mkl-devel-dpcpp 2024.1.0__py2.py3-none-manylinux1_x86_64.whl → 2024.2.1__py2.py3-none-manylinux1_x86_64.whl

{mkl_devel_dpcpp-2024.1.0.data → mkl_devel_dpcpp-2024.2.1.data}/data/include/oneapi/mkl/lapack/lapack.hpp

@@ -1,5 +1,5 @@
 /*******************************************************************************
-* Copyright 2023 Intel Corporation.
+* Copyright 2024 Intel Corporation.
 *
 * This software and the related documents are Intel copyrighted  materials,  and
 * your use of  them is  governed by the  express license  under which  they were
@@ -11,21 +11,15 @@
 * or implied  warranties,  other  than those  that are  expressly stated  in the
 * License.
 *******************************************************************************/
+#ifndef _ONEAPI_MKL_LAPACK_LAPACK_HPP__
+#define _ONEAPI_MKL_LAPACK_LAPACK_HPP__
 
-#ifndef _LAPACK_LAPACK_HPP__
-#define _LAPACK_LAPACK_HPP__
-
-//
-// DPC++ MKL LAPACK API
-//
-
-#include <cstdint>
 #include <complex>
+#include <cstdint>
 #include <sycl/sycl.hpp>
 
-#include "oneapi/mkl/export.hpp"
-#include "mkl_types.h"
 #include "oneapi/mkl/types.hpp"
+#include "oneapi/mkl/export.hpp"
 
 namespace oneapi {
 namespace mkl {
@@ -40,8 +34,6 @@ DLL_EXPORT void potrf(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n
 DLL_EXPORT void potrf(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, sycl::buffer<double> &a, std::int64_t lda, sycl::buffer<double> &scratchpad, std::int64_t scratchpad_size);
 DLL_EXPORT void potrf(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, sycl::buffer<std::complex<float>>  &a, std::int64_t lda, sycl::buffer<std::complex<float>>  &scratchpad, std::int64_t scratchpad_size);
 DLL_EXPORT void potrf(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, sycl::buffer<std::complex<double>> &a, std::int64_t lda, sycl::buffer<std::complex<double>> &scratchpad, std::int64_t scratchpad_size);
-template <typename data_t, oneapi::mkl::lapack::internal::is_floating_point<data_t> = nullptr>
-std::int64_t potrf_scratchpad_size(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, std::int64_t lda);
 
 // potrs
 DLL_EXPORT sycl::event potrs(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, std::int64_t nrhs, float  *a, std::int64_t lda, float  *b, std::int64_t ldb, float  *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
@@ -52,8 +44,6 @@ DLL_EXPORT void potrs(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n
 DLL_EXPORT void potrs(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, std::int64_t nrhs, sycl::buffer<double> &a, std::int64_t lda, sycl::buffer<double> &b, std::int64_t ldb, sycl::buffer<double> &scratchpad, std::int64_t scratchpad_size);
 DLL_EXPORT void potrs(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, std::int64_t nrhs, sycl::buffer<std::complex<float>>  &a, std::int64_t lda, sycl::buffer<std::complex<float>>  &b, std::int64_t ldb, sycl::buffer<std::complex<float>>  &scratchpad, std::int64_t scratchpad_size);
 DLL_EXPORT void potrs(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, std::int64_t nrhs, sycl::buffer<std::complex<double>> &a, std::int64_t lda, sycl::buffer<std::complex<double>> &b, std::int64_t ldb, sycl::buffer<std::complex<double>> &scratchpad, std::int64_t scratchpad_size);
-template <typename data_t, oneapi::mkl::lapack::internal::is_floating_point<data_t> = nullptr>
-std::int64_t potrs_scratchpad_size(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, std::int64_t nrhs, std::int64_t lda, std::int64_t ldb);
 
 // potri
 DLL_EXPORT sycl::event potri(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, float  *a, std::int64_t lda, float  *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
@@ -64,28 +54,18 @@ DLL_EXPORT void potri(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n
 DLL_EXPORT void potri(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, sycl::buffer<double> &a, std::int64_t lda, sycl::buffer<double> &scratchpad, std::int64_t scratchpad_size);
 DLL_EXPORT void potri(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, sycl::buffer<std::complex<float>>  &a, std::int64_t lda, sycl::buffer<std::complex<float>>  &scratchpad, std::int64_t scratchpad_size);
 DLL_EXPORT void potri(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, sycl::buffer<std::complex<double>> &a, std::int64_t lda, sycl::buffer<std::complex<double>> &scratchpad, std::int64_t scratchpad_size);
-template <typename data_t, oneapi::mkl::lapack::internal::is_floating_point<data_t> = nullptr>
-std::int64_t potri_scratchpad_size(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, std::int64_t lda);
 
 // trtri
 DLL_EXPORT sycl::event trtri(sycl::queue &queue, oneapi::mkl::uplo uplo, oneapi::mkl::diag diag, std::int64_t n, float  *a, std::int64_t lda, float  *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
 DLL_EXPORT sycl::event trtri(sycl::queue &queue, oneapi::mkl::uplo uplo, oneapi::mkl::diag diag, std::int64_t n, double *a, std::int64_t lda, double *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
 DLL_EXPORT sycl::event trtri(sycl::queue &queue, oneapi::mkl::uplo uplo, oneapi::mkl::diag diag, std::int64_t n, std::complex<float>  *a, std::int64_t lda, std::complex<float>  *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
 DLL_EXPORT sycl::event trtri(sycl::queue &queue, oneapi::mkl::uplo uplo, oneapi::mkl::diag diag, std::int64_t n, std::complex<double> *a, std::int64_t lda, std::complex<double> *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
-template <typename data_t, oneapi::mkl::lapack::internal::is_floating_point<data_t> = nullptr>
-std::int64_t trtri_scratchpad_size(sycl::queue &queue, oneapi::mkl::uplo uplo, oneapi::mkl::diag diag, std::int64_t n, std::int64_t lda);
 
 // gesv
 DLL_EXPORT sycl::event gesv(sycl::queue &queue, std::int64_t n, std::int64_t nrhs, float *a, std::int64_t lda, std::int64_t *ipiv, float *b, std::int64_t ldb, float *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
 DLL_EXPORT sycl::event gesv(sycl::queue &queue, std::int64_t n, std::int64_t nrhs, double *a, std::int64_t lda, std::int64_t *ipiv, double *b, std::int64_t ldb, double *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
 DLL_EXPORT sycl::event gesv(sycl::queue &queue, std::int64_t n, std::int64_t nrhs, std::complex<float> *a, std::int64_t lda, std::int64_t *ipiv, std::complex<float> *b, std::int64_t ldb, std::complex<float> *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
 DLL_EXPORT sycl::event gesv(sycl::queue &queue, std::int64_t n, std::int64_t nrhs, std::complex<double> *a, std::int64_t lda, std::int64_t *ipiv, std::complex<double> *b, std::int64_t ldb, std::complex<double> *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
-template <typename data_t, oneapi::mkl::lapack::internal::is_floating_point<data_t> = nullptr>
-std::int64_t gesv_scratchpad_size(sycl::queue &queue, std::int64_t n, std::int64_t nrhs, std::int64_t lda, std::int64_t ldb);
-
-// gebrd: scratchpad size query
-template <typename fp_type, internal::is_floating_point<fp_type> = nullptr>
-int64_t gebrd_scratchpad_size(sycl::queue &queue, int64_t m, int64_t n, int64_t lda);
 
 // gebrd: USM API
 DLL_EXPORT sycl::event gebrd(sycl::queue &queue, int64_t m, int64_t n, std::complex<float> *a, int64_t lda, float *d,
@@ -119,10 +99,6 @@ DLL_EXPORT void gebrd(sycl::queue &queue, int64_t m, int64_t n, sycl::buffer<std
                       sycl::buffer<std::complex<double>> &taup, sycl::buffer<std::complex<double>> &scratchpad,
                       int64_t scratchpad_size);
 
