intel(R) SSE42)
+ * \tparam sse42Container Implementation for Intel(R) Streaming SIMD Extensions 4.2 (Intel(R) SSE4.2)
  * \tparam avx2Container Implementation for Intel(R) Advanced Vector Extensions 2 (Intel(R) AVX2)
  * \tparam avx512Container Implementation for Intel(R) Xeon(R) processors based on Intel AVX-512
  */

cpp/daal/include/algorithms/algorithm_container_base_common.h

@@ -49,7 +49,7 @@ namespace interface1
  *
  * \tparam mode Computation mode of the algorithm, \ref ComputeMode
  * \tparam sse2Container Implementation for Intel(R) Streaming SIMD Extensions 2 (Intel(R) SSE2)
- * \tparam sse42Container Implementation for Intel(R) Streaming SIMD Extensions 42 (Intel(R) SSE42)
+ * \tparam sse42Container Implementation for Intel(R) Streaming SIMD Extensions 4.2 (Intel(R) SSE4.2)
  * \tparam avx2Container Implementation for Intel(R) Advanced Vector Extensions 2 (Intel(R) AVX2)
  * \tparam avx512Container Implementation for Intel(R) Xeon(R) processors based on Intel AVX-512
  */

cpp/daal/include/services/daal_defines.h

@@ -130,7 +130,7 @@ namespace daal
 {
 /**
 * <a name="DAAL-ENUM-COMPUTEMODE"></a>
-* Computation modes of Intel(R) oneDAL algorithms
+* Computation modes of Intel(R) oneAPI Data Analytics Library (oneDAL) algorithms
 */
 enum ComputeMode
 {

cpp/daal/include/services/daal_string.h

@@ -17,7 +17,7 @@
 
 /*
 //++
-// Intel(R) oneDAL string class.
+// Intel(R) oneAPI Data Analytics Library (oneDAL) string class.
 //--
 */
 

cpp/daal/include/services/error_handling.h

@@ -17,7 +17,7 @@
 
 /*
 //++
-// Handling errors in Intel(R) oneDAL.
+// Handling errors in Intel(R) oneAPI Data Analytics Library (oneDAL).
 //--
 */
 

cpp/daal/include/services/error_id.h

@@ -17,7 +17,7 @@
 
 /*
 //++
-// Data types for error handling in Intel(R) oneDAL.
+// Data types for error handling in Intel(R) oneAPI Data Analytics Library (oneDAL).
 //--
 */
 

cpp/daal/include/services/error_indexes.h

@@ -17,7 +17,7 @@
 
 /*
 //++
-// Details of errors in Intel(R) oneDAL.
+// Details of errors in Intel(R) oneAPI Data Analytics Library (oneDAL).
 //--
 */
 

cpp/daal/include/services/internal/execution_context.h

@@ -58,7 +58,7 @@ class ExecutionContext : public Base
  // This branch is needed to avoid problems with deleting SYCL entities
  // after SYCL RT static objects are already released.
  // This is caused by "C++ static initialization order fiasco" problem between
- // oneDAL static Environment object and internal static contexts of SYCL RT.
+ // Intel(R) oneAPI Data Analytics Library (oneDAL) static Environment object and internal static contexts of SYCL RT.
  // Here we solve this temporary with a small memory leak.
  // TODO: remove this after complete transition to DPC++ kernels.
  if (needEmptyDeleter)

cpp/daal/include/services/internal/sycl/math/mkl_blas.h

@@ -46,7 +46,7 @@ namespace interface1
 
 /**
  * <a name="DAAL-CLASS-ONEAPI-INTERNAL__MKLGEMM"></a>
- * \brief Adapter for MKL GEMM routine
+ * \brief Adapter for Intel(R) MKL GEMM routine
  */
 template <typename algorithmFPType>
 struct MKLGemm
@@ -112,7 +112,7 @@ struct MKLGemm
 
 /**
  * <a name="DAAL-CLASS-ONEAPI-INTERNAL__MKLSYRK"></a>
- * \brief Adapter for MKL SYRK routine
+ * \brief Adapter for Intel(R) MKL SYRK routine
  */
 template <typename algorithmFPType>
 struct MKLSyrk
@@ -171,7 +171,7 @@ struct MKLSyrk
 
 /**
  * <a name="DAAL-CLASS-ONEAPI-INTERNAL__MKLAXPY"></a>
- * \brief Adapter for MKL AXPY routine
+ * \brief Adapter for Intel(R) MKL AXPY routine
  */
 template <typename algorithmFPType>
 struct MKLAxpy

cpp/daal/include/services/internal/sycl/math/mkl_dal_utils.h

@@ -17,7 +17,7 @@
 
 /*
 //++
-// Utility functions for DAL wrappers over MKL routines.
+// Utility functions for DAL wrappers over Intel(R) MKL routines.
 //--
 */
 

cpp/daal/include/services/internal/sycl/math/mkl_lapack.h

@@ -45,7 +45,7 @@ namespace interface1
 
 /**
  * <a name="DAAL-CLASS-ONEAPI-INTERNAL__MKLPOTRF"></a>
- * \brief Adapter for MKL POTRF routine
+ * \brief Adapter for Intel(R) MKL POTRF routine
  */
 template <typename algorithmFPType>
 struct MKLPotrf
@@ -98,7 +98,7 @@ struct MKLPotrf
 
