Merge pull request #1890 from IntelPython/fix-gh-1887

Check pointer alignment when copying from strided array to C-contiguous array

Author: ndgrigorian

CHANGELOG.md

@@ -18,8 +18,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Fixed
 * Fix for `tensor.result_type` when all inputs are Python built-in scalars [gh-1877](https://github.com/IntelPython/dpctl/pull/1877)
-
 * Improved error in constructors `tensor.full` and `tensor.full_like` when provided a non-numeric fill value [gh-1878](https://github.com/IntelPython/dpctl/pull/1878)
+* Added a check for pointer alignment when copying to C-contiguous memory [gh-1890](https://github.com/IntelPython/dpctl/pull/1890)
 
 ### Maintenance
 

dpctl/tensor/libtensor/include/kernels/copy_as_contiguous.hpp

@@ -44,8 +44,8 @@ namespace copy_as_contig
 
 template <typename T,
           typename IndexerT,
-          int vec_sz = 4,
-          int n_vecs = 2,
+          std::uint32_t vec_sz = 4u,
+          std::uint32_t n_vecs = 2u,
           bool enable_sg_loadstore = true>
 class CopyAsCContigFunctor
 {
@@ -66,53 +66,63 @@ class CopyAsCContigFunctor
 
     void operator()(sycl::nd_item<1> ndit) const
     {
+        static_assert(vec_sz > 0);
+        static_assert(n_vecs > 0);
+        static_assert(vec_sz * n_vecs < (std::uint32_t(1) << 8));
+
+        constexpr std::uint8_t elems_per_wi =
+            static_cast<std::uint8_t>(vec_sz * n_vecs);
+
         using dpctl::tensor::type_utils::is_complex;
         if constexpr (!enable_sg_loadstore || is_complex<T>::value) {
-            const std::uint32_t sgSize =
+            const std::uint16_t sgSize =
                 ndit.get_sub_group().get_local_range()[0];
             const std::size_t gid = ndit.get_global_linear_id();
 
-            const std::size_t base =
-                (gid / sgSize) * sgSize * n_vecs * vec_sz + (gid % sgSize);
-            for (size_t offset = base;
-                 offset < std::min(nelems, base + sgSize * (n_vecs * vec_sz));
-                 offset += sgSize)
-            {
+            // base = (gid / sgSize) * sgSize * elems_per_wi + (gid % sgSize)
+            // gid % sgSize == gid - (gid / sgSize) * sgSize
+            const std::size_t elems_per_sg = sgSize * (elems_per_wi - 1);
+            const std::size_t base = (gid / sgSize) * elems_per_sg + gid;
+            const std::size_t offset_max =
+                std::min(nelems, base + sgSize * elems_per_wi);
+
+            for (size_t offset = base; offset < offset_max; offset += sgSize) {
                 auto src_offset = src_indexer(offset);
                 dst_p[offset] = src_p[src_offset];
             }
         }
         else {
             auto sg = ndit.get_sub_group();
-            const std::uint32_t sgSize = sg.get_local_range()[0];
-            const size_t base = n_vecs * vec_sz *
-                                (ndit.get_group(0) * ndit.get_local_range(0) +
-                                 sg.get_group_id()[0] * sgSize);
+            const std::uint16_t sgSize = sg.get_max_local_range()[0];
+            const size_t base =
+                elems_per_wi * (ndit.get_group(0) * ndit.get_local_range(0) +
+                                sg.get_group_id()[0] * sgSize);
 
-            if (base + n_vecs * vec_sz * sgSize < nelems) {
+            if (base + elems_per_wi * sgSize < nelems) {
                 sycl::vec<T, vec_sz> dst_vec;
 
 #pragma unroll
-                for (std::uint32_t it = 0; it < n_vecs * vec_sz; it += vec_sz) {
+                for (std::uint8_t it = 0; it < elems_per_wi; it += vec_sz) {
+                    const size_t block_start_id = base + it * sgSize;
                     auto dst_multi_ptr = sycl::address_space_cast<
                         sycl::access::address_space::global_space,
-                        sycl::access::decorated::yes>(
-                        &dst_p[base + it * sgSize]);
+                        sycl::access::decorated::yes>(&dst_p[block_start_id]);
 