-// geqrf: scratchpad size query
-template <typename fp_type, internal::is_floating_point<fp_type> = nullptr>
-int64_t geqrf_scratchpad_size(sycl::queue &queue, int64_t m, int64_t n, int64_t lda);
-
 // geqrf: USM API
 DLL_EXPORT sycl::event geqrf(sycl::queue &queue, int64_t m, int64_t n, std::complex<float> *a, int64_t lda,
                                  std::complex<float> *tau, std::complex<float> *scratchpad, int64_t scratchpad_size,
@@ -149,11 +125,6 @@ DLL_EXPORT void geqrf(sycl::queue &queue, int64_t m, int64_t n, sycl::buffer<std
                       int64_t lda, sycl::buffer<std::complex<double>> &tau,
                       sycl::buffer<std::complex<double>> &scratchpad, int64_t scratchpad_size);
 
-// gesvd_cmplx: scratchpad size query
-template <typename fp_type, internal::is_complex_floating_point<fp_type> = nullptr>
-int64_t gesvd_scratchpad_size(sycl::queue &queue, oneapi::mkl::jobsvd jobu, oneapi::mkl::jobsvd jobvt, int64_t m,
-                              int64_t n, int64_t lda, int64_t ldu, int64_t ldvt);
-
 // gesvd_cmplx: USM API
 DLL_EXPORT sycl::event gesvd(sycl::queue &queue, oneapi::mkl::jobsvd jobu, oneapi::mkl::jobsvd jobvt, int64_t m,
                              int64_t n, std::complex<float> *a, int64_t lda, float *s, std::complex<float> *u,
@@ -174,11 +145,6 @@ DLL_EXPORT void gesvd(sycl::queue &queue, oneapi::mkl::jobsvd jobu, oneapi::mkl:
                       sycl::buffer<std::complex<double>> &u, int64_t ldu, sycl::buffer<std::complex<double>> &vt,
                       int64_t ldvt, sycl::buffer<std::complex<double>> &scratchpad, int64_t scratchpad_size);
 
-// gesvd_real: scratchpad size query
-template <typename fp_type, internal::is_real_floating_point<fp_type> = nullptr>
-int64_t gesvd_scratchpad_size(sycl::queue &queue, oneapi::mkl::jobsvd jobu, oneapi::mkl::jobsvd jobvt, int64_t m,
-                              int64_t n, int64_t lda, int64_t ldu, int64_t ldvt);
-
 // gesvd_real: USM API
 DLL_EXPORT sycl::event gesvd(sycl::queue &queue, oneapi::mkl::jobsvd jobu, oneapi::mkl::jobsvd jobvt, int64_t m,
                              int64_t n, double *a, int64_t lda, double *s, double *u, int64_t ldu, double *vt,
@@ -198,12 +164,7 @@ DLL_EXPORT void gesvd(sycl::queue &queue, oneapi::mkl::jobsvd jobu, oneapi::mkl:
                       sycl::buffer<float> &a, int64_t lda, sycl::buffer<float> &s, sycl::buffer<float> &u, int64_t ldu,
                       sycl::buffer<float> &vt, int64_t ldvt, sycl::buffer<float> &scratchpad, int64_t scratchpad_size);
 
-// gesvda_batch_strided: scratchpad size query
-template <typename fp_type, internal::is_floating_point<fp_type> = nullptr>
-int64_t gesvda_batch_scratchpad_size(sycl::queue &queue, int64_t m, int64_t n, int64_t lda, int64_t stride_a,
-                                     int64_t stride_s, int64_t ldu, int64_t stride_u, int64_t ldvt, int64_t stride_vt,
-                                     int64_t batch_size);
- 
+
 // gesvda_batch_strided: USM API
 DLL_EXPORT sycl::event gesvda_batch(sycl::queue &queue, int64_t *iparm, int64_t *irank, int64_t m, int64_t n,
                                     std::complex<float> *a, int64_t lda, int64_t stride_a, float *s, int64_t stride_s,
@@ -253,9 +214,6 @@ DLL_EXPORT void gesvda_batch(sycl::queue &queue, sycl::buffer<int64_t> &iparm, s
                              int64_t ldu, int64_t stride_u, sycl::buffer<std::complex<double>> &vt, int64_t ldvt,
                              int64_t stride_vt, double tolerance, sycl::buffer<double> &residual, int64_t batch_size,
                              sycl::buffer<std::complex<double>> &scratchpad, int64_t scratchpad_size);
-// getrf: scratchpad size query
-template <typename fp_type, internal::is_floating_point<fp_type> = nullptr>
-int64_t getrf_scratchpad_size(sycl::queue &queue, int64_t m, int64_t n, int64_t lda);
 
 // getrf: USM API
 DLL_EXPORT sycl::event getrf(sycl::queue &queue, int64_t m, int64_t n, std::complex<float> *a, int64_t lda,
@@ -283,11 +241,6 @@ DLL_EXPORT void getrf(sycl::queue &queue, int64_t m, int64_t n, sycl::buffer<std
                       sycl::buffer<int64_t> &ipiv, sycl::buffer<std::complex<double>> &scratchpad,
                       int64_t scratchpad_size);
 
-// getrf_batch_group: scratchpad size query
-template <typename fp_type, internal::is_floating_point<fp_type> = nullptr>
-int64_t getrf_batch_scratchpad_size(sycl::queue &queue, int64_t* m, int64_t* n, int64_t* lda, int64_t group_count,
-                                    int64_t* group_sizes);
-
 // getrf_batch_group: USM API
 DLL_EXPORT sycl::event getrf_batch(sycl::queue &queue, int64_t *m, int64_t *n, std::complex<float> **a, int64_t *lda,
                                    int64_t **ipiv, int64_t group_count, int64_t *group_sizes,
@@ -304,11 +257,6 @@ DLL_EXPORT sycl::event getrf_batch(sycl::queue &queue, int64_t *m, int64_t *n, s
                                    std::complex<double> *scratchpad, int64_t scratchpad_size,
                                    const std::vector<sycl::event> &event_list = {});
 
-// getrf_batch_strided: scratchpad size query
-template <typename fp_type, internal::is_floating_point<fp_type> = nullptr>
-int64_t getrf_batch_scratchpad_size(sycl::queue &queue, int64_t m, int64_t n, int64_t lda, int64_t stride_a,
-                                    int64_t stride_ipiv, int64_t batch_size);
-
 // getrf_batch_strided: USM API
 DLL_EXPORT sycl::event getrf_batch(sycl::queue &queue, int64_t m, int64_t n, std::complex<float> *a, int64_t lda,
                                    int64_t stride_a, int64_t *ipiv, int64_t stride_ipiv, int64_t batch_size,
@@ -340,10 +288,6 @@ DLL_EXPORT void getrf_batch(sycl::queue &queue, int64_t m, int64_t n, sycl::buff
                             int64_t batch_size, sycl::buffer<std::complex<double>> &scratchpad,
                             int64_t scratchpad_size);
 
-// getrfnp: scratchpad size query
-template <typename fp_type, internal::is_floating_point<fp_type> = nullptr>
-int64_t getrfnp_scratchpad_size(sycl::queue &queue, int64_t m, int64_t n, int64_t lda);
-
 // getrfnp: USM API
 DLL_EXPORT sycl::event getrfnp(sycl::queue &queue, int64_t m, int64_t n, std::complex<float> *a, int64_t lda,
                                std::complex<float> *scratchpad, int64_t scratchpad_size,
@@ -366,10 +310,6 @@ DLL_EXPORT void getrfnp(sycl::queue &queue, int64_t m, int64_t n, sycl::buffer<f
 DLL_EXPORT void getrfnp(sycl::queue &queue, int64_t m, int64_t n, sycl::buffer<std::complex<double>> &a, int64_t lda,
                         sycl::buffer<std::complex<double>> &scratchpad, int64_t scratchpad_size);
 
-// getrfnp_batch_group: scratchpad size query
-template <typename fp_type, internal::is_floating_point<fp_type> = nullptr>
-int64_t getrfnp_batch_scratchpad_size(sycl::queue &queue, int64_t* m, int64_t* n, int64_t* lda, int64_t group_count,
-                                      int64_t* group_sizes);
 
 // getrfnp_batch_group: USM API
 DLL_EXPORT sycl::event getrfnp_batch(sycl::queue &queue, int64_t *m, int64_t *n, std::complex<float> **a, int64_t *lda,
@@ -385,11 +325,6 @@ DLL_EXPORT sycl::event getrfnp_batch(sycl::queue &queue, int64_t *m, int64_t *n,
                                      int64_t group_count, int64_t *group_sizes, std::complex<double> *scratchpad,
                                      int64_t scratchpad_size, const std::vector<sycl::event> &event_list = {});
 
