Move _tensor_reductions_impl extension and use it for dpnp (#2794)

This PR completely moves `_tensor_reductions_impl` pybind11 extension
into `dpctl_ext.tensor` and extends dpctl_ext.tensor Python API with the
functions: `all, any, diff, argmax, argmin, count_nonzero, logsumexp, max. min, prod, reduce_hypot and sum`
 reusing them in dpnp

Author: vlad-perevezentsev

dpctl_ext/tensor/CMakeLists.txt

@@ -69,6 +69,19 @@ set(_accumulator_sources
     ${CMAKE_CURRENT_SOURCE_DIR}/libtensor/source/accumulators/cumulative_prod.cpp
     ${CMAKE_CURRENT_SOURCE_DIR}/libtensor/source/accumulators/cumulative_sum.cpp
 )
+set(_reduction_sources
+    ${CMAKE_CURRENT_SOURCE_DIR}/libtensor/source/reductions/reduction_common.cpp
+    ${CMAKE_CURRENT_SOURCE_DIR}/libtensor/source/reductions/all.cpp
+    ${CMAKE_CURRENT_SOURCE_DIR}/libtensor/source/reductions/any.cpp
+    ${CMAKE_CURRENT_SOURCE_DIR}/libtensor/source/reductions/argmax.cpp
+    ${CMAKE_CURRENT_SOURCE_DIR}/libtensor/source/reductions/argmin.cpp
+    ${CMAKE_CURRENT_SOURCE_DIR}/libtensor/source/reductions/logsumexp.cpp
+    ${CMAKE_CURRENT_SOURCE_DIR}/libtensor/source/reductions/max.cpp
+    ${CMAKE_CURRENT_SOURCE_DIR}/libtensor/source/reductions/min.cpp
+    ${CMAKE_CURRENT_SOURCE_DIR}/libtensor/source/reductions/prod.cpp
+    ${CMAKE_CURRENT_SOURCE_DIR}/libtensor/source/reductions/reduce_hypot.cpp
+    ${CMAKE_CURRENT_SOURCE_DIR}/libtensor/source/reductions/sum.cpp
+)
 set(_sorting_sources
     ${CMAKE_CURRENT_SOURCE_DIR}/libtensor/source/sorting/isin.cpp
     ${CMAKE_CURRENT_SOURCE_DIR}/libtensor/source/sorting/merge_sort.cpp
@@ -82,6 +95,10 @@ set(_tensor_accumulation_impl_sources
     ${CMAKE_CURRENT_SOURCE_DIR}/libtensor/source/tensor_accumulation.cpp
     ${_accumulator_sources}
 )
+set(_tensor_reductions_impl_sources
+    ${CMAKE_CURRENT_SOURCE_DIR}/libtensor/source/tensor_reductions.cpp
+    ${_reduction_sources}
+)
 set(_tensor_sorting_impl_sources
     ${CMAKE_CURRENT_SOURCE_DIR}/libtensor/source/tensor_sorting.cpp
     ${_sorting_sources}
@@ -114,6 +131,12 @@ add_sycl_to_target(TARGET ${python_module_name} SOURCES ${_tensor_accumulation_i
 target_link_libraries(${python_module_name} PRIVATE ${_static_lib_trgt})
 list(APPEND _py_trgts ${python_module_name})
 
+set(python_module_name _tensor_reductions_impl)
+pybind11_add_module(${python_module_name} MODULE ${_tensor_reductions_impl_sources})
+add_sycl_to_target(TARGET ${python_module_name} SOURCES ${_tensor_reductions_impl_sources})
+target_link_libraries(${python_module_name} PRIVATE ${_static_lib_trgt})
+list(APPEND _py_trgts ${python_module_name})
+
 set(python_module_name _tensor_sorting_impl)
 pybind11_add_module(${python_module_name} MODULE ${_tensor_sorting_impl_sources})
 add_sycl_to_target(TARGET ${python_module_name} SOURCES ${_tensor_sorting_impl_sources})
@@ -135,7 +158,7 @@ set(_no_fast_math_sources
 list(
     APPEND _no_fast_math_sources
     # ${_elementwise_sources}
-    # ${_reduction_sources}
+    ${_reduction_sources}
     ${_sorting_sources}
     # ${_linalg_sources}
     ${_accumulator_sources}

dpctl_ext/tensor/__init__.py

@@ -78,6 +78,17 @@
     tile,
     unstack,
 )
+from ._reduction import (
+    argmax,
+    argmin,
+    count_nonzero,
+    logsumexp,
+    max,
+    min,
+    prod,
+    reduce_hypot,
+    sum,
+)
 from ._reshape import reshape
 from ._search_functions import where
 from ._searchsorted import searchsorted
@@ -90,9 +101,14 @@
 )
 from ._sorting import argsort, sort, top_k
 from ._type_utils import can_cast, finfo, iinfo, isdtype, result_type
+from ._utility_functions import all, any, diff
 
 __all__ = [
+    "all",
+    "any",
     "arange",
+    "argmax",
+    "argmin",
     "argsort",
     "asarray",
     "asnumpy",
@@ -102,10 +118,12 @@
     "can_cast",
     "concat",
     "copy",
+    "count_nonzero",
     "clip",
     "cumulative_logsumexp",
     "cumulative_prod",
     "cumulative_sum",
+    "diff",
     "empty",
     "empty_like",
     "extract",
@@ -120,15 +138,20 @@
     "isdtype",
     "isin",
     "linspace",
+    "logsumexp",
+    "max",
     "meshgrid",
+    "min",
     "moveaxis",
     "permute_dims",
     "nonzero",
     "ones",
     "ones_like",
     "place",
+    "prod",
     "put",
     "put_along_axis",
+    "reduce_hypot",
     "repeat",
     "reshape",
     "result_type",
@@ -137,6 +160,7 @@
     "sort",
     "squeeze",
     "stack",
+    "sum",
     "swapaxes",
     "take",
     "take_along_axis",

dpctl_ext/tensor/_manipulation_functions.py

@@ -624,7 +624,7 @@ def repeat(x, repeats, /, *, axis=None):
                     "'repeats' array must be broadcastable to the size of "
                     "the repeated axis"
                 )
-            if not dpt.all(repeats >= 0):
+            if not dpt_ext.all(repeats >= 0):
                 raise ValueError("'repeats' elements must be positive")
 
     elif isinstance(repeats, (tuple, list, range)):
@@ -646,7 +646,7 @@ def repeat(x, repeats, /, *, axis=None):
             repeats = dpt_ext.asarray(
                 repeats, dtype=dpt.int64, usm_type=usm_type, sycl_queue=exec_q
             )
-            if not dpt.all(repeats >= 0):
+            if not dpt_ext.all(repeats >= 0):
                 raise ValueError("`repeats` elements must be positive")
     else:
         raise TypeError(