[half]Add ifdefs around usages of half

sbalint98 · sbalint98 · commit 2e8f8d256943 · 2021-05-21T15:07:01.000+02:00
Use DISABLE_HALF_RUTINES, use cmake config to set
diff --git a/include/oneapi/mkl/blas.hxx b/include/oneapi/mkl/blas.hxx
@@ -261,7 +261,7 @@ static inline void gemm(cl::sycl::queue &queue, transpose transa, transpose tran
                  c, ldc);
     gemm_postcondition(queue, transa, transb, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
 }
-
+#ifndef DISABLE_HALF_RUTINES
 static inline void gemm(cl::sycl::queue &queue, transpose transa, transpose transb, std::int64_t m,
                         std::int64_t n, std::int64_t k, half alpha, cl::sycl::buffer<half, 1> &a,
                         std::int64_t lda, cl::sycl::buffer<half, 1> &b, std::int64_t ldb, half beta,
@@ -281,7 +281,7 @@ static inline void gemm(cl::sycl::queue &queue, transpose transa, transpose tran
                  c, ldc);
     gemm_postcondition(queue, transa, transb, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
 }
-
+#endif
 static inline void gemm_batch(cl::sycl::queue &queue, transpose transa, transpose transb,
                               std::int64_t m, std::int64_t n, std::int64_t k, float alpha,
                               cl::sycl::buffer<float, 1> &a, std::int64_t lda,
diff --git a/include/oneapi/mkl/blas/detail/blas_ct_backends.hpp b/include/oneapi/mkl/blas/detail/blas_ct_backends.hpp
@@ -25,6 +25,7 @@
 #include <cstdint>
 
 #include "oneapi/mkl/types.hpp"
+#include "oneapi/mkl/detail/config.hpp"
 #include "oneapi/mkl/detail/backend_selector.hpp"
 
 namespace oneapi {
diff --git a/include/oneapi/mkl/blas/detail/blas_ct_backends.hxx b/include/oneapi/mkl/blas/detail/blas_ct_backends.hxx
@@ -413,7 +413,7 @@ static inline void gemm(backend_selector<backend::BACKEND> selector, transpose t
                         std::int64_t lda, cl::sycl::buffer<std::complex<double>, 1> &b,
                         std::int64_t ldb, std::complex<double> beta,
                         cl::sycl::buffer<std::complex<double>, 1> &c, std::int64_t ldc);
-
+#ifndef DISABLE_HALF_RUTINES
 static inline void gemm(backend_selector<backend::BACKEND> selector, transpose transa,
                         transpose transb, std::int64_t m, std::int64_t n, std::int64_t k,
                         half alpha, cl::sycl::buffer<half, 1> &a, std::int64_t lda,
@@ -425,7 +425,7 @@ static inline void gemm(backend_selector<backend::BACKEND> selector, transpose t
                         float alpha, cl::sycl::buffer<half, 1> &a, std::int64_t lda,
                         cl::sycl::buffer<half, 1> &b, std::int64_t ldb, float beta,
                         cl::sycl::buffer<float, 1> &c, std::int64_t ldc);
-
+#endif
 static inline void herk(backend_selector<backend::BACKEND> selector, uplo upper_lower,
                         transpose trans, std::int64_t n, std::int64_t k, float alpha,
                         cl::sycl::buffer<std::complex<float>, 1> &a, std::int64_t lda, float beta,
diff --git a/include/oneapi/mkl/blas/detail/blas_loader.hpp b/include/oneapi/mkl/blas/detail/blas_loader.hpp
@@ -24,6 +24,7 @@
 #include <cstdint>
 #include <CL/sycl.hpp>
 
+#include "oneapi/mkl/detail/config.hpp"
 #include "oneapi/mkl/types.hpp"
 