-// getrfnp_batch_strided: scratchpad size query
-template <typename fp_type, internal::is_floating_point<fp_type> = nullptr>
-int64_t getrfnp_batch_scratchpad_size(sycl::queue &queue, int64_t m, int64_t n, int64_t lda, int64_t stride_a,
-                                      int64_t batch_size);
-
 // getrfnp_batch_strided: USM API
 DLL_EXPORT sycl::event getrfnp_batch(sycl::queue &queue, int64_t m, int64_t n, std::complex<float> *a, int64_t lda,
                                      int64_t stride_a, int64_t batch_size, std::complex<float> *scratchpad,
@@ -418,10 +353,6 @@ DLL_EXPORT void getrfnp_batch(sycl::queue &queue, int64_t m, int64_t n, sycl::bu
                               int64_t lda, int64_t stride_a, int64_t batch_size,
                               sycl::buffer<std::complex<double>> &scratchpad, int64_t scratchpad_size);
 
-// getri: scratchpad size query
-template <typename fp_type, internal::is_floating_point<fp_type> = nullptr>
-int64_t getri_scratchpad_size(sycl::queue &queue, int64_t n, int64_t lda);
-
 // getri: USM API
 DLL_EXPORT sycl::event getri(sycl::queue &queue, int64_t n, std::complex<float> *a, int64_t lda, int64_t *ipiv,
                              std::complex<float> *scratchpad, int64_t scratchpad_size,
@@ -446,11 +377,6 @@ DLL_EXPORT void getri(sycl::queue &queue, int64_t n, sycl::buffer<std::complex<d
                       sycl::buffer<int64_t> &ipiv, sycl::buffer<std::complex<double>> &scratchpad,
                       int64_t scratchpad_size);
 
-// getri_oop_batch_strided: scratchpad size query
-template <typename fp_type, internal::is_floating_point<fp_type> = nullptr>
-int64_t getri_batch_scratchpad_size(sycl::queue &queue, int64_t n, int64_t lda, int64_t stride_a,
-                                    int64_t stride_ipiv, int64_t ldainv, int64_t stride_ainv, int64_t batch_size);
-
 // getri_oop_batch_strided: USM API
 DLL_EXPORT sycl::event getri_batch(sycl::queue &queue, int64_t n, std::complex<float> *a, int64_t lda,
                                        int64_t stride_a, int64_t *ipiv, int64_t stride_ipiv, std::complex<float> *ainv,
@@ -493,11 +419,6 @@ DLL_EXPORT void getri_batch(sycl::queue &queue, int64_t n, sycl::buffer<std::com
                             int64_t batch_size, sycl::buffer<std::complex<double>> &scratchpad,
                             int64_t scratchpad_size);
 
-// getrs: scratchpad size query
-template <typename fp_type, internal::is_floating_point<fp_type> = nullptr>
-int64_t getrs_scratchpad_size(sycl::queue &queue, oneapi::mkl::transpose trans, int64_t n, int64_t nrhs, int64_t lda,
-                              int64_t ldb);
-
 // getrs: USM API
 DLL_EXPORT sycl::event getrs(sycl::queue &queue, oneapi::mkl::transpose trans, int64_t n, int64_t nrhs,
                              std::complex<float> *a, int64_t lda, int64_t *ipiv, std::complex<float> *b, int64_t ldb,
@@ -530,12 +451,6 @@ DLL_EXPORT void getrs(sycl::queue &queue, oneapi::mkl::transpose trans, int64_t
                       sycl::buffer<std::complex<double>> &b, int64_t ldb,
                       sycl::buffer<std::complex<double>> &scratchpad, int64_t scratchpad_size);
 
-// getrs_batch_strided: scratchpad size query
-template <typename fp_type, internal::is_floating_point<fp_type> = nullptr>
-int64_t getrs_batch_scratchpad_size(sycl::queue &queue, oneapi::mkl::transpose trans, int64_t n, int64_t nrhs,
-                                    int64_t lda, int64_t stride_a, int64_t stride_ipiv, int64_t ldb, int64_t stride_b,
-                                    int64_t batch_size);
-
 // getrs_batch_strided: USM API
 DLL_EXPORT sycl::event getrs_batch(sycl::queue &queue, oneapi::mkl::transpose trans, int64_t n, int64_t nrhs,
                                        std::complex<float> *a, int64_t lda, int64_t stride_a, int64_t *ipiv,
@@ -579,11 +494,6 @@ DLL_EXPORT void getrs_batch(sycl::queue &queue, oneapi::mkl::transpose trans, in
                             int64_t batch_size, sycl::buffer<std::complex<double>> &scratchpad,
                             int64_t scratchpad_size);
 
-// getrsnp_batch_strided: scratchpad size query
-template <typename fp_type, internal::is_floating_point<fp_type> = nullptr>
-int64_t getrsnp_batch_scratchpad_size(sycl::queue &queue, oneapi::mkl::transpose trans, int64_t n, int64_t nrhs,
-                                      int64_t lda, int64_t stride_a, int64_t ldb, int64_t stride_b, int64_t batch_size);
-
 // getrsnp_batch_strided: USM API
 DLL_EXPORT sycl::event getrsnp_batch(sycl::queue &queue, oneapi::mkl::transpose trans, int64_t n, int64_t nrhs,
                                      std::complex<float> *a, int64_t lda, int64_t stride_a, std::complex<float> *b,
@@ -621,11 +531,6 @@ DLL_EXPORT void getrsnp_batch(sycl::queue &queue, oneapi::mkl::transpose trans,
                               sycl::buffer<std::complex<double>> &b, int64_t ldb, int64_t stride_b, int64_t batch_size,
                               sycl::buffer<std::complex<double>> &scratchpad, int64_t scratchpad_size);
 
-// heev: scratchpad size query
-template <typename fp_type, internal::is_complex_floating_point<fp_type> = nullptr>
-int64_t heev_scratchpad_size(sycl::queue &queue, oneapi::mkl::compz jobz, oneapi::mkl::uplo uplo, int64_t n,
-                             int64_t lda);
-
 // heev: USM API
 DLL_EXPORT sycl::event heev(sycl::queue &queue, oneapi::mkl::compz jobz, oneapi::mkl::uplo uplo, int64_t n,
                             std::complex<float> *a, int64_t lda, float *w, std::complex<float> *scratchpad,
@@ -642,11 +547,6 @@ DLL_EXPORT void heev(sycl::queue &queue, oneapi::mkl::compz jobz, oneapi::mkl::u
                      sycl::buffer<std::complex<double>> &a, int64_t lda, sycl::buffer<double> &w,
                      sycl::buffer<std::complex<double>> &scratchpad, int64_t scratchpad_size);
 
-// heevd: scratchpad size query
-template <typename fp_type, internal::is_complex_floating_point<fp_type> = nullptr>
-int64_t heevd_scratchpad_size(sycl::queue &queue, oneapi::mkl::job jobz, oneapi::mkl::uplo uplo, int64_t n,
-                              int64_t lda);
-
 // heevd: USM API
 DLL_EXPORT sycl::event heevd(sycl::queue &queue, oneapi::mkl::job jobz, oneapi::mkl::uplo uplo, int64_t n,
                              std::complex<float> *a, int64_t lda, float *w, std::complex<float> *scratchpad,
@@ -663,13 +563,6 @@ DLL_EXPORT void heevd(sycl::queue &queue, oneapi::mkl::job jobz, oneapi::mkl::up
                       sycl::buffer<std::complex<double>> &a, int64_t lda, sycl::buffer<double> &w,
                       sycl::buffer<std::complex<double>> &scratchpad, int64_t scratchpad_size);
 
