Merge pull request #1025 from IntelPython/master

Merge remote-tracking branch 'origin/master' into gold/2021

Author: xaleryb

conda-recipe/conda_build_config.yaml

@@ -1,2 +1,2 @@
 numpy:
-    - 1.21
+    - 1.22

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

@@ -129,6 +129,7 @@ sycl::event lin_space_step_impl(sycl::queue exec_q,
                                 char *array_data,
                                 const std::vector<sycl::event> &depends)
 {
+    dpctl::tensor::type_utils::validate_type_for_device<Ty>(exec_q);
     sycl::event lin_space_step_event = exec_q.submit([&](sycl::handler &cgh) {
         cgh.depends_on(depends);
         cgh.parallel_for<linear_sequence_step_kernel<Ty>>(
@@ -270,6 +271,8 @@ sycl::event lin_space_affine_impl(sycl::queue exec_q,
                                   char *array_data,
                                   const std::vector<sycl::event> &depends)
 {
+    dpctl::tensor::type_utils::validate_type_for_device<Ty>(exec_q);
+
     bool device_supports_doubles = exec_q.get_device().has(sycl::aspect::fp64);
     sycl::event lin_space_affine_event = exec_q.submit([&](sycl::handler &cgh) {
         cgh.depends_on(depends);
@@ -378,6 +381,7 @@ sycl::event full_contig_impl(sycl::queue q,
                              char *dst_p,
                              const std::vector<sycl::event> &depends)
 {
+    dpctl::tensor::type_utils::validate_type_for_device<dstTy>(q);
     sycl::event fill_ev = q.submit([&](sycl::handler &cgh) {
         cgh.depends_on(depends);
         dstTy *p = reinterpret_cast<dstTy *>(dst_p);
@@ -496,6 +500,7 @@ sycl::event eye_impl(sycl::queue exec_q,
                      char *array_data,
                      const std::vector<sycl::event> &depends)
 {
+    dpctl::tensor::type_utils::validate_type_for_device<Ty>(exec_q);
     sycl::event eye_event = exec_q.submit([&](sycl::handler &cgh) {
         cgh.depends_on(depends);
         cgh.parallel_for<eye_kernel<Ty>>(
@@ -576,6 +581,8 @@ sycl::event tri_impl(sycl::queue exec_q,
     Ty *src = reinterpret_cast<Ty *>(src_p);
     Ty *dst = reinterpret_cast<Ty *>(dst_p);
 
+    dpctl::tensor::type_utils::validate_type_for_device<Ty>(exec_q);
+
     sycl::event tri_ev = exec_q.submit([&](sycl::handler &cgh) {
         cgh.depends_on(depends);
         cgh.depends_on(additional_depends);

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

@@ -215,6 +215,9 @@ copy_and_cast_generic_impl(sycl::queue q,
                            const std::vector<sycl::event> &depends,
                            const std::vector<sycl::event> &additional_depends)
 {
+    dpctl::tensor::type_utils::validate_type_for_device<dstTy>(q);
+    dpctl::tensor::type_utils::validate_type_for_device<srcTy>(q);
+
     sycl::event copy_and_cast_ev = q.submit([&](sycl::handler &cgh) {
         cgh.depends_on(depends);
         cgh.depends_on(additional_depends);
@@ -317,6 +320,9 @@ copy_and_cast_nd_specialized_impl(sycl::queue q,
                                   py::ssize_t dst_offset,
                                   const std::vector<sycl::event> &depends)
 {
+    dpctl::tensor::type_utils::validate_type_for_device<dstTy>(q);
+    dpctl::tensor::type_utils::validate_type_for_device<srcTy>(q);
+
     sycl::event copy_and_cast_ev = q.submit([&](sycl::handler &cgh) {
         cgh.depends_on(depends);
         cgh.parallel_for<copy_cast_spec_kernel<srcTy, dstTy, nd>>(
@@ -486,6 +492,10 @@ void copy_and_cast_from_host_impl(
     const std::vector<sycl::event> &additional_depends)
 {
     py::ssize_t nelems_range = src_max_nelem_offset - src_min_nelem_offset + 1;
+
+    dpctl::tensor::type_utils::validate_type_for_device<dstTy>(q);
+    dpctl::tensor::type_utils::validate_type_for_device<srcTy>(q);
+
     sycl::buffer<srcTy, 1> npy_buf(
         reinterpret_cast<const srcTy *>(host_src_p) + src_min_nelem_offset,
         sycl::range<1>(nelems_range), {sycl::property::buffer::use_host_ptr{}});
@@ -637,6 +647,8 @@ copy_for_reshape_generic_impl(sycl::queue q,
                               char *dst_p,
                               const std::vector<sycl::event> &depends)
 {
+    dpctl::tensor::type_utils::validate_type_for_device<Ty>(q);
+
     sycl::event copy_for_reshape_ev = q.submit([&](sycl::handler &cgh) {
         cgh.depends_on(depends);
         cgh.parallel_for<copy_for_reshape_generic_kernel<Ty>>(

dpctl/tensor/libtensor/include/utils/type_utils.hpp

@@ -23,7 +23,9 @@
 //===----------------------------------------------------------------------===//
 
 #pragma once
+#include <CL/sycl.hpp>
 #include <complex>
+#include <exception>
 
 namespace dpctl
 {
@@ -68,6 +70,29 @@ template <typename dstTy, typename srcTy> dstTy convert_impl(const srcTy &v)
     }
 }
 
+template <typename T> void validate_type_for_device(const sycl::device &d)
+{
+    if constexpr (std::is_same_v<T, double>) {
+        if (!d.has(sycl::aspect::fp64)) {
+            throw std::runtime_error("Device " +
+                                     d.get_info<sycl::info::device::name>() +
+                                     " does not support type 'double'");
+        }
+    }
+    else if constexpr (std::is_same_v<T, sycl::half>) {
+        if (!d.has(sycl::aspect::fp16)) {
+            throw std::runtime_error("Device " +
+                                     d.get_info<sycl::info::device::name>() +
+                                     " does not support type 'half'");
+        }
+    }
+}
+
+template <typename T> void validate_type_for_device(const sycl::queue &q)
+{
+    validate_type_for_device<T>(q.get_device());
+}
+
 } // namespace type_utils
 } // namespace tensor
 } // namespace dpctl

dpctl/tensor/libtensor/source/device_support_queries.cpp

@@ -51,7 +51,9 @@ std::string _default_device_fp_type(sycl::device d)
 
 std::string _default_device_int_type(sycl::device)
 {
-    return "i8";
+    return "l"; // code for numpy.dtype('long') to be consisent
+                // with NumPy's default integer type across
+                // platforms.
 }
 
 std::string _default_device_complex_type(sycl::device d)