Merge pull request #1520 from IntelPython/use_dpctl_round_func

use_dpctl_round_func_in_dpnp

Author: vtavana

dpnp/backend/extensions/vm/round.hpp

@@ -0,0 +1,78 @@
+//*****************************************************************************
+// Copyright (c) 2023, Intel Corporation
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are met:
+// - Redistributions of source code must retain the above copyright notice,
+//   this list of conditions and the following disclaimer.
+// - Redistributions in binary form must reproduce the above copyright notice,
+//   this list of conditions and the following disclaimer in the documentation
+//   and/or other materials provided with the distribution.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+// AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+// IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+// ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
+// LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+// CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+// SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+// INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+// CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+// ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
+// THE POSSIBILITY OF SUCH DAMAGE.
+//*****************************************************************************
+
+#pragma once
+
+#include <CL/sycl.hpp>
+
+#include "common.hpp"
+#include "types_matrix.hpp"
+
+namespace dpnp
+{
+namespace backend
+{
+namespace ext
+{
+namespace vm
+{
+template <typename T>
+sycl::event round_contig_impl(sycl::queue exec_q,
+                              const std::int64_t n,
+                              const char *in_a,
+                              char *out_y,
+                              const std::vector<sycl::event> &depends)
+{
+    type_utils::validate_type_for_device<T>(exec_q);
+
+    const T *a = reinterpret_cast<const T *>(in_a);
+    T *y = reinterpret_cast<T *>(out_y);
+
+    return mkl_vm::rint(exec_q,
+                        n, // number of elements to be calculated
+                        a, // pointer `a` containing input vector of size n
+                        y, // pointer `y` to the output vector of size n
+                        depends);
+}
+
+template <typename fnT, typename T>
+struct RoundContigFactory
+{
+    fnT get()
+    {
+        if constexpr (std::is_same_v<
+                          typename types::RoundOutputType<T>::value_type, void>)
+        {
+            return nullptr;
+        }
+        else {
+            return round_contig_impl<T>;
+        }
+    }
+};
+} // namespace vm
+} // namespace ext
+} // namespace backend
+} // namespace dpnp

dpnp/backend/extensions/vm/types_matrix.hpp

@@ -203,6 +203,21 @@ struct MulOutputType
         dpctl_td_ns::DefaultResultEntry<void>>::result_type;
 };
 
+/**
+ * @brief A factory to define pairs of supported types for which
+ * MKL VM library provides support in oneapi::mkl::vm::rint<T> function.
+ *
+ * @tparam T Type of input vector `a` and of result vector `y`.
+ */
+template <typename T>
+struct RoundOutputType
+{
+    using value_type = typename std::disjunction<
+        dpctl_td_ns::TypeMapResultEntry<T, double, double>,
+        dpctl_td_ns::TypeMapResultEntry<T, float, float>,
+        dpctl_td_ns::DefaultResultEntry<void>>::result_type;
+};
+
 /**
  * @brief A factory to define pairs of supported types for which
  * MKL VM library provides support in oneapi::mkl::vm::sin<T> function.

dpnp/backend/extensions/vm/vm_py.cpp

@@ -39,6 +39,7 @@
 #include "floor.hpp"
 #include "ln.hpp"
 #include "mul.hpp"
+#include "round.hpp"
 #include "sin.hpp"
 #include "sqr.hpp"
 #include "sqrt.hpp"
@@ -60,6 +61,7 @@ static unary_impl_fn_ptr_t floor_dispatch_vector[dpctl_td_ns::num_types];
 static unary_impl_fn_ptr_t conj_dispatch_vector[dpctl_td_ns::num_types];
 static unary_impl_fn_ptr_t ln_dispatch_vector[dpctl_td_ns::num_types];
 static binary_impl_fn_ptr_t mul_dispatch_vector[dpctl_td_ns::num_types];
+static unary_impl_fn_ptr_t round_dispatch_vector[dpctl_td_ns::num_types];
 static unary_impl_fn_ptr_t sin_dispatch_vector[dpctl_td_ns::num_types];
 static unary_impl_fn_ptr_t sqr_dispatch_vector[dpctl_td_ns::num_types];
 static unary_impl_fn_ptr_t sqrt_dispatch_vector[dpctl_td_ns::num_types];
@@ -301,6 +303,34 @@ PYBIND11_MODULE(_vm_impl, m)
               py::arg("dst"));
     }
 