 /**
  * <a name="DAAL-CLASS-ONEAPI-INTERNAL__MKLPOTRS></a>
- * \brief Adapter for MKL POTRS routine
+ * \brief Adapter for Intel(R) MKL POTRS routine
  */
 template <typename algorithmFPType>
 struct MKLPotrs

cpp/daal/include/services/library_version_info.h

@@ -17,7 +17,7 @@
 
 /*
 //++
-// Intel(R) oneDAL version information.
+// Intel(R) oneAPI Data Analytics Library (oneDAL) version information.
 //--
 */
 

cpp/daal/src/externals/core_threading_win_dll.cpp

@@ -837,7 +837,7 @@ DAAL_EXPORT void * _getThreadPinner(bool create_pinner, void (*read_topo)(int &,
  #define CALL_RET_FUNC_FROM_DLL_CPU_MIC(ret_type, fn_dpref, fn_cpu, fn_name, argdecl, argcall)
 #endif
 
-/* Used directly in Intel oneDAL */
+/* Used directly in Intel(R) oneAPI Data Analytics Library (oneDAL) */
 CALL_VOID_FUNC_FROM_DLL(fpk_blas_, dsyrk,
  (const char * uplo, const char * trans, const DAAL_INT * n, const DAAL_INT * k, const double * alpha, const double * a,
  const DAAL_INT * lda, const double * beta, double * c, const DAAL_INT * ldc),
@@ -1109,7 +1109,7 @@ typedef signed int Ipp32s;
 typedef float Ipp32f;
 typedef double Ipp64f;
 
-/* Used in Intel oneDAL via SS */
+/* Used in Intel(R) oneAPI Data Analytics Library (oneDAL) via SS */
 CALL_RET_FUNC_FROM_DLL(IppStatus, fpk_dft_, ippsSortRadixAscend_64f_I, (Ipp64f * pSrcDst, Ipp64f * pTmp, Ipp32s len), (pSrcDst, pTmp, len));
 CALL_RET_FUNC_FROM_DLL(IppStatus, fpk_dft_, ippsSortRadixAscend_32f_I, (Ipp32f * pSrcDst, Ipp32f * pTmp, Ipp32s len), (pSrcDst, pTmp, len));
 

cpp/oneapi/dal/algo/objective_function/backend/gpu/compute_kernel_dense_batch_impl_dpc.cpp

@@ -118,28 +118,25 @@ void add_regularization(sycl::queue& q_,
 template <typename Float>
 sycl::event value_and_gradient_iter(sycl::queue& q_,
  std::int64_t p,
- const pr::ndarray<Float, 1>& params_nd,
- const pr::ndarray<Float, 2>& data_nd,
- const pr::ndarray<std::int32_t, 1>& responses_nd,
- const pr::ndarray<Float, 1>& probabilities,
- pr::ndarray<Float, 1>& out,
- pr::ndarray<Float, 1>& ans,
+ const pr::ndview<Float, 2>& data_nd,
+ const pr::ndview<std::int32_t, 1>& responses_nd,
+ const pr::ndview<Float, 1>& probabilities,
+ pr::ndview<Float, 1>& out,
+ pr::ndview<Float, 1>& ans,
  bool fit_intercept,
  sycl::event& prev_iter) {
  auto fill_event = fill(q_, out, Float(0), {});
 
- auto out_loss = out.slice(0, 1);
- auto out_gradient = out.slice(1, p + 1);
+ auto out_loss = out.get_slice(0, 1);
+ auto out_gradient = out.get_slice(1, p + 2);
+ auto out_gradient_suf = fit_intercept ? out_gradient : out_gradient.get_slice(1, p + 1);
 
  auto loss_event = compute_logloss_with_der(q_,
- params_nd,
  data_nd,
  responses_nd,
  probabilities,
  out_loss,
- out_gradient,
- Float(0),
- Float(0),
+ out_gradient_suf,
  fit_intercept,
  { fill_event });
 