+                    const size_t elem_id0 = block_start_id + sg.get_local_id();
 #pragma unroll
-                    for (std::uint32_t k = 0; k < vec_sz; k++) {
-                        ssize_t src_offset = src_indexer(
-                            base + (it + k) * sgSize + sg.get_local_id());
+                    for (std::uint8_t k = 0; k < vec_sz; k++) {
+                        const size_t elem_id = elem_id0 + k * sgSize;
+                        const ssize_t src_offset = src_indexer(elem_id);
                         dst_vec[k] = src_p[src_offset];
                     }
                     sg.store<vec_sz>(dst_multi_ptr, dst_vec);
                 }
             }
             else {
-                for (size_t k = base + sg.get_local_id()[0]; k < nelems;
-                     k += sgSize)
-                {
-                    ssize_t src_offset = src_indexer(k);
+                const size_t lane_id = sg.get_local_id()[0];
+                const size_t k0 = base + lane_id;
+                for (size_t k = k0; k < nelems; k += sgSize) {
+                    const ssize_t src_offset = src_indexer(k);
                     dst_p[k] = src_p[src_offset];
                 }
             }
@@ -121,36 +131,23 @@ class CopyAsCContigFunctor
 };
 
 template <typename T,
-          typename IndexT,
-          int vec_sz,
-          int n_vecs,
-          bool enable_sgload>
-class as_contig_krn;
-
-template <typename T>
-sycl::event
-as_c_contiguous_array_generic_impl(sycl::queue &exec_q,
-                                   size_t nelems,
-                                   int nd,
-                                   const ssize_t *shape_and_strides,
-                                   const char *src_p,
-                                   char *dst_p,
-                                   const std::vector<sycl::event> &depends)
+          typename IndexerT,
+          std::uint32_t vec_sz,
+          std::uint32_t n_vecs,
+          bool enable_sg_load,
+          typename KernelName>
+sycl::event submit_c_contiguous_copy(sycl::queue &exec_q,
+                                     size_t nelems,
+                                     const T *src,
+                                     T *dst,
+                                     const IndexerT &src_indexer,
+                                     const std::vector<sycl::event> &depends)
 {
-    dpctl::tensor::type_utils::validate_type_for_device<T>(exec_q);
-
-    const T *src_tp = reinterpret_cast<const T *>(src_p);
-    T *dst_tp = reinterpret_cast<T *>(dst_p);
-
-    using IndexerT = dpctl::tensor::offset_utils::StridedIndexer;
-    const IndexerT src_indexer(nd, ssize_t(0), shape_and_strides);
+    static_assert(vec_sz > 0);
+    static_assert(n_vecs > 0);
+    static_assert(vec_sz * n_vecs < (std::uint32_t(1) << 8));
 
     constexpr std::size_t preferred_lws = 256;
-    constexpr std::uint32_t n_vecs = 2;
-    constexpr std::uint32_t vec_sz = 4;
-    constexpr bool enable_sg_load = true;
-    using KernelName =
-        as_contig_krn<T, IndexerT, vec_sz, n_vecs, enable_sg_load>;
 
     const auto &kernel_id = sycl::get_kernel_id<KernelName>();
 
@@ -167,9 +164,11 @@ as_c_contiguous_array_generic_impl(sycl::queue &exec_q,
     const std::size_t lws =
         ((preferred_lws + max_sg_size - 1) / max_sg_size) * max_sg_size;
 