-// heevx: scratchpad size query
-template <typename fp_type, internal::is_complex_floating_point<fp_type> = nullptr>
-int64_t heevx_scratchpad_size(sycl::queue &queue, oneapi::mkl::compz jobz, oneapi::mkl::rangev range,
-                              oneapi::mkl::uplo uplo, int64_t n, int64_t lda, typename fp_type::value_type vl,
-                              typename fp_type::value_type vu, int64_t il, int64_t iu,
-                              typename fp_type::value_type abstol, int64_t ldz);
-
 // heevx: USM API
 DLL_EXPORT sycl::event heevx(sycl::queue &queue, oneapi::mkl::compz jobz, oneapi::mkl::rangev range,
                              oneapi::mkl::uplo uplo, int64_t n, std::complex<float> *a, int64_t lda, float vl, float vu,
@@ -694,11 +587,6 @@ DLL_EXPORT void heevx(sycl::queue &queue, oneapi::mkl::compz jobz, oneapi::mkl::
                       sycl::buffer<std::complex<double>> &z, int64_t ldz,
                       sycl::buffer<std::complex<double>> &scratchpad, int64_t scratchpad_size);
 
-// hegvd: scratchpad size query
-template <typename fp_type, internal::is_complex_floating_point<fp_type> = nullptr>
-int64_t hegvd_scratchpad_size(sycl::queue &queue, int64_t itype, oneapi::mkl::job jobz, oneapi::mkl::uplo uplo,
-                              int64_t n, int64_t lda, int64_t ldb);
-
 // hegvd: USM API
 DLL_EXPORT sycl::event hegvd(sycl::queue &queue, int64_t itype, oneapi::mkl::job jobz, oneapi::mkl::uplo uplo,
                              int64_t n, std::complex<float> *a, int64_t lda, std::complex<float> *b, int64_t ldb,
@@ -719,13 +607,6 @@ DLL_EXPORT void hegvd(sycl::queue &queue, int64_t itype, oneapi::mkl::job jobz,
                       int64_t ldb, sycl::buffer<double> &w, sycl::buffer<std::complex<double>> &scratchpad,
                       int64_t scratchpad_size);
 
-// hegvx: scratchpad size query
-template <typename fp_type, internal::is_complex_floating_point<fp_type> = nullptr>
-int64_t hegvx_scratchpad_size(sycl::queue &queue, int64_t itype, oneapi::mkl::compz jobz, oneapi::mkl::rangev range,
-                              oneapi::mkl::uplo uplo, int64_t n, int64_t lda, int64_t ldb,
-                              typename fp_type::value_type vl, typename fp_type::value_type vu, int64_t il, int64_t iu,
-                              typename fp_type::value_type abstol, int64_t ldz);
-
 // hegvx: USM API
 DLL_EXPORT sycl::event hegvx(sycl::queue &queue, int64_t itype, oneapi::mkl::compz jobz, oneapi::mkl::rangev range,
                              oneapi::mkl::uplo uplo, int64_t n, std::complex<float> *a, int64_t lda,
@@ -754,10 +635,6 @@ DLL_EXPORT void hegvx(sycl::queue &queue, int64_t itype, oneapi::mkl::compz jobz
                       sycl::buffer<std::complex<double>> &z, int64_t ldz,
                       sycl::buffer<std::complex<double>> &scratchpad, int64_t scratchpad_size);
 
-// hetrd: scratchpad size query
-template <typename fp_type, internal::is_complex_floating_point<fp_type> = nullptr>
-int64_t hetrd_scratchpad_size(sycl::queue &queue, oneapi::mkl::uplo uplo, int64_t n, int64_t lda);
-
 // hetrd: USM API
 DLL_EXPORT sycl::event hetrd(sycl::queue &queue, oneapi::mkl::uplo uplo, int64_t n, std::complex<float> *a, int64_t lda,
                              float *d, float *e, std::complex<float> *tau, std::complex<float> *scratchpad,
@@ -782,20 +659,12 @@ DLL_EXPORT sycl::event hetrf(sycl::queue &queue, oneapi::mkl::uplo uplo, std::in
 DLL_EXPORT sycl::event hetrf(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, std::complex<double> *a, std::int64_t lda, std::int64_t *ipiv, std::complex<double> *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
 DLL_EXPORT void hetrf(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, sycl::buffer<std::complex<float>>  &a, std::int64_t lda, sycl::buffer<std::int64_t> &ipiv, sycl::buffer<std::complex<float>>  &scratchpad, std::int64_t scratchpad_size);
 DLL_EXPORT void hetrf(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, sycl::buffer<std::complex<double>> &a, std::int64_t lda, sycl::buffer<std::int64_t> &ipiv, sycl::buffer<std::complex<double>> &scratchpad, std::int64_t scratchpad_size);
-template <typename data_t, oneapi::mkl::lapack::internal::is_complex_floating_point<data_t> = nullptr>
-std::int64_t hetrf_scratchpad_size(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, std::int64_t lda);
 
 // orgbr
 DLL_EXPORT sycl::event orgbr(sycl::queue &queue, oneapi::mkl::generate vec, std::int64_t m, std::int64_t n, std::int64_t k, float  *a, std::int64_t lda, float  *tau, float  *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
 DLL_EXPORT sycl::event orgbr(sycl::queue &queue, oneapi::mkl::generate vec, std::int64_t m, std::int64_t n, std::int64_t k, double *a, std::int64_t lda, double *tau, double *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
 DLL_EXPORT void orgbr(sycl::queue &queue, oneapi::mkl::generate vec, std::int64_t m, std::int64_t n, std::int64_t k, sycl::buffer<float>  &a, std::int64_t lda, sycl::buffer<float>  &tau, sycl::buffer<float>  &scratchpad, std::int64_t scratchpad_size);
 DLL_EXPORT void orgbr(sycl::queue &queue, oneapi::mkl::generate vec, std::int64_t m, std::int64_t n, std::int64_t k, sycl::buffer<double> &a, std::int64_t lda, sycl::buffer<double> &tau, sycl::buffer<double> &scratchpad, std::int64_t scratchpad_size);
-template <typename data_t, oneapi::mkl::lapack::internal::is_real_floating_point<data_t> = nullptr>
-std::int64_t orgbr_scratchpad_size(sycl::queue &queue, oneapi::mkl::generate vect, std::int64_t m, std::int64_t n, std::int64_t k, std::int64_t lda);
-
-// orgqr: scratchpad size query
-template <typename fp_type, internal::is_real_floating_point<fp_type> = nullptr>
-int64_t orgqr_scratchpad_size(sycl::queue &queue, int64_t m, int64_t n, int64_t k, int64_t lda);
 
 // orgqr: USM API
 DLL_EXPORT sycl::event orgqr(sycl::queue &queue, int64_t m, int64_t n, int64_t k, double *a, int64_t lda, double *tau,
@@ -811,11 +680,6 @@ DLL_EXPORT void orgqr(sycl::queue &queue, int64_t m, int64_t n, int64_t k, sycl:
 DLL_EXPORT void orgqr(sycl::queue &queue, int64_t m, int64_t n, int64_t k, sycl::buffer<float> &a, int64_t lda,
                       sycl::buffer<float> &tau, sycl::buffer<float> &scratchpad, int64_t scratchpad_size);
 
-// ormqr: scratchpad size query
-template <typename fp_type, internal::is_real_floating_point<fp_type> = nullptr>
-int64_t ormqr_scratchpad_size(sycl::queue &queue, oneapi::mkl::side side, oneapi::mkl::transpose trans, int64_t m,
-                              int64_t n, int64_t k, int64_t lda, int64_t ldc);
-
 // ormqr: USM API
 DLL_EXPORT sycl::event ormqr(sycl::queue &queue, oneapi::mkl::side side, oneapi::mkl::transpose trans, int64_t m,
                              int64_t n, int64_t k, double *a, int64_t lda, double *tau, double *c, int64_t ldc,
@@ -834,10 +698,6 @@ DLL_EXPORT void ormqr(sycl::queue &queue, oneapi::mkl::side side, oneapi::mkl::t
                       int64_t k, sycl::buffer<float> &a, int64_t lda, sycl::buffer<float> &tau, sycl::buffer<float> &c,
                       int64_t ldc, sycl::buffer<float> &scratchpad, int64_t scratchpad_size);
 
-// steqr: scratchpad size query
-template <typename fp_type, internal::is_floating_point<fp_type> = nullptr>
-int64_t steqr_scratchpad_size(sycl::queue &queue, oneapi::mkl::compz compz, int64_t n, int64_t ldz);
-
 // steqr: USM API
 DLL_EXPORT sycl::event steqr(sycl::queue &queue, oneapi::mkl::compz compz, int64_t n, float *d, float *e,
                              std::complex<float> *z, int64_t ldz, std::complex<float> *scratchpad,
@@ -866,11 +726,6 @@ DLL_EXPORT void steqr(sycl::queue &queue, oneapi::mkl::compz compz, int64_t n, s
                       sycl::buffer<double> &e, sycl::buffer<std::complex<double>> &z, int64_t ldz,
                       sycl::buffer<std::complex<double>> &scratchpad, int64_t scratchpad_size);
 