 #include "oneapi/mkl/detail/export.hpp"
diff --git a/include/oneapi/mkl/blas/detail/blas_loader.hxx b/include/oneapi/mkl/blas/detail/blas_loader.hxx
@@ -396,6 +396,7 @@ ONEMKL_EXPORT void gemm(oneapi::mkl::device libkey, cl::sycl::queue &queue, tran
                         std::int64_t lda, cl::sycl::buffer<std::complex<double>, 1> &b,
                         std::int64_t ldb, std::complex<double> beta,
                         cl::sycl::buffer<std::complex<double>, 1> &c, std::int64_t ldc);
+#ifndef DISABLE_HALF_RUTINES
 ONEMKL_EXPORT void gemm(oneapi::mkl::device libkey, cl::sycl::queue &queue, transpose transa,
                         transpose transb, std::int64_t m, std::int64_t n, std::int64_t k,
                         half alpha, cl::sycl::buffer<half, 1> &a, std::int64_t lda,
@@ -406,6 +407,7 @@ ONEMKL_EXPORT void gemm(oneapi::mkl::device libkey, cl::sycl::queue &queue, tran
                         float alpha, cl::sycl::buffer<half, 1> &a, std::int64_t lda,
                         cl::sycl::buffer<half, 1> &b, std::int64_t ldb, float beta,
                         cl::sycl::buffer<float, 1> &c, std::int64_t ldc);
+#endif
 
 ONEMKL_EXPORT void syr2(oneapi::mkl::device libkey, cl::sycl::queue &queue, uplo upper_lower,
                         std::int64_t n, float alpha, cl::sycl::buffer<float, 1> &x,
diff --git a/include/oneapi/mkl/blas/detail/cublas/blas_ct.hpp b/include/oneapi/mkl/blas/detail/cublas/blas_ct.hpp
@@ -26,6 +26,7 @@
 
 #include "oneapi/mkl/types.hpp"
 #include "oneapi/mkl/detail/backend_selector.hpp"
+#include "oneapi/mkl/detail/config.hpp"
 #include "oneapi/mkl/blas/detail/cublas/onemkl_blas_cublas.hpp"
 #include "oneapi/mkl/blas/detail/blas_ct_backends.hpp"
 
diff --git a/include/oneapi/mkl/blas/detail/cublas/blas_ct.hxx b/include/oneapi/mkl/blas/detail/cublas/blas_ct.hxx
@@ -743,7 +743,7 @@ void gemm(backend_selector<backend::cublas> selector, transpose transa, transpos
     gemm_postcondition(selector.get_queue(), transa, transb, m, n, k, alpha, a, lda, b, ldb, beta,
                        c, ldc);
 }
-
+#ifndef DISABLE_HALF_RUTINES
 void gemm(backend_selector<backend::cublas> selector, transpose transa, transpose transb,
           std::int64_t m, std::int64_t n, std::int64_t k, half alpha, cl::sycl::buffer<half, 1> &a,
           std::int64_t lda, cl::sycl::buffer<half, 1> &b, std::int64_t ldb, half beta,
@@ -767,7 +767,7 @@ void gemm(backend_selector<backend::cublas> selector, transpose transa, transpos
     gemm_postcondition(selector.get_queue(), transa, transb, m, n, k, alpha, a, lda, b, ldb, beta,
                        c, ldc);
 }
-
+#endif
 void syr2(backend_selector<backend::cublas> selector, uplo upper_lower, std::int64_t n, float alpha,
           cl::sycl::buffer<float, 1> &x, std::int64_t incx, cl::sycl::buffer<float, 1> &y,
           std::int64_t incy, cl::sycl::buffer<float, 1> &a, std::int64_t lda) {
diff --git a/include/oneapi/mkl/blas/detail/cublas/onemkl_blas_cublas.hpp b/include/oneapi/mkl/blas/detail/cublas/onemkl_blas_cublas.hpp
@@ -23,6 +23,7 @@
 #include <cstdint>
 #include <string>
 #include "oneapi/mkl/types.hpp"
+#include "oneapi/mkl/detail/config.hpp"
 