-    constexpr std::uint32_t nelems_per_wi = n_vecs * vec_sz;
-    size_t n_groups =
-        (nelems + nelems_per_wi * lws - 1) / (nelems_per_wi * lws);
+    constexpr std::uint8_t nelems_per_wi = n_vecs * vec_sz;
+
+    const size_t nelems_per_group = nelems_per_wi * lws;
+    const size_t n_groups =
+        (nelems + nelems_per_group - 1) / (nelems_per_group);
 
     sycl::event copy_ev = exec_q.submit([&](sycl::handler &cgh) {
         cgh.depends_on(depends);
@@ -181,8 +180,62 @@ as_c_contiguous_array_generic_impl(sycl::queue &exec_q,
         cgh.parallel_for<KernelName>(
             sycl::nd_range<1>(gRange, lRange),
             CopyAsCContigFunctor<T, IndexerT, vec_sz, n_vecs, enable_sg_load>(
-                nelems, src_tp, dst_tp, src_indexer));
+                nelems, src, dst, src_indexer));
     });
+    return copy_ev;
+}
+
+template <typename T,
+          typename IndexT,
+          std::uint32_t vec_sz,
+          std::uint32_t n_vecs,
+          bool enable_sgload>
+class as_contig_krn;
+
+template <typename T>
+sycl::event
+as_c_contiguous_array_generic_impl(sycl::queue &exec_q,
+                                   size_t nelems,
+                                   int nd,
+                                   const ssize_t *shape_and_strides,
+                                   const char *src_p,
+                                   char *dst_p,
+                                   const std::vector<sycl::event> &depends)
+{
+    dpctl::tensor::type_utils::validate_type_for_device<T>(exec_q);
+
+    const T *src_tp = reinterpret_cast<const T *>(src_p);
+    T *dst_tp = reinterpret_cast<T *>(dst_p);
+
+    using IndexerT = dpctl::tensor::offset_utils::StridedIndexer;
+    const IndexerT src_indexer(nd, ssize_t(0), shape_and_strides);
+
+    constexpr std::uint32_t vec_sz = 4u;
+    constexpr std::uint32_t n_vecs = 2u;
+
+    using dpctl::tensor::kernels::alignment_utils::
+        disabled_sg_loadstore_wrapper_krn;
+    using dpctl::tensor::kernels::alignment_utils::is_aligned;
+    using dpctl::tensor::kernels::alignment_utils::required_alignment;
+
+    sycl::event copy_ev;
+    if (is_aligned<required_alignment>(dst_p)) {
+        constexpr bool enable_sg_load = true;
+        using KernelName =
+            as_contig_krn<T, IndexerT, vec_sz, n_vecs, enable_sg_load>;
+        copy_ev = submit_c_contiguous_copy<T, IndexerT, vec_sz, n_vecs,
+                                           enable_sg_load, KernelName>(
+            exec_q, nelems, src_tp, dst_tp, src_indexer, depends);
+    }
+    else {
+        constexpr bool disable_sg_load = false;
+        using InnerKernelName =
+            as_contig_krn<T, IndexerT, vec_sz, n_vecs, disable_sg_load>;
+        using KernelName = disabled_sg_loadstore_wrapper_krn<InnerKernelName>;
+        copy_ev = submit_c_contiguous_copy<T, IndexerT, vec_sz, n_vecs,
+                                           disable_sg_load, KernelName>(
+            exec_q, nelems, src_tp, dst_tp, src_indexer, depends);
+    }
 
     return copy_ev;
 }

dpctl/tests/test_usm_ndarray_ctor.py

@@ -2632,3 +2632,19 @@ def test_full_functions_raise_type_error():
     x = dpt.ones(1, dtype="i4")
     with pytest.raises(TypeError):
         dpt.full_like(x, "0")
+
+
+@pytest.mark.parametrize("dt", _all_dtypes)
+def test_setitem_copy_as_contig_alignment(dt):
+    q = get_queue_or_skip()
+    skip_if_dtype_not_supported(dt, q)
+
+    dtype_ = dpt.dtype(dt)
+    n0, n1 = 8, 23
+
+    x = dpt.zeros((n0, n1), dtype=dtype_, sycl_queue=q)
+
+    vals = dpt.ones(n1, dtype=dtype_, sycl_queue=q)[dpt.newaxis, :]
+    x[1:, ...] = vals
+    assert dpt.all(x[0] == 0)
+    assert dpt.all(x[1:, :] == vals)