-// syev: scratchpad size query
-template <typename fp_type, internal::is_real_floating_point<fp_type> = nullptr>
-int64_t syev_scratchpad_size(sycl::queue &queue, oneapi::mkl::compz jobz, oneapi::mkl::uplo uplo, int64_t n,
-                             int64_t lda);
-
 // syev: USM API
 DLL_EXPORT sycl::event syev(sycl::queue &queue, oneapi::mkl::compz jobz, oneapi::mkl::uplo uplo, int64_t n, double *a,
                             int64_t lda, double *w, double *scratchpad, int64_t scratchpad_size,
@@ -887,11 +742,6 @@ DLL_EXPORT void syev(sycl::queue &queue, oneapi::mkl::compz jobz, oneapi::mkl::u
                      sycl::buffer<float> &a, int64_t lda, sycl::buffer<float> &w, sycl::buffer<float> &scratchpad,
                      int64_t scratchpad_size);
 
-// syevd: scratchpad size query
-template <typename fp_type, internal::is_real_floating_point<fp_type> = nullptr>
-int64_t syevd_scratchpad_size(sycl::queue &queue, oneapi::mkl::job jobz, oneapi::mkl::uplo uplo, int64_t n,
-                              int64_t lda);
-
 // syevd: USM API
 DLL_EXPORT sycl::event syevd(sycl::queue &queue, oneapi::mkl::job jobz, oneapi::mkl::uplo uplo, int64_t n, double *a,
                              int64_t lda, double *w, double *scratchpad, int64_t scratchpad_size,
@@ -908,12 +758,6 @@ DLL_EXPORT void syevd(sycl::queue &queue, oneapi::mkl::job jobz, oneapi::mkl::up
                       sycl::buffer<float> &a, int64_t lda, sycl::buffer<float> &w, sycl::buffer<float> &scratchpad,
                       int64_t scratchpad_size);
 
-// syevx: scratchpad size query
-template <typename fp_type, internal::is_real_floating_point<fp_type> = nullptr>
-int64_t syevx_scratchpad_size(sycl::queue &queue, oneapi::mkl::compz jobz, oneapi::mkl::rangev range,
-                              oneapi::mkl::uplo uplo, int64_t n, int64_t lda, fp_type vl, fp_type vu, int64_t il,
-                              int64_t iu, fp_type abstol, int64_t ldz);
-
 // syevx: USM API
 DLL_EXPORT sycl::event syevx(sycl::queue &queue, oneapi::mkl::compz jobz, oneapi::mkl::rangev range,
                              oneapi::mkl::uplo uplo, int64_t n, double *a, int64_t lda, double vl, double vu,
@@ -935,11 +779,6 @@ DLL_EXPORT void syevx(sycl::queue &queue, oneapi::mkl::compz jobz, oneapi::mkl::
                       float abstol, sycl::buffer<int64_t> &m, sycl::buffer<float> &w, sycl::buffer<float> &z,
                       int64_t ldz, sycl::buffer<float> &scratchpad, int64_t scratchpad_size);
 
-// sygvd: scratchpad size query
-template <typename fp_type, internal::is_real_floating_point<fp_type> = nullptr>
-int64_t sygvd_scratchpad_size(sycl::queue &queue, int64_t itype, oneapi::mkl::job jobz, oneapi::mkl::uplo uplo,
-                              int64_t n, int64_t lda, int64_t ldb);
-
 // sygvd: USM API
 DLL_EXPORT sycl::event sygvd(sycl::queue &queue, int64_t itype, oneapi::mkl::job jobz, oneapi::mkl::uplo uplo,
                              int64_t n, double *a, int64_t lda, double *b, int64_t ldb, double *w, double *scratchpad,
@@ -956,12 +795,6 @@ DLL_EXPORT void sygvd(sycl::queue &queue, int64_t itype, oneapi::mkl::job jobz,
                       sycl::buffer<float> &a, int64_t lda, sycl::buffer<float> &b, int64_t ldb, sycl::buffer<float> &w,
                       sycl::buffer<float> &scratchpad, int64_t scratchpad_size);
 
-// sygvx: scratchpad size query
-template <typename fp_type, internal::is_real_floating_point<fp_type> = nullptr>
-int64_t sygvx_scratchpad_size(sycl::queue &queue, int64_t itype, oneapi::mkl::compz jobz, oneapi::mkl::rangev range,
-                              oneapi::mkl::uplo uplo, int64_t n, int64_t lda, int64_t ldb, fp_type vl, fp_type vu,
-                              int64_t il, int64_t iu, fp_type abstol, int64_t ldz);
-
 // sygvx: USM API
 DLL_EXPORT sycl::event sygvx(sycl::queue &queue, int64_t itype, oneapi::mkl::compz jobz, oneapi::mkl::rangev range,
                              oneapi::mkl::uplo uplo, int64_t n, double *a, int64_t lda, double *b, int64_t ldb,
@@ -986,10 +819,6 @@ DLL_EXPORT void sygvx(sycl::queue &queue, int64_t itype, oneapi::mkl::compz jobz
                       sycl::buffer<float> &w, sycl::buffer<float> &z, int64_t ldz, sycl::buffer<float> &scratchpad,
                       int64_t scratchpad_size);
 
-// sytrd: scratchpad size query
-template <typename fp_type, internal::is_real_floating_point<fp_type> = nullptr>
-int64_t sytrd_scratchpad_size(sycl::queue &queue, oneapi::mkl::uplo uplo, int64_t n, int64_t lda);
-
 // sytrd: USM API
 DLL_EXPORT sycl::event sytrd(sycl::queue &queue, oneapi::mkl::uplo uplo, int64_t n, double *a, int64_t lda, double *d,
                              double *e, double *tau, double *scratchpad, int64_t scratchpad_size,
@@ -1006,11 +835,6 @@ DLL_EXPORT void sytrd(sycl::queue &queue, oneapi::mkl::uplo uplo, int64_t n, syc
                       sycl::buffer<float> &d, sycl::buffer<float> &e, sycl::buffer<float> &tau,
                       sycl::buffer<float> &scratchpad, int64_t scratchpad_size);
 
-// trtrs: scratchpad size query
-template <typename fp_type, internal::is_floating_point<fp_type> = nullptr>
-int64_t trtrs_scratchpad_size(sycl::queue &queue, oneapi::mkl::uplo uplo, oneapi::mkl::transpose trans,
-                              oneapi::mkl::diag diag, int64_t n, int64_t nrhs, int64_t lda, int64_t ldb);
-
 // trtrs: USM API
 DLL_EXPORT sycl::event trtrs(sycl::queue &queue, oneapi::mkl::uplo uplo, oneapi::mkl::transpose trans,
                              oneapi::mkl::diag diag, int64_t n, int64_t nrhs, std::complex<float> *a, int64_t lda,
@@ -1050,12 +874,6 @@ DLL_EXPORT sycl::event ungbr(sycl::queue &queue, oneapi::mkl::generate vec, std:
 DLL_EXPORT sycl::event ungbr(sycl::queue &queue, oneapi::mkl::generate vec, std::int64_t m, std::int64_t n, std::int64_t k, std::complex<double> *a, std::int64_t lda, std::complex<double> *tau, std::complex<double> *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
 DLL_EXPORT void ungbr(sycl::queue &queue, oneapi::mkl::generate vec, std::int64_t m, std::int64_t n, std::int64_t k, sycl::buffer<std::complex<float>>  &a, std::int64_t lda, sycl::buffer<std::complex<float>>  &tau, sycl::buffer<std::complex<float>>  &scratchpad, std::int64_t scratchpad_size);
 DLL_EXPORT void ungbr(sycl::queue &queue, oneapi::mkl::generate vec, std::int64_t m, std::int64_t n, std::int64_t k, sycl::buffer<std::complex<double>> &a, std::int64_t lda, sycl::buffer<std::complex<double>> &tau, sycl::buffer<std::complex<double>> &scratchpad, std::int64_t scratchpad_size);
-template <typename data_t, oneapi::mkl::lapack::internal::is_complex_floating_point<data_t> = nullptr>
-std::int64_t ungbr_scratchpad_size(sycl::queue &queue, oneapi::mkl::generate vect, std::int64_t m, std::int64_t n, std::int64_t k, std::int64_t lda);
-
-// ungqr: scratchpad size query
-template <typename fp_type, internal::is_complex_floating_point<fp_type> = nullptr>
-int64_t ungqr_scratchpad_size(sycl::queue &queue, int64_t m, int64_t n, int64_t k, int64_t lda);
 