 namespace oneapi {
 namespace mkl {
diff --git a/include/oneapi/mkl/blas/detail/cublas/onemkl_blas_cublas.hxx b/include/oneapi/mkl/blas/detail/cublas/onemkl_blas_cublas.hxx
@@ -496,7 +496,7 @@ void gemm(cl::sycl::queue &queue, transpose transa, transpose transb, std::int64
           cl::sycl::buffer<std::complex<double>, 1> &a, std::int64_t lda,
           cl::sycl::buffer<std::complex<double>, 1> &b, std::int64_t ldb, std::complex<double> beta,
           cl::sycl::buffer<std::complex<double>, 1> &c, std::int64_t ldc);
-
+#ifndef DISABLE_HALF_RUTINES
 void gemm(cl::sycl::queue &queue, transpose transa, transpose transb, std::int64_t m,
           std::int64_t n, std::int64_t k, half alpha, cl::sycl::buffer<half, 1> &a,
           std::int64_t lda, cl::sycl::buffer<half, 1> &b, std::int64_t ldb, half beta,
@@ -506,7 +506,7 @@ void gemm(cl::sycl::queue &queue, transpose transa, transpose transb, std::int64
           std::int64_t n, std::int64_t k, float alpha, cl::sycl::buffer<half, 1> &a,
           std::int64_t lda, cl::sycl::buffer<half, 1> &b, std::int64_t ldb, float beta,
           cl::sycl::buffer<float, 1> &c, std::int64_t ldc);
-
+#endif
 void hemm(cl::sycl::queue &queue, side left_right, uplo upper_lower, std::int64_t m, std::int64_t n,
           std::complex<float> alpha, cl::sycl::buffer<std::complex<float>, 1> &a, std::int64_t lda,
           cl::sycl::buffer<std::complex<float>, 1> &b, std::int64_t ldb, std::complex<float> beta,
diff --git a/include/oneapi/mkl/blas/detail/mklcpu/blas_ct.hpp b/include/oneapi/mkl/blas/detail/mklcpu/blas_ct.hpp
@@ -25,6 +25,7 @@
 #include <cstdint>
 
 #include "oneapi/mkl/types.hpp"
+#include "oneapi/mkl/detail/config.hpp"
 #include "oneapi/mkl/detail/backend_selector.hpp"
 