+    // UnaryUfunc: ==== Round(x) ====
+    {
+        vm_ext::init_ufunc_dispatch_vector<unary_impl_fn_ptr_t,
+                                           vm_ext::RoundContigFactory>(
+            round_dispatch_vector);
+
+        auto round_pyapi = [&](sycl::queue exec_q, arrayT src, arrayT dst,
+                               const event_vecT &depends = {}) {
+            return vm_ext::unary_ufunc(exec_q, src, dst, depends,
+                                       round_dispatch_vector);
+        };
+        m.def("_round", round_pyapi,
+              "Call `rint` function from OneMKL VM library to compute "
+              "the rounded value of vector elements",
+              py::arg("sycl_queue"), py::arg("src"), py::arg("dst"),
+              py::arg("depends") = py::list());
+
+        auto round_need_to_call_pyapi = [&](sycl::queue exec_q, arrayT src,
+                                            arrayT dst) {
+            return vm_ext::need_to_call_unary_ufunc(exec_q, src, dst,
+                                                    round_dispatch_vector);
+        };
+        m.def("_mkl_round_to_call", round_need_to_call_pyapi,
+              "Check input arguments to answer if `rint` function from "
+              "OneMKL VM library can be used",
+              py::arg("sycl_queue"), py::arg("src"), py::arg("dst"));
+    }
+
     // UnaryUfunc: ==== Sin(x) ====
     {
         vm_ext::init_ufunc_dispatch_vector<unary_impl_fn_ptr_t,

dpnp/backend/include/dpnp_iface_fptr.hpp

@@ -101,8 +101,6 @@ enum class DPNPFuncName : size_t
     DPNP_FN_ARGSORT_EXT,  /**< Used in numpy.argsort() impl, requires extra
                              parameters */
     DPNP_FN_AROUND,       /**< Used in numpy.around() impl  */
-    DPNP_FN_AROUND_EXT,   /**< Used in numpy.around() impl, requires extra
-                             parameters */
     DPNP_FN_ASTYPE,       /**< Used in numpy.astype() impl  */
     DPNP_FN_ASTYPE_EXT,   /**< Used in numpy.astype() impl, requires extra
                              parameters */

dpnp/backend/kernels/dpnp_krnl_mathematical.cpp

@@ -109,15 +109,6 @@ template <typename _DataType>
 void (*dpnp_around_default_c)(const void *, void *, const size_t, const int) =
     dpnp_around_c<_DataType>;
 
-template <typename _DataType>
-DPCTLSyclEventRef (*dpnp_around_ext_c)(DPCTLSyclQueueRef,
-                                       const void *,
-                                       void *,
-                                       const size_t,
-                                       const int,
-                                       const DPCTLEventVectorRef) =
-    dpnp_around_c<_DataType>;
-
 template <typename _KernelNameSpecialization1,
           typename _KernelNameSpecialization2>
 class dpnp_elemwise_absolute_c_kernel;
@@ -1184,15 +1175,6 @@ void func_map_init_mathematical(func_map_t &fmap)
     fmap[DPNPFuncName::DPNP_FN_AROUND][eft_DBL][eft_DBL] = {
         eft_DBL, (void *)dpnp_around_default_c<double>};
 
-    fmap[DPNPFuncName::DPNP_FN_AROUND_EXT][eft_INT][eft_INT] = {
-        eft_INT, (void *)dpnp_around_ext_c<int32_t>};
-    fmap[DPNPFuncName::DPNP_FN_AROUND_EXT][eft_LNG][eft_LNG] = {
-        eft_LNG, (void *)dpnp_around_ext_c<int64_t>};
-    fmap[DPNPFuncName::DPNP_FN_AROUND_EXT][eft_FLT][eft_FLT] = {
-        eft_FLT, (void *)dpnp_around_ext_c<float>};
-    fmap[DPNPFuncName::DPNP_FN_AROUND_EXT][eft_DBL][eft_DBL] = {
-        eft_DBL, (void *)dpnp_around_ext_c<double>};
-
     fmap[DPNPFuncName::DPNP_FN_CROSS][eft_INT][eft_INT] = {
         eft_INT, (void *)dpnp_cross_default_c<int32_t, int32_t, int32_t>};
     fmap[DPNPFuncName::DPNP_FN_CROSS][eft_INT][eft_LNG] = {

dpnp/dpnp_algo/dpnp_algo.pxd

@@ -58,8 +58,6 @@ cdef extern from "dpnp_iface_fptr.hpp" namespace "DPNPFuncName":  # need this na
         DPNP_FN_ARGMIN_EXT
         DPNP_FN_ARGSORT
         DPNP_FN_ARGSORT_EXT
-        DPNP_FN_AROUND
-        DPNP_FN_AROUND_EXT
         DPNP_FN_ASTYPE
         DPNP_FN_ASTYPE_EXT
         DPNP_FN_CBRT

dpnp/dpnp_algo/dpnp_algo.pyx

@@ -170,7 +170,7 @@ cpdef utils.dpnp_descriptor dpnp_flatten(utils.dpnp_descriptor x1):
 
 cpdef utils.dpnp_descriptor dpnp_init_val(shape, dtype, value):
     """
-    same as dpnp_full(). TODO remove code dumplication
+    same as dpnp_full(). TODO remove code duplication
     """
     cdef DPNPFuncType param1_type = dpnp_dtype_to_DPNPFuncType(dtype)
 

dpnp/dpnp_algo/dpnp_algo_mathematical.pxi

@@ -38,7 +38,6 @@ and the rest of the library
 __all__ += [
     "dpnp_absolute",
     "dpnp_arctan2",
-    "dpnp_around",
     "dpnp_copysign",
     "dpnp_cross",
     "dpnp_cumprod",
@@ -72,9 +71,6 @@ ctypedef c_dpctl.DPCTLSyclEventRef(*fptr_1in_2out_t)(c_dpctl.DPCTLSyclQueueRef,
 ctypedef c_dpctl.DPCTLSyclEventRef(*ftpr_custom_trapz_2in_1out_with_2size_t)(c_dpctl.DPCTLSyclQueueRef,
                                                                              void *, void * , void * , double, size_t, size_t,
                                                                              const c_dpctl.DPCTLEventVectorRef)
-ctypedef c_dpctl.DPCTLSyclEventRef(*ftpr_custom_around_1in_1out_t)(c_dpctl.DPCTLSyclQueueRef,
-                                                                   const void * , void * , const size_t, const int,
-                                                                   const c_dpctl.DPCTLEventVectorRef)
 
 
 cpdef utils.dpnp_descriptor dpnp_absolute(utils.dpnp_descriptor x1):
@@ -120,38 +116,6 @@ cpdef utils.dpnp_descriptor dpnp_arctan2(utils.dpnp_descriptor x1_obj,
     return call_fptr_2in_1out_strides(DPNP_FN_ARCTAN2_EXT, x1_obj, x2_obj, dtype, out, where, func_name="arctan2")
 