 // ungqr: USM API
 DLL_EXPORT sycl::event ungqr(sycl::queue &queue, int64_t m, int64_t n, int64_t k, std::complex<float> *a, int64_t lda,
@@ -1073,11 +891,6 @@ DLL_EXPORT void ungqr(sycl::queue &queue, int64_t m, int64_t n, int64_t k, sycl:
                       int64_t lda, sycl::buffer<std::complex<double>> &tau,
                       sycl::buffer<std::complex<double>> &scratchpad, int64_t scratchpad_size);
 
-// unmqr: scratchpad size query
-template <typename fp_type, internal::is_complex_floating_point<fp_type> = nullptr>
-int64_t unmqr_scratchpad_size(sycl::queue &queue, oneapi::mkl::side side, oneapi::mkl::transpose trans, int64_t m,
-                              int64_t n, int64_t k, int64_t lda, int64_t ldc);
-
 // unmqr: USM API
 DLL_EXPORT sycl::event unmqr(sycl::queue &queue, oneapi::mkl::side side, oneapi::mkl::transpose trans, int64_t m,
                              int64_t n, int64_t k, std::complex<float> *a, int64_t lda, std::complex<float> *tau,
@@ -1107,24 +920,18 @@ DLL_EXPORT void gerqf(sycl::queue &queue, std::int64_t m, std::int64_t n, sycl::
 DLL_EXPORT void gerqf(sycl::queue &queue, std::int64_t m, std::int64_t n, sycl::buffer<double> &a, std::int64_t lda, sycl::buffer<double> &tau, sycl::buffer<double> &scratchpad, std::int64_t scratchpad_size);
 DLL_EXPORT void gerqf(sycl::queue &queue, std::int64_t m, std::int64_t n, sycl::buffer<std::complex<float>>  &a, std::int64_t lda, sycl::buffer<std::complex<float>>  &tau, sycl::buffer<std::complex<float>>  &scratchpad, std::int64_t scratchpad_size);
 DLL_EXPORT void gerqf(sycl::queue &queue, std::int64_t m, std::int64_t n, sycl::buffer<std::complex<double>> &a, std::int64_t lda, sycl::buffer<std::complex<double>> &tau, sycl::buffer<std::complex<double>> &scratchpad, std::int64_t scratchpad_size);
-template <typename data_t, oneapi::mkl::lapack::internal::is_floating_point<data_t> = nullptr>
-std::int64_t gerqf_scratchpad_size(sycl::queue &queue, std::int64_t m, std::int64_t n, std::int64_t lda);
 
 // ormrq
 DLL_EXPORT sycl::event ormrq(sycl::queue &queue, oneapi::mkl::side side, oneapi::mkl::transpose trans, std::int64_t m, std::int64_t n, std::int64_t k, float  *a, std::int64_t lda, float  *tau, float  *c, std::int64_t ldc, float  *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
 DLL_EXPORT sycl::event ormrq(sycl::queue &queue, oneapi::mkl::side side, oneapi::mkl::transpose trans, std::int64_t m, std::int64_t n, std::int64_t k, double *a, std::int64_t lda, double *tau, double *c, std::int64_t ldc, double *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
 DLL_EXPORT void ormrq(sycl::queue &queue, oneapi::mkl::side side, oneapi::mkl::transpose trans, std::int64_t m, std::int64_t n, std::int64_t k, sycl::buffer<float>  &a, std::int64_t lda, sycl::buffer<float>  &tau, sycl::buffer<float>  &c, std::int64_t ldc, sycl::buffer<float>  &scratchpad, std::int64_t scratchpad_size);
 DLL_EXPORT void ormrq(sycl::queue &queue, oneapi::mkl::side side, oneapi::mkl::transpose trans, std::int64_t m, std::int64_t n, std::int64_t k, sycl::buffer<double> &a, std::int64_t lda, sycl::buffer<double> &tau, sycl::buffer<double> &c, std::int64_t ldc, sycl::buffer<double> &scratchpad, std::int64_t scratchpad_size);
-template <typename data_t, oneapi::mkl::lapack::internal::is_real_floating_point<data_t> = nullptr>
-std::int64_t ormrq_scratchpad_size(sycl::queue &queue, oneapi::mkl::side side, oneapi::mkl::transpose trans, std::int64_t m, std::int64_t n, std::int64_t k, std::int64_t lda, std::int64_t ldc);
 
 // unmrq
 DLL_EXPORT sycl::event unmrq(sycl::queue &queue, oneapi::mkl::side side, oneapi::mkl::transpose trans, std::int64_t m, std::int64_t n, std::int64_t k, std::complex<float>  *a, std::int64_t lda, std::complex<float>  *tau, std::complex<float>  *c, std::int64_t ldc, std::complex<float>  *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
 DLL_EXPORT sycl::event unmrq(sycl::queue &queue, oneapi::mkl::side side, oneapi::mkl::transpose trans, std::int64_t m, std::int64_t n, std::int64_t k, std::complex<double> *a, std::int64_t lda, std::complex<double> *tau, std::complex<double> *c, std::int64_t ldc, std::complex<double> *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
 DLL_EXPORT void unmrq(sycl::queue &queue, oneapi::mkl::side side, oneapi::mkl::transpose trans, std::int64_t m, std::int64_t n, std::int64_t k, sycl::buffer<std::complex<float>>  &a, std::int64_t lda, sycl::buffer<std::complex<float>>  &tau, sycl::buffer<std::complex<float>>  &c, std::int64_t ldc, sycl::buffer<std::complex<float>>  &scratchpad, std::int64_t scratchpad_size);
 DLL_EXPORT void unmrq(sycl::queue &queue, oneapi::mkl::side side, oneapi::mkl::transpose trans, std::int64_t m, std::int64_t n, std::int64_t k, sycl::buffer<std::complex<double>> &a, std::int64_t lda, sycl::buffer<std::complex<double>> &tau, sycl::buffer<std::complex<double>> &c, std::int64_t ldc, sycl::buffer<std::complex<double>> &scratchpad, std::int64_t scratchpad_size);
-template <typename data_t, oneapi::mkl::lapack::internal::is_complex_floating_point<data_t> = nullptr>
-std::int64_t unmrq_scratchpad_size(sycl::queue &queue, oneapi::mkl::side side, oneapi::mkl::transpose trans, std::int64_t m, std::int64_t n, std::int64_t k, std::int64_t lda, std::int64_t ldc);
 
 // sytrf
 DLL_EXPORT sycl::event sytrf(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, float  *a, std::int64_t lda, std::int64_t *ipiv, float  *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
@@ -1135,40 +942,36 @@ DLL_EXPORT void sytrf(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n
 DLL_EXPORT void sytrf(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, sycl::buffer<double> &a, std::int64_t lda, sycl::buffer<std::int64_t> &ipiv, sycl::buffer<double> &scratchpad, std::int64_t scratchpad_size);
 DLL_EXPORT void sytrf(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, sycl::buffer<std::complex<float>>  &a, std::int64_t lda, sycl::buffer<std::int64_t> &ipiv, sycl::buffer<std::complex<float>>  &scratchpad, std::int64_t scratchpad_size);
 DLL_EXPORT void sytrf(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, sycl::buffer<std::complex<double>> &a, std::int64_t lda, sycl::buffer<std::int64_t> &ipiv, sycl::buffer<std::complex<double>> &scratchpad, std::int64_t scratchpad_size);
-template <typename data_t, oneapi::mkl::lapack::internal::is_floating_point<data_t> = nullptr>
-std::int64_t sytrf_scratchpad_size(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, std::int64_t lda);
 
