Backprt gh-2333 (#2343)

This PR backports of #2333 from development branch to
`maintenance/0.17.x`.

Author: antonwolfy

CHANGELOG.md

@@ -103,6 +103,7 @@ Furthermore, a number of issues relating to running on NVIDIA GPUs have been res
 * Resolved an issue with wrong result returned by `dpnp.tensordot` for integer data types [#2296](https://github.com/IntelPython/dpnp/pull/2296)
 * Resolved `ValueError` exception raised by `dpnp.linalg.qr` with non-contiguous input array [#2314](https://github.com/IntelPython/dpnp/pull/2314)
 * Resolved an issue with wrong result returned by `dpnp.fft.fftn` and `dpnp.fft.rfftn` when running on NVIDIA GPU [#2332](https://github.com/IntelPython/dpnp/pull/2332)
+* Added a workaround to prevent a memory corruption in `dpnp.correlate` [#2333](https://github.com/IntelPython/dpnp/pull/2333)
 
 
 ## [0.16.1] - 12/06/2024

dpnp/backend/extensions/statistics/sliding_window1d.hpp

@@ -436,7 +436,7 @@ class PaddedSpan : public Span<T, SizeT>
     using size_type = SizeT;
 
     PaddedSpan(T *const data, const SizeT size, const SizeT pad)
-        : Span<T>(data, size), pad_(pad)
+        : Span<T, SizeT>(data, size), pad_(pad)
     {
     }
 
@@ -574,9 +574,20 @@ void submit_sliding_window1d(const PaddedSpan<const T, SizeT> &a,
             }
 
             auto *const out_ptr = out.begin();
-            auto *const out_end = out.end();
-            results.store(&out_ptr[glid],
-                          [out_end](auto &&ptr) { return ptr < out_end; });
+            // auto *const out_end = out.end();
+
+            auto y_start = glid;
+            auto y_stop =
+                std::min(y_start + WorkPI * results.size_x(), out.size());
+            uint32_t i = 0;
+            for (uint32_t y = y_start; y < y_stop; y += results.size_x()) {
+                out_ptr[y] = results[i++];
+            }
+            // while the code itself seems to be valid, inside correlate
+            // kernel it results in memory corruption. Further investigation
+            // is needed. SAT-7693
+            // corruption results.store(&out_ptr[glid],
+            //               [out_end](auto &&ptr) { return ptr < out_end; });
         });
 }
 
@@ -635,9 +646,20 @@ void submit_sliding_window1d_small_kernel(const PaddedSpan<const T, SizeT> &a,
                           red);
 
             auto *const out_ptr = out.begin();
-            auto *const out_end = out.end();
-            results.store(&out_ptr[glid],
-                          [out_end](auto &&ptr) { return ptr < out_end; });
+            // auto *const out_end = out.end();
+
+            auto y_start = glid;
+            auto y_stop =
+                std::min(y_start + WorkPI * results.size_x(), out.size());
+            uint32_t i = 0;
+            for (uint32_t y = y_start; y < y_stop; y += results.size_x()) {
+                out_ptr[y] = results[i++];
+            }
+            // while the code itself seems to be valid, inside correlate
+            // kernel it results in memory corruption. Further investigation
+            // is needed. SAT-7693
+            // corruption results.store(&out_ptr[glid],
+            //               [out_end](auto &&ptr) { return ptr < out_end; });
         });
 }
 

dpnp/tests/test_usm_type.py

@@ -800,9 +800,6 @@ def test_1in_1out(func, data, usm_type):
 @pytest.mark.parametrize("usm_type_x", list_of_usm_types, ids=list_of_usm_types)
 @pytest.mark.parametrize("usm_type_y", list_of_usm_types, ids=list_of_usm_types)
 def test_2in_1out(func, data1, data2, usm_type_x, usm_type_y):
-    if func == "correlate" and is_win_platform():
-        pytest.skip("due to SAT-7693")
-
     x = dp.array(data1, usm_type=usm_type_x)
     y = dp.array(data2, usm_type=usm_type_y)
     z = getattr(dp, func)(x, y)