[SYCL][CUDA] Enables sub_group_sycl2020 for CUDA (#316)

sub_group_sycl2020.cpp test is currently marked as unsupported for CUDA
due to missing support for kernel-bundles. This patch adds a special
CUDA path for avoiding the use of kernel-bundles, using the known size
of sub-groups instead of querying for it.

Signed-off-by: Steffen Larsen <[email protected]>

Author: steffenlarsen

SYCL/SubGroup/sub_groups_sycl2020.cpp

@@ -1,55 +1,69 @@
 // RUN: %clangxx -fsycl -fsycl-targets=%sycl_triple %s -o %t.out
 // RUN: %CPU_RUN_PLACEHOLDER %t.out
 // RUN: %GPU_RUN_PLACEHOLDER %t.out
-//
-// TODO enable test on CUDA once kernel_bundle is supported
-// UNSUPPORTED: cuda
 
 #include <sycl/sycl.hpp>
 
 class TestKernel;
+class TestKernelCUDA;
 
 int main() {
   sycl::queue Q;
   sycl::buffer<int, 3> Buf{sycl::range{3, 32, 32}};
 
-  sycl::kernel_id TestKernelