Merge branch 'master' into feature/eig_kernel

Author: shssf

dpnp/backend/backend_iface.hpp

@@ -171,22 +171,6 @@ INP_DLLEXPORT void custom_sum_c(void* array, void* result, size_t size);
 template <typename _DataType>
 INP_DLLEXPORT void custom_prod_c(void* array, void* result, size_t size);
 
-/**
- * @ingroup BACKEND_API
- * @brief MKL implementation of dot function
- *
- * @param [in]  array1  Input array.
- *
- * @param [in]  array2  Input array.
- *
- * @param [out] result1 Output array.
- *
- * @param [in]  size    Number of elements in input arrays.
- *
- */
-template <typename _DataType>
-INP_DLLEXPORT void mkl_blas_dot_c(void* array1, void* array2, void* result1, size_t size);
-
 /**
  * @ingroup BACKEND_API
  * @brief Custom implementation of eig function

dpnp/backend/backend_iface_fptr.cpp

@@ -91,11 +91,11 @@ void* get_backend_function_name(const char* func_name, const char* type_name)
     {
         if (!strncmp(type_name, supported_type1_name, strlen(supported_type1_name)))
         {
-            return reinterpret_cast<void*>(mkl_blas_dot_c<double>);
+            return reinterpret_cast<void*>(custom_blas_dot_c<double>);
         }
         else if (!strncmp(type_name, supported_type2_name, strlen(supported_type2_name)))
         {
-            return reinterpret_cast<void*>(mkl_blas_dot_c<float>);
+            return reinterpret_cast<void*>(custom_blas_dot_c<float>);
         }
         else if (!strncmp(type_name, supported_type3_name, strlen(supported_type3_name)))
         {
@@ -312,8 +312,8 @@ static func_map_t func_map_init()
 
     fmap[DPNPFuncName::DPNP_FN_DOT][eft_INT][eft_INT] = {eft_INT, (void*)custom_blas_dot_c<int>};
     fmap[DPNPFuncName::DPNP_FN_DOT][eft_LNG][eft_LNG] = {eft_LNG, (void*)custom_blas_dot_c<long>};
-    fmap[DPNPFuncName::DPNP_FN_DOT][eft_FLT][eft_FLT] = {eft_FLT, (void*)mkl_blas_dot_c<float>};
-    fmap[DPNPFuncName::DPNP_FN_DOT][eft_DBL][eft_DBL] = {eft_DBL, (void*)mkl_blas_dot_c<double>};
+    fmap[DPNPFuncName::DPNP_FN_DOT][eft_FLT][eft_FLT] = {eft_FLT, (void*)custom_blas_dot_c<float>};
+    fmap[DPNPFuncName::DPNP_FN_DOT][eft_DBL][eft_DBL] = {eft_DBL, (void*)custom_blas_dot_c<double>};
 
     fmap[DPNPFuncName::DPNP_FN_EIG][eft_INT][eft_INT] = {eft_DBL, (void*)custom_lapack_syevd_c<double>};
     fmap[DPNPFuncName::DPNP_FN_EIG][eft_LNG][eft_LNG] = {eft_DBL, (void*)custom_lapack_syevd_c<double>};

dpnp/backend/custom_kernels.cpp

@@ -133,35 +133,58 @@ void custom_blas_dot_c(void* array1_in, void* array2_in, void* result1, size_t s
     _DataType* array_2 = reinterpret_cast<_DataType*>(array2_in);
     _DataType* result = reinterpret_cast<_DataType*>(result1);
 
-    _DataType* local_mem = reinterpret_cast<_DataType*>(dpnp_memory_alloc_c(size * sizeof(_DataType)));
+    if (!size)
+    {
+        return;
+    }
 
-    // what about reduction??
-    cl::sycl::range<1> gws(size);
-    event = DPNP_QUEUE.submit([&](cl::sycl::handler& cgh) {
-            cgh.parallel_for<class custom_blas_dot_c_kernel<_DataType> >(gws, [=](cl::sycl::id<1> global_id)
-            {
-                const size_t index = global_id[0];
-                local_mem[index] = array_1[index] * array_2[index];
-            }     // kernel lambda
-            );    // parallel_for
-    }             // task lambda
-    );            // queue.submit
+    if constexpr (std::is_same<_DataType, double>::value || std::is_same<_DataType, float>::value)
+    {
+        event = mkl_blas::dot(DPNP_QUEUE,
+                              size,
+                              array_1,
+                              1, // array_1 stride
+                              array_2,
+                              1, // array_2 stride
+                              result);
+        event.wait();
+    }
+    else
+    {
+        _DataType* local_mem = reinterpret_cast<_DataType*>(dpnp_memory_alloc_c(size * sizeof(_DataType)));
 
-    event.wait();
+        // what about reduction??
+        cl::sycl::range<1> gws(size);
 
-    auto policy = oneapi::dpl::execution::make_device_policy<class custom_blas_dot_c_kernel<_DataType>>(DPNP_QUEUE);
+        auto kernel_parallel_for_func = [=](cl::sycl::id<1> global_id) {
+            const size_t index = global_id[0];
+            local_mem[index] = array_1[index] * array_2[index];
+        };
+
+        auto kernel_func = [&](cl::sycl::handler& cgh) {
+            cgh.parallel_for<class custom_blas_dot_c_kernel<_DataType>>(gws, kernel_parallel_for_func);
+        };
+
+        event = DPNP_QUEUE.submit(kernel_func);
 
-    _DataType accumulator = 0;
-    accumulator = std::reduce(policy, local_mem, local_mem + size, _DataType(0), std::plus<_DataType>());
-    policy.queue().wait();
+        event.wait();
+
+        auto policy = oneapi::dpl::execution::make_device_policy<class custom_blas_dot_c_kernel<_DataType>>(DPNP_QUEUE);
+
+        _DataType accumulator = 0;
+        accumulator = std::reduce(policy, local_mem, local_mem + size, _DataType(0), std::plus<_DataType>());
+        policy.queue().wait();
 
-    result[0] = accumulator;
+        result[0] = accumulator;
 
-    free(local_mem, DPNP_QUEUE);
+        free(local_mem, DPNP_QUEUE);
+    }
 }
 
-template void custom_blas_dot_c<long>(void* array1_in, void* array2_in, void* result1, size_t size);
 template void custom_blas_dot_c<int>(void* array1_in, void* array2_in, void* result1, size_t size);
+template void custom_blas_dot_c<long>(void* array1_in, void* array2_in, void* result1, size_t size);
+template void custom_blas_dot_c<float>(void* array1_in, void* array2_in, void* result1, size_t size);
+template void custom_blas_dot_c<double>(void* array1_in, void* array2_in, void* result1, size_t size);
 
 template <typename _DataType>
 void custom_lapack_syevd_c(void* array_in, void* result1, size_t size)