 // orgtr
 DLL_EXPORT sycl::event orgtr(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, float  *a, std::int64_t lda, float  *tau, float  *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
 DLL_EXPORT sycl::event orgtr(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, double *a, std::int64_t lda, double *tau, double *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
 DLL_EXPORT void orgtr(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, sycl::buffer<float>  &a, std::int64_t lda, sycl::buffer<float>  &tau, sycl::buffer<float>  &scratchpad, std::int64_t scratchpad_size);
 DLL_EXPORT void orgtr(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, sycl::buffer<double> &a, std::int64_t lda, sycl::buffer<double> &tau, sycl::buffer<double> &scratchpad, std::int64_t scratchpad_size);
-template <typename data_t, oneapi::mkl::lapack::internal::is_real_floating_point<data_t> = nullptr>
-std::int64_t orgtr_scratchpad_size(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, std::int64_t lda);
 
 // ungtr
 DLL_EXPORT sycl::event ungtr(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, std::complex<float>  *a, std::int64_t lda, std::complex<float>  *tau, std::complex<float>  *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
 DLL_EXPORT sycl::event ungtr(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, std::complex<double> *a, std::int64_t lda, std::complex<double> *tau, std::complex<double> *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
 DLL_EXPORT void ungtr(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, sycl::buffer<std::complex<float>>  &a, std::int64_t lda, sycl::buffer<std::complex<float>>  &tau, sycl::buffer<std::complex<float>>  &scratchpad, std::int64_t scratchpad_size);
 DLL_EXPORT void ungtr(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, sycl::buffer<std::complex<double>> &a, std::int64_t lda, sycl::buffer<std::complex<double>> &tau, sycl::buffer<std::complex<double>> &scratchpad, std::int64_t scratchpad_size);
-template <typename data_t, oneapi::mkl::lapack::internal::is_complex_floating_point<data_t> = nullptr>
-std::int64_t ungtr_scratchpad_size(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, std::int64_t lda);
 
 // ormtr
 DLL_EXPORT sycl::event ormtr(sycl::queue &queue, oneapi::mkl::side side, oneapi::mkl::uplo uplo, oneapi::mkl::transpose trans, std::int64_t m, std::int64_t n, float  *a, std::int64_t lda, float  *tau, float  *c, std::int64_t ldc, float  *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
 DLL_EXPORT sycl::event ormtr(sycl::queue &queue, oneapi::mkl::side side, oneapi::mkl::uplo uplo, oneapi::mkl::transpose trans, std::int64_t m, std::int64_t n, double *a, std::int64_t lda, double *tau, double *c, std::int64_t ldc, double *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
 DLL_EXPORT void ormtr(sycl::queue &queue, oneapi::mkl::side side, oneapi::mkl::uplo uplo, oneapi::mkl::transpose trans, std::int64_t m, std::int64_t n, sycl::buffer<float>  &a, std::int64_t lda, sycl::buffer<float>  &tau, sycl::buffer<float>  &c, std::int64_t ldc, sycl::buffer<float>  &scratchpad, std::int64_t scratchpad_size);
 DLL_EXPORT void ormtr(sycl::queue &queue, oneapi::mkl::side side, oneapi::mkl::uplo uplo, oneapi::mkl::transpose trans, std::int64_t m, std::int64_t n, sycl::buffer<double> &a, std::int64_t lda, sycl::buffer<double> &tau, sycl::buffer<double> &c, std::int64_t ldc, sycl::buffer<double> &scratchpad, std::int64_t scratchpad_size);
-template <typename data_t, oneapi::mkl::lapack::internal::is_real_floating_point<data_t> = nullptr>
-std::int64_t ormtr_scratchpad_size(sycl::queue &queue, oneapi::mkl::side side, oneapi::mkl::uplo uplo, oneapi::mkl::transpose trans, std::int64_t m, std::int64_t n, std::int64_t lda, std::int64_t ldc);
 
 // unmtr
 DLL_EXPORT sycl::event unmtr(sycl::queue &queue, oneapi::mkl::side side, oneapi::mkl::uplo uplo, oneapi::mkl::transpose trans, std::int64_t m, std::int64_t n, std::complex<float>  *a, std::int64_t lda, std::complex<float>  *tau, std::complex<float>  *c, std::int64_t ldc, std::complex<float>  *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
 DLL_EXPORT sycl::event unmtr(sycl::queue &queue, oneapi::mkl::side side, oneapi::mkl::uplo uplo, oneapi::mkl::transpose trans, std::int64_t m, std::int64_t n, std::complex<double> *a, std::int64_t lda, std::complex<double> *tau, std::complex<double> *c, std::int64_t ldc, std::complex<double> *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
 DLL_EXPORT void unmtr(sycl::queue &queue, oneapi::mkl::side side, oneapi::mkl::uplo uplo, oneapi::mkl::transpose trans, std::int64_t m, std::int64_t n, sycl::buffer<std::complex<float>>  &a, std::int64_t lda, sycl::buffer<std::complex<float>>  &tau, sycl::buffer<std::complex<float>>  &c, std::int64_t ldc, sycl::buffer<std::complex<float>>  &scratchpad, std::int64_t scratchpad_size);
 DLL_EXPORT void unmtr(sycl::queue &queue, oneapi::mkl::side side, oneapi::mkl::uplo uplo, oneapi::mkl::transpose trans, std::int64_t m, std::int64_t n, sycl::buffer<std::complex<double>> &a, std::int64_t lda, sycl::buffer<std::complex<double>> &tau, sycl::buffer<std::complex<double>> &c, std::int64_t ldc, sycl::buffer<std::complex<double>> &scratchpad, std::int64_t scratchpad_size);
-template <typename data_t, oneapi::mkl::lapack::internal::is_complex_floating_point<data_t> = nullptr>
-std::int64_t unmtr_scratchpad_size(sycl::queue &queue, oneapi::mkl::side side, oneapi::mkl::uplo uplo, oneapi::mkl::transpose trans, std::int64_t m, std::int64_t n, std::int64_t lda, std::int64_t ldc);
+
+// gels
+DLL_EXPORT sycl::event gels(sycl::queue &queue, oneapi::mkl::transpose trans, std::int64_t m,  std::int64_t n, std::int64_t nrhs, float *a, std::int64_t lda, float *b, std::int64_t ldb, float *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
+DLL_EXPORT sycl::event gels(sycl::queue &queue, oneapi::mkl::transpose trans, std::int64_t m,  std::int64_t n, std::int64_t nrhs, double *a, std::int64_t lda, double *b, std::int64_t ldb, double *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
+DLL_EXPORT sycl::event gels(sycl::queue &queue, oneapi::mkl::transpose trans, std::int64_t m,  std::int64_t n, std::int64_t nrhs, std::complex<float> *a, std::int64_t lda, std::complex<float> *b, std::int64_t ldb, std::complex<float> *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
+DLL_EXPORT sycl::event gels(sycl::queue &queue, oneapi::mkl::transpose trans, std::int64_t m,  std::int64_t n, std::int64_t nrhs, std::complex<double> *a, std::int64_t lda, std::complex<double> *b, std::int64_t ldb, std::complex<double> *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
 
 //
 // DPC++ MKL LAPACK batch group API
@@ -1226,32 +1029,6 @@ DLL_EXPORT sycl::event gels_batch(sycl::queue &queue, oneapi::mkl::transpose *tr
 DLL_EXPORT sycl::event gels_batch(sycl::queue &queue, oneapi::mkl::transpose *trans, std::int64_t *m,  std::int64_t *n, std::int64_t *nrhs, std::complex<float> **a, std::int64_t *lda,std::complex<float> **b, std::int64_t *ldb, std::int64_t group_count, std::int64_t *group_sizes, std::complex<float> *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
 DLL_EXPORT sycl::event gels_batch(sycl::queue &queue, oneapi::mkl::transpose *trans, std::int64_t *m,  std::int64_t *n, std::int64_t *nrhs, std::complex<double> **a, std::int64_t *lda, std::complex<double> **b, std::int64_t *ldb, std::int64_t group_count, std::int64_t *group_sizes, std::complex<double> *scratchpad, std::int64_t scratchpad_size, const std::vector<sycl::event> &events = {});
 