 
-cpdef utils.dpnp_descriptor dpnp_around(utils.dpnp_descriptor x1, int decimals):
-
-    cdef DPNPFuncType param1_type = dpnp_dtype_to_DPNPFuncType(x1.dtype)
-
-    cdef DPNPFuncData kernel_data = get_dpnp_function_ptr(DPNP_FN_AROUND_EXT, param1_type, param1_type)
-
-    x1_obj = x1.get_array()
-
-    # ceate result array with type given by FPTR data
-    cdef shape_type_c result_shape = x1.shape
-    cdef utils.dpnp_descriptor result = utils.create_output_descriptor(result_shape,
-                                                                       kernel_data.return_type,
-                                                                       None,
-                                                                       device=x1_obj.sycl_device,
-                                                                       usm_type=x1_obj.usm_type,
-                                                                       sycl_queue=x1_obj.sycl_queue)
-
-    result_sycl_queue = result.get_array().sycl_queue
-
-    cdef c_dpctl.SyclQueue q = <c_dpctl.SyclQueue> result_sycl_queue
-    cdef c_dpctl.DPCTLSyclQueueRef q_ref = q.get_queue_ref()
-
-    cdef ftpr_custom_around_1in_1out_t func = <ftpr_custom_around_1in_1out_t > kernel_data.ptr
-
-    cdef c_dpctl.DPCTLSyclEventRef event_ref = func(q_ref, x1.get_data(), result.get_data(), x1.size, decimals, NULL)
-
-    with nogil: c_dpctl.DPCTLEvent_WaitAndThrow(event_ref)
-    c_dpctl.DPCTLEvent_Delete(event_ref)
-
-    return result
-
-
 cpdef utils.dpnp_descriptor dpnp_copysign(utils.dpnp_descriptor x1_obj,
                                           utils.dpnp_descriptor x2_obj,
                                           object dtype=None,

dpnp/dpnp_algo/dpnp_elementwise_common.py

@@ -69,6 +69,7 @@
     "dpnp_not_equal",
     "dpnp_remainder",
     "dpnp_right_shift",
+    "dpnp_round",
     "dpnp_sin",
     "dpnp_sqrt",
     "dpnp_square",
@@ -1545,6 +1546,57 @@ def dpnp_right_shift(x1, x2, out=None, order="K"):
     return dpnp_array._create_from_usm_ndarray(res_usm)
 
 
+_round_docstring = """
+round(x, out=None, order='K')
+Rounds each element `x_i` of the input array `x` to
+the nearest integer-valued number.
+Args:
+    x (dpnp.ndarray):
+        Input array, expected to have numeric data type.
+    out ({None, dpnp.ndarray}, optional):
+        Output array to populate. Array must have the correct
+        shape and the expected data type.
+    order ("C","F","A","K", optional): memory layout of the new
+        output array, if parameter `out` is `None`.
+        Default: "K".
+Return:
+    dpnp.ndarray:
+        An array containing the element-wise rounded value. The data type
+        of the returned array is determined by the Type Promotion Rules.
+"""
+
+
+def _call_round(src, dst, sycl_queue, depends=None):
+    """A callback to register in UnaryElementwiseFunc class of dpctl.tensor"""
+
+    if depends is None:
+        depends = []
+
+    if vmi._mkl_round_to_call(sycl_queue, src, dst):
+        # call pybind11 extension for round() function from OneMKL VM
+        return vmi._round(sycl_queue, src, dst, depends)
+    return ti._round(src, dst, sycl_queue, depends)
+
+
+round_func = UnaryElementwiseFunc(
+    "round", ti._round_result_type, _call_round, _round_docstring
+)
+
+
+def dpnp_round(x, out=None, order="K"):
+    """
+    Invokes round() function from pybind11 extension of OneMKL VM if possible.
+
+    Otherwise fully relies on dpctl.tensor implementation for round() function.
+    """
+    # dpctl.tensor only works with usm_ndarray
+    x1_usm = dpnp.get_usm_ndarray(x)
+    out_usm = None if out is None else dpnp.get_usm_ndarray(out)
+
+    res_usm = round_func(x1_usm, out=out_usm, order=order)
+    return dpnp_array._create_from_usm_ndarray(res_usm)
+
+
 _sign_docstring = """
 sign(x, out=None, order="K")