 #include "oneapi/mkl/blas/detail/blas_ct_backends.hpp"
diff --git a/include/oneapi/mkl/blas/detail/mklcpu/blas_ct.hxx b/include/oneapi/mkl/blas/detail/mklcpu/blas_ct.hxx
@@ -743,7 +743,7 @@ void gemm(backend_selector<backend::mklcpu> selector, transpose transa, transpos
     gemm_postcondition(selector.get_queue(), transa, transb, m, n, k, alpha, a, lda, b, ldb, beta,
                        c, ldc);
 }
-
+#ifndef DISABLE_HALF_RUTINES
 void gemm(backend_selector<backend::mklcpu> selector, transpose transa, transpose transb,
           std::int64_t m, std::int64_t n, std::int64_t k, half alpha, cl::sycl::buffer<half, 1> &a,
           std::int64_t lda, cl::sycl::buffer<half, 1> &b, std::int64_t ldb, half beta,
@@ -767,7 +767,7 @@ void gemm(backend_selector<backend::mklcpu> selector, transpose transa, transpos
     gemm_postcondition(selector.get_queue(), transa, transb, m, n, k, alpha, a, lda, b, ldb, beta,
                        c, ldc);
 }
-
+#endif
 void syr2(backend_selector<backend::mklcpu> selector, uplo upper_lower, std::int64_t n, float alpha,
           cl::sycl::buffer<float, 1> &x, std::int64_t incx, cl::sycl::buffer<float, 1> &y,
           std::int64_t incy, cl::sycl::buffer<float, 1> &a, std::int64_t lda) {
diff --git a/include/oneapi/mkl/blas/detail/mklgpu/blas_ct.hxx b/include/oneapi/mkl/blas/detail/mklgpu/blas_ct.hxx
@@ -743,7 +743,7 @@ void gemm(backend_selector<backend::mklgpu> selector, transpose transa, transpos
     gemm_postcondition(selector.get_queue(), transa, transb, m, n, k, alpha, a, lda, b, ldb, beta,
                        c, ldc);
 }
-
+#ifndef DISABLE_HALF_RUTINES
 void gemm(backend_selector<backend::mklgpu> selector, transpose transa, transpose transb,
           std::int64_t m, std::int64_t n, std::int64_t k, half alpha, cl::sycl::buffer<half, 1> &a,
           std::int64_t lda, cl::sycl::buffer<half, 1> &b, std::int64_t ldb, half beta,
@@ -767,7 +767,7 @@ void gemm(backend_selector<backend::mklgpu> selector, transpose transa, transpos
     gemm_postcondition(selector.get_queue(), transa, transb, m, n, k, alpha, a, lda, b, ldb, beta,
                        c, ldc);
 }
-
+#endif
 void syr2(backend_selector<backend::mklgpu> selector, uplo upper_lower, std::int64_t n, float alpha,
           cl::sycl::buffer<float, 1> &x, std::int64_t incx, cl::sycl::buffer<float, 1> &y,
           std::int64_t incy, cl::sycl::buffer<float, 1> &a, std::int64_t lda) {
diff --git a/include/oneapi/mkl/blas/detail/netlib/blas_ct.hxx b/include/oneapi/mkl/blas/detail/netlib/blas_ct.hxx
@@ -743,7 +743,7 @@ void gemm(backend_selector<backend::netlib> selector, transpose transa, transpos
     gemm_postcondition(selector.get_queue(), transa, transb, m, n, k, alpha, a, lda, b, ldb, beta,
                        c, ldc);
 }
-
+#ifndef DISABLE_HALF_RUTINES
 void gemm(backend_selector<backend::netlib> selector, transpose transa, transpose transb,
           std::int64_t m, std::int64_t n, std::int64_t k, half alpha, cl::sycl::buffer<half, 1> &a,
           std::int64_t lda, cl::sycl::buffer<half, 1> &b, std::int64_t ldb, half beta,
@@ -767,7 +767,7 @@ void gemm(backend_selector<backend::netlib> selector, transpose transa, transpos
     gemm_postcondition(selector.get_queue(), transa, transb, m, n, k, alpha, a, lda, b, ldb, beta,
                        c, ldc);
 }
-
+#endif
 void syr2(backend_selector<backend::netlib> selector, uplo upper_lower, std::int64_t n, float alpha,
           cl::sycl::buffer<float, 1> &x, std::int64_t incx, cl::sycl::buffer<float, 1> &y,
           std::int64_t incy, cl::sycl::buffer<float, 1> &a, std::int64_t lda) {
diff --git a/include/oneapi/mkl/blas/detail/onemkl_blas_backends.hxx b/include/oneapi/mkl/blas/detail/onemkl_blas_backends.hxx
@@ -46,7 +46,7 @@ ONEMKL_EXPORT void gemm(cl::sycl::queue &queue, oneapi::mkl::transpose transa,
                         cl::sycl::buffer<std::complex<double>, 1> &b, std::int64_t ldb,
                         std::complex<double> beta, cl::sycl::buffer<std::complex<double>, 1> &c,
                         std::int64_t ldc);
-
+#ifndef DISABLE_HALF_RUTINES
 ONEMKL_EXPORT void gemm(cl::sycl::queue &queue, oneapi::mkl::transpose transa,
                         oneapi::mkl::transpose transb, std::int64_t m, std::int64_t n,
                         std::int64_t k, half alpha, cl::sycl::buffer<half, 1> &a, std::int64_t lda,
@@ -58,7 +58,7 @@ ONEMKL_EXPORT void gemm(cl::sycl::queue &queue, oneapi::mkl::transpose transa,
                         std::int64_t k, float alpha, cl::sycl::buffer<half, 1> &a, std::int64_t lda,
                         cl::sycl::buffer<half, 1> &b, std::int64_t ldb, float beta,
                         cl::sycl::buffer<float, 1> &c, std::int64_t ldc);
-
+#endif
 ONEMKL_EXPORT void symm(cl::sycl::queue &queue, oneapi::mkl::side left_right,
                         oneapi::mkl::uplo upper_lower, std::int64_t m, std::int64_t n, float alpha,
                         cl::sycl::buffer<float, 1> &a, std::int64_t lda,
diff --git a/include/oneapi/mkl/blas/predicates.hpp b/include/oneapi/mkl/blas/predicates.hpp
@@ -26,6 +26,7 @@
 