@@ -156,26 +153,15 @@ sycl::event value_and_gradient_iter(sycl::queue& q_,
 
 template <typename Float>
 sycl::event value_iter(sycl::queue& q_,
- std::int64_t p,
- const pr::ndarray<Float, 1>& params_nd,
- const pr::ndarray<Float, 2>& data_nd,
- const pr::ndarray<std::int32_t, 1>& responses_nd,
- const pr::ndarray<Float, 1>& probabilities,
- pr::ndarray<Float, 1>& out_loss,
- pr::ndarray<Float, 1>& ans_loss,
+ const pr::ndview<std::int32_t, 1>& responses_nd,
+ const pr::ndview<Float, 1>& probabilities,
+ pr::ndview<Float, 1>& out_loss,
+ pr::ndview<Float, 1>& ans_loss,
  bool fit_intercept,
  sycl::event& prev_iter) {
  auto fill_event = fill(q_, out_loss, Float(0), {});
- auto loss_event = compute_logloss(q_,
- params_nd,
- data_nd,
- responses_nd,
- probabilities,
- out_loss,
- Float(0),
- Float(0),
- fit_intercept,
- { fill_event });
+ auto loss_event =
+ compute_logloss(q_, responses_nd, probabilities, out_loss, fit_intercept, { fill_event });
  const auto* const out_ptr = out_loss.get_data();
  auto* const ans_loss_ptr = ans_loss.get_mutable_data();
  return q_.submit([&](sycl::handler& cgh) {
@@ -189,7 +175,6 @@ sycl::event value_iter(sycl::queue& q_,
 template <typename Float>
 sycl::event gradient_iter(sycl::queue& q_,
  std::int64_t p,
- const pr::ndarray<Float, 1>& params_nd,
  const pr::ndarray<Float, 2>& data_nd,
  const pr::ndarray<std::int32_t, 1>& responses_nd,
  const pr::ndarray<Float, 1>& probabilities,
@@ -198,14 +183,12 @@ sycl::event gradient_iter(sycl::queue& q_,
  bool fit_intercept,
  sycl::event& prev_iter) {
  auto fill_event = fill(q_, out_gradient, Float(0), {});
+ auto out_grad_suf = fit_intercept ? out_gradient : out_gradient.get_slice(1, p + 1);
  auto grad_event = compute_derivative(q_,
- params_nd,
  data_nd,
  responses_nd,
  probabilities,
- out_gradient,
- Float(0),
- Float(0),
+ out_grad_suf,
  fit_intercept,
  { fill_event });
  grad_event.wait_and_throw();
@@ -225,7 +208,6 @@ sycl::event gradient_iter(sycl::queue& q_,
 template <typename Float>
 sycl::event hessian_iter(sycl::queue& q_,
  std::int64_t p,
- const pr::ndarray<Float, 1>& params_nd,
  const pr::ndarray<Float, 2>& data_nd,
  const pr::ndarray<std::int32_t, 1>& responses_nd,
  const pr::ndarray<Float, 1>& probabilities,
@@ -235,7 +217,6 @@ sycl::event hessian_iter(sycl::queue& q_,
  sycl::event& prev_iter) {
  auto fill_event = fill(q_, out_hessian, Float(0), {});
  auto hess_event = compute_hessian(q_,
- params_nd,
  data_nd,
  responses_nd,
  probabilities,
@@ -282,6 +263,7 @@ result_t compute_kernel_dense_batch_impl<Float>::operator()(
  const bk::uniform_blocking blocking(n, bsz);
 
  const auto params_nd = pr::table2ndarray_1d<Float>(q_, params, alloc::device);
+ const auto params_nd_suf = fit_intercept ? params_nd : params_nd.slice(1, p);
  const auto* const params_ptr = params_nd.get_data();
 
  const auto responses_nd_big = pr::table2ndarray_1d<std::int32_t>(q_, responses, alloc::device);
@@ -326,14 +308,13 @@ result_t compute_kernel_dense_batch_impl<Float>::operator()(
  const auto responses_nd = responses_nd_big.slice(first, cursize);
 
  sycl::event prob_e =
- compute_probabilities(q_, params_nd, data_nd, probabilities, fit_intercept, {});
+ compute_probabilities(q_, params_nd_suf, data_nd, probabilities, fit_intercept, {});
  prob_e.wait_and_throw();
 
  if (desc.get_result_options().test(result_options::value) &&
  desc.get_result_options().test(result_options::gradient)) {
  prev_logloss_e = value_and_gradient_iter(q_,
  p,
- params_nd,
  data_nd,
  responses_nd,
  probabilities,
@@ -345,9 +326,6 @@ result_t compute_kernel_dense_batch_impl<Float>::operator()(
  else {
  if (desc.get_result_options().test(result_options::value)) {
  prev_logloss_e = value_iter(q_,
- p,
- params_nd,
- data_nd,
  responses_nd,
  probabilities,
  out_loss,
@@ -358,7 +336,6 @@ result_t compute_kernel_dense_batch_impl<Float>::operator()(
  if (desc.get_result_options().test(result_options::gradient)) {
  prev_grad_e = gradient_iter(q_,
  p,
- params_nd,
  data_nd,
  responses_nd,
  probabilities,
@@ -371,7 +348,6 @@ result_t compute_kernel_dense_batch_impl<Float>::operator()(
  if (desc.get_result_options().test(result_options::hessian)) {
  prev_hess_e = hessian_iter(q_,
  p,
- params_nd,
  data_nd,
  responses_nd,
  probabilities,

cpp/oneapi/dal/backend/primitives/objective_function/BUILD

@@ -10,6 +10,7 @@ dal_module(
  dal_deps = [
  "@onedal//cpp/oneapi/dal/backend/primitives:common",
  "@onedal//cpp/oneapi/dal/backend/primitives:blas",
+ "@onedal//cpp/oneapi/dal/backend/primitives/optimizers",
  ],
 )