-template <typename data_t, oneapi::mkl::lapack::internal::is_floating_point<data_t> = nullptr>
-std::int64_t potrf_batch_scratchpad_size(sycl::queue &queue, oneapi::mkl::uplo *uplo, std::int64_t *n, std::int64_t *lda, std::int64_t group_count, std::int64_t *group_sizes);
-template <typename data_t, oneapi::mkl::lapack::internal::is_floating_point<data_t> = nullptr>
-std::int64_t potrs_batch_scratchpad_size(sycl::queue &queue, oneapi::mkl::uplo *uplo, std::int64_t *n, std::int64_t *nrhs, std::int64_t *lda, std::int64_t *ldb, std::int64_t group_count, std::int64_t *group_sizes);
-template <typename data_t, oneapi::mkl::lapack::internal::is_floating_point<data_t> = nullptr>
-std::int64_t geinv_batch_scratchpad_size(sycl::queue &queue, std::int64_t *n, std::int64_t *lda, std::int64_t group_count, std::int64_t *group_sizes);
-template <typename data_t, oneapi::mkl::lapack::internal::is_floating_point<data_t> = nullptr>
-std::int64_t getrs_batch_scratchpad_size(sycl::queue &queue, oneapi::mkl::transpose *trans, std::int64_t *n, std::int64_t *nrhs, std::int64_t *lda, std::int64_t *ldb, std::int64_t group_count, std::int64_t *group_sizes);
-template <typename data_t, oneapi::mkl::lapack::internal::is_floating_point<data_t> = nullptr>
-std::int64_t getri_batch_scratchpad_size(sycl::queue &queue, std::int64_t *n, std::int64_t *lda, std::int64_t group_count, std::int64_t *group_sizes);
-template <typename data_t, oneapi::mkl::lapack::internal::is_floating_point<data_t> = nullptr>
-std::int64_t geqrf_batch_scratchpad_size(sycl::queue &queue, std::int64_t *m, std::int64_t *n, std::int64_t *lda, std::int64_t group_count, std::int64_t *group_sizes);
-template <typename data_t, oneapi::mkl::lapack::internal::is_real_floating_point<data_t> = nullptr>
-std::int64_t orgqr_batch_scratchpad_size(sycl::queue &queue, std::int64_t *m, std::int64_t *n, std::int64_t *k, std::int64_t *lda, std::int64_t group_count, std::int64_t *group_sizes);
-template <typename data_t, oneapi::mkl::lapack::internal::is_complex_floating_point<data_t> = nullptr>
-std::int64_t ungqr_batch_scratchpad_size(sycl::queue &queue, std::int64_t *m, std::int64_t *n, std::int64_t *k, std::int64_t *lda, std::int64_t group_count, std::int64_t *group_sizes);
-
-template <typename data_t, oneapi::mkl::lapack::internal::is_real_floating_point<data_t> = nullptr>
-std::int64_t ormqr_batch_scratchpad_size(sycl::queue &queue, oneapi::mkl::side *side, oneapi::mkl::transpose *trans, std::int64_t *m, std::int64_t *n, std::int64_t *k, std::int64_t *lda, std::int64_t *ldc, std::int64_t group_count, std::int64_t *group_sizes);
-template <typename data_t, oneapi::mkl::lapack::internal::is_complex_floating_point<data_t> = nullptr>
-std::int64_t unmqr_batch_scratchpad_size(sycl::queue &queue, oneapi::mkl::side *side, oneapi::mkl::transpose *trans, std::int64_t *m, std::int64_t *n, std::int64_t *k, std::int64_t *lda, std::int64_t *ldc, std::int64_t group_count, std::int64_t *group_sizes);
-template <typename data_t, oneapi::mkl::lapack::internal::is_floating_point<data_t> = nullptr>
-std::int64_t trtrs_batch_scratchpad_size(sycl::queue &queue, oneapi::mkl::uplo *uplo, oneapi::mkl::transpose *trans, oneapi::mkl::diag *diag, std::int64_t *n, std::int64_t *nrhs, std::int64_t *lda, std::int64_t *ldb, std::int64_t group_count, std::int64_t *group_sizes);
-template <typename data_t, oneapi::mkl::lapack::internal::is_floating_point<data_t> = nullptr>
-std::int64_t gels_batch_scratchpad_size(sycl::queue &queue, oneapi::mkl::transpose *trans, int64_t *m, std::int64_t *n, std::int64_t *nrhs, std::int64_t *lda, std::int64_t *ldb, std::int64_t group_count, std::int64_t *group_sizes);
-
 //
 // DPC++ MKL LAPACK batch stride API
 //
@@ -1311,24 +1088,8 @@ DLL_EXPORT void gels_batch(sycl::queue &queue, oneapi::mkl::transpose trans, std
 DLL_EXPORT void gels_batch(sycl::queue &queue, oneapi::mkl::transpose trans, std::int64_t m, std::int64_t n, std::int64_t nrhs, sycl::buffer<std::complex<float>> &a, std::int64_t lda, std::int64_t stridea, sycl::buffer<std::complex<float>> &b, std::int64_t ldb, std::int64_t strideb, std::int64_t batchsize, sycl::buffer<std::complex<float>> &scratchpad, std::int64_t scratchpad_size);
 DLL_EXPORT void gels_batch(sycl::queue &queue, oneapi::mkl::transpose trans, std::int64_t m, std::int64_t n, std::int64_t nrhs, sycl::buffer<std::complex<double>> &a, std::int64_t lda, std::int64_t stridea, sycl::buffer<std::complex<double>> &b, std::int64_t ldb, std::int64_t strideb, std::int64_t batchsize, sycl::buffer<std::complex<double>> &scratchpad, std::int64_t scratchpad_size);
 
-template <typename data_t, oneapi::mkl::lapack::internal::is_floating_point<data_t> = nullptr>
-std::int64_t potrf_batch_scratchpad_size(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, std::int64_t lda, std::int64_t stride_a, std::int64_t batch_size);
-template <typename data_t, oneapi::mkl::lapack::internal::is_floating_point<data_t> = nullptr>
-std::int64_t potrs_batch_scratchpad_size(sycl::queue &queue, oneapi::mkl::uplo uplo, std::int64_t n, std::int64_t nrhs, std::int64_t lda, std::int64_t stride_a, std::int64_t ldb, std::int64_t stride_b, std::int64_t batch_size);
-template <typename data_t, oneapi::mkl::lapack::internal::is_floating_point<data_t> = nullptr>
-std::int64_t geqrf_batch_scratchpad_size(sycl::queue &queue, std::int64_t m, std::int64_t n, std::int64_t lda, std::int64_t stride_a, std::int64_t stride_tau, std::int64_t batch_size);
-template <typename data_t, oneapi::mkl::lapack::internal::is_real_floating_point<data_t> = nullptr>
-std::int64_t orgqr_batch_scratchpad_size(sycl::queue &queue, std::int64_t m, std::int64_t n, std::int64_t k, std::int64_t lda, std::int64_t stride_a, std::int64_t stride_tau, std::int64_t batch_size);
-template <typename data_t, oneapi::mkl::lapack::internal::is_complex_floating_point<data_t> = nullptr>
-std::int64_t ungqr_batch_scratchpad_size(sycl::queue &queue, std::int64_t m, std::int64_t n, std::int64_t k, std::int64_t lda, std::int64_t stride_a, std::int64_t stride_tau, std::int64_t batch_size);
-template <typename data_t, oneapi::mkl::lapack::internal::is_floating_point<data_t> = nullptr>
-std::int64_t getri_batch_scratchpad_size(sycl::queue &queue, std::int64_t n, std::int64_t lda, std::int64_t stride_a, std::int64_t stride_ipiv, std::int64_t batch_size);
-template <typename data_t, oneapi::mkl::lapack::internal::is_floating_point<data_t> = nullptr>
-std::int64_t gels_batch_scratchpad_size(sycl::queue &queue, oneapi::mkl::transpose trans, std::int64_t m, std::int64_t n, std::int64_t nrhs, std::int64_t lda, std::int64_t stride_a, std::int64_t ldb, std::int64_t stride_b, std::int64_t batch_size);
-
-} //namespace lapack
-} //namespace mkl
+} // namespace lapack
+} // namespace mkl
 } // namespace oneapi
 
-
-#endif  // _LAPACK_LAPACK_HPP__
+#endif  // _ONEAPI_MKL_LAPACK_LAPACK_HPP__