 #include "oneapi/mkl/exceptions.hpp"
 #include "oneapi/mkl/types.hpp"
+#include "oneapi/mkl/detail/config.hpp"
 
 namespace oneapi {
 namespace mkl {
diff --git a/include/oneapi/mkl/blas/predicates.hxx b/include/oneapi/mkl/blas/predicates.hxx
@@ -1518,6 +1518,7 @@ inline void gemm_postcondition(cl::sycl::queue &queue, transpose transa, transpo
 #endif
 }
 
+#ifndef DISABLE_HALF_RUTINES
 inline void gemm_precondition(cl::sycl::queue &queue, transpose transa, transpose transb,
                               std::int64_t m, std::int64_t n, std::int64_t k, half alpha,
                               cl::sycl::buffer<half, 1> &a, std::int64_t lda,
@@ -1557,7 +1558,7 @@ inline void gemm_postcondition(cl::sycl::queue &queue, transpose transa, transpo
     /* add postchecks to queue here for input args.  */
 #endif
 }
-
+#endif
 inline void syr2_precondition(cl::sycl::queue &queue, uplo upper_lower, std::int64_t n, float alpha,
                               cl::sycl::buffer<float, 1> &x, std::int64_t incx,
                               cl::sycl::buffer<float, 1> &y, std::int64_t incy,
@@ -4749,7 +4750,7 @@ inline void gemm_postcondition(cl::sycl::queue &queue, transpose transa, transpo
     /* add postchecks to queue here for input args.  */
 #endif
 }
-
+#ifndef DISABLE_HALF_RUTINES
 inline void gemm_precondition(cl::sycl::queue &queue, transpose transa, transpose transb,
                               std::int64_t m, std::int64_t n, std::int64_t k, half alpha,
                               const half *a, std::int64_t lda, const half *b, std::int64_t ldb,
@@ -4769,7 +4770,7 @@ inline void gemm_postcondition(cl::sycl::queue &queue, transpose transa, transpo
     /* add postchecks to queue here for input args.  */
 #endif
 }
-
+#endif
 inline void syr2_precondition(cl::sycl::queue &queue, uplo upper_lower, std::int64_t n, float alpha,
                               const float *x, std::int64_t incx, const float *y, std::int64_t incy,
                               float *a, std::int64_t lda,
diff --git a/include/oneapi/mkl/detail/backend_selector.hpp b/include/oneapi/mkl/detail/backend_selector.hpp
@@ -27,6 +27,8 @@
 namespace oneapi {
 namespace mkl {
 
+using namespace cl;
+
 template <backend Backend>
 class backend_selector {
 public:
diff --git a/src/blas/backends/cublas/cublas_level3.cpp b/src/blas/backends/cublas/cublas_level3.cpp
@@ -109,10 +109,10 @@ inline void gemm(Func func, DATATYPE_A DT_A, DATATYPE_B DT_B, DATATYPE_C DT_C,
         gemm(CUBLAS_ROUTINE, CUDADATATYPE_A, CUDADATATYPE_B, CUDADATATYPE_C, queue, transa,      \
              transb, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);                              \
     }
-
+#ifndef DISABLE_HALF_RUTINES
 GEMM_EX_LAUNCHER(half, half, float, cublasGemmEx, CUDA_R_16F, CUDA_R_16F, CUDA_R_32F)
 GEMM_EX_LAUNCHER(half, half, half, cublasGemmEx, CUDA_R_16F, CUDA_R_16F, CUDA_R_16F)
-
+#endif
 #undef GEMM_EX_LAUNCHER
 
 template <typename Func, typename T>
@@ -465,14 +465,14 @@ GEMM_LAUNCHER_USM(std::complex<float>, cublasCgemm)
 GEMM_LAUNCHER_USM(std::complex<double>, cublasZgemm)
 
 #undef GEMM_LAUNCHER_USM
-
+#ifndef DISABLE_HALF_RUTINES
 cl::sycl::event gemm(cl::sycl::queue &queue, transpose transa, transpose transb, std::int64_t m,
                      std::int64_t n, std::int64_t k, half alpha, const half *a, std::int64_t lda,
                      const half *b, std::int64_t ldb, half beta, half *c, std::int64_t ldc,
                      const cl::sycl::vector_class<cl::sycl::event> &dependencies) {
     throw unimplemented("blas", "gemm", "for column_major layout");
 }
-
+#endif
 template <typename Func, typename T>
 inline cl::sycl::event symm(Func func, cl::sycl::queue &queue, side left_right, uplo upper_lower,
                             int64_t m, int64_t n, T alpha, const T *a, int64_t lda, const T *b,
@@ -860,10 +860,10 @@ inline void gemm(Func func, DATATYPE_A DT_A, DATATYPE_B DT_B, DATATYPE_C DT_C,
         gemm(CUBLAS_ROUTINE, CUDADATATYPE_A, CUDADATATYPE_B, CUDADATATYPE_C, queue, transa,      \
              transb, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);                              \
     }
-
+#ifndef DISABLE_HALF_RUTINES
 GEMM_EX_LAUNCHER(half, half, float, cublasGemmEx, CUDA_R_16F, CUDA_R_16F, CUDA_R_32F)
 GEMM_EX_LAUNCHER(half, half, half, cublasGemmEx, CUDA_R_16F, CUDA_R_16F, CUDA_R_16F)
-
+#endif
 #undef GEMM_EX_LAUNCHER
 
 template <typename Func, typename T>
@@ -1065,14 +1065,14 @@ GEMM_LAUNCHER_USM(std::complex<float>, cublasCgemm)
 GEMM_LAUNCHER_USM(std::complex<double>, cublasZgemm)
 
 #undef GEMM_LAUNCHER_USM
-
+#ifndef DISABLE_HALF_RUTINES
 cl::sycl::event gemm(cl::sycl::queue &queue, transpose transa, transpose transb, std::int64_t m,
                      std::int64_t n, std::int64_t k, half alpha, const half *a, std::int64_t lda,
                      const half *b, std::int64_t ldb, half beta, half *c, std::int64_t ldc,
                      const cl::sycl::vector_class<cl::sycl::event> &dependencies) {
     throw unimplemented("blas", "gemm", "for row_major layout");
 }
-
+#endif
 template <typename Func, typename T>
 inline cl::sycl::event symm(Func func, cl::sycl::queue &queue, side left_right, uplo upper_lower,
                             int64_t m, int64_t n, T alpha, const T *a, int64_t lda, const T *b,
diff --git a/src/blas/backends/cublas/mkl_blas_cublas_wrappers.cpp b/src/blas/backends/cublas/mkl_blas_cublas_wrappers.cpp
@@ -147,8 +147,10 @@ extern "C" blas_function_table_t mkl_blas_table = {
     oneapi::mkl::blas::cublas::column_major::gemm,
     oneapi::mkl::blas::cublas::column_major::gemm,
     oneapi::mkl::blas::cublas::column_major::gemm,
+#ifndef DISABLE_HALF_RUTINES
     oneapi::mkl::blas::cublas::column_major::gemm,
     oneapi::mkl::blas::cublas::column_major::gemm,
+#endif
     oneapi::mkl::blas::cublas::column_major::hemm,
     oneapi::mkl::blas::cublas::column_major::hemm,
     oneapi::mkl::blas::cublas::column_major::herk,
@@ -478,8 +480,10 @@ extern "C" blas_function_table_t mkl_blas_table = {
     oneapi::mkl::blas::cublas::row_major::gemm,
     oneapi::mkl::blas::cublas::row_major::gemm,
     oneapi::mkl::blas::cublas::row_major::gemm,
+#ifndef DISABLE_HALF_RUTINES
     oneapi::mkl::blas::cublas::row_major::gemm,
     oneapi::mkl::blas::cublas::row_major::gemm,
+#endif
     oneapi::mkl::blas::cublas::row_major::hemm,
     oneapi::mkl::blas::cublas::row_major::hemm,
     oneapi::mkl::blas::cublas::row_major::herk,
diff --git a/src/blas/backends/mklcpu/mklcpu_level3.cxx b/src/blas/backends/mklcpu/mklcpu_level3.cxx
@@ -96,7 +96,7 @@ void gemm(cl::sycl::queue &queue, transpose transa, transpose transb, int64_t m,
         });
     });
 }
-
+#ifndef DISABLE_HALF_RUTINES
 void gemm(cl::sycl::queue &queue, transpose transa, transpose transb, int64_t m, int64_t n,
           int64_t k, half alpha, cl::sycl::buffer<half, 1> &a, int64_t lda,
           cl::sycl::buffer<half, 1> &b, int64_t ldb, half beta, cl::sycl::buffer<half, 1> &c,
@@ -140,7 +140,6 @@ void gemm(cl::sycl::queue &queue, transpose transa, transpose transb, int64_t m,
         });
     });
 }
-
 void gemm(cl::sycl::queue &queue, transpose transa, transpose transb, int64_t m, int64_t n,
           int64_t k, float alpha, cl::sycl::buffer<half, 1> &a, int64_t lda,
           cl::sycl::buffer<half, 1> &b, int64_t ldb, float beta, cl::sycl::buffer<float, 1> &c,
@@ -173,7 +172,7 @@ void gemm(cl::sycl::queue &queue, transpose transa, transpose transb, int64_t m,
         });
     });
 }
-
+#endif
 void hemm(cl::sycl::queue &queue, side left_right, uplo upper_lower, int64_t m, int64_t n,
           std::complex<float> alpha, cl::sycl::buffer<std::complex<float>, 1> &a, int64_t lda,
           cl::sycl::buffer<std::complex<float>, 1> &b, int64_t ldb, std::complex<float> beta,
diff --git a/src/blas/backends/mklcpu/mklcpu_wrappers.cpp b/src/blas/backends/mklcpu/mklcpu_wrappers.cpp
@@ -148,8 +148,10 @@ extern "C" ONEMKL_EXPORT blas_function_table_t mkl_blas_table = {
     oneapi::mkl::blas::mklcpu::column_major::gemm,
     oneapi::mkl::blas::mklcpu::column_major::gemm,
     oneapi::mkl::blas::mklcpu::column_major::gemm,
+    #ifndef DISABLE_HALF_RUTINES
     oneapi::mkl::blas::mklcpu::column_major::gemm,
     oneapi::mkl::blas::mklcpu::column_major::gemm,
+    #endif
     oneapi::mkl::blas::mklcpu::column_major::hemm,
     oneapi::mkl::blas::mklcpu::column_major::hemm,
     oneapi::mkl::blas::mklcpu::column_major::herk,
@@ -479,8 +481,10 @@ extern "C" ONEMKL_EXPORT blas_function_table_t mkl_blas_table = {
     oneapi::mkl::blas::mklcpu::row_major::gemm,
     oneapi::mkl::blas::mklcpu::row_major::gemm,
     oneapi::mkl::blas::mklcpu::row_major::gemm,
+    #ifndef DISABLE_HALF_RUTINES
     oneapi::mkl::blas::mklcpu::row_major::gemm,
     oneapi::mkl::blas::mklcpu::row_major::gemm,
+    #endif
     oneapi::mkl::blas::mklcpu::row_major::hemm,
     oneapi::mkl::blas::mklcpu::row_major::hemm,
     oneapi::mkl::blas::mklcpu::row_major::herk,
diff --git a/src/blas/backends/mklgpu/mklgpu_common.hpp b/src/blas/backends/mklgpu/mklgpu_common.hpp
diff --git a/src/blas/backends/mklgpu/mklgpu_level3.cxx b/src/blas/backends/mklgpu/mklgpu_level3.cxx
diff --git a/src/blas/backends/mklgpu/mklgpu_wrappers.cpp b/src/blas/backends/mklgpu/mklgpu_wrappers.cpp
diff --git a/src/blas/blas_loader.cpp b/src/blas/blas_loader.cpp
diff --git a/src/blas/function_table.hpp b/src/blas/function_table.hpp
diff --git a/src/config.hpp.in b/src/config.hpp.in
diff --git a/src/rng/backends/mklcpu/mrg32k3a.cpp b/src/rng/backends/mklcpu/mrg32k3a.cpp
diff --git a/src/rng/backends/mklcpu/philox4x32x10.cpp b/src/rng/backends/mklcpu/philox4x32x10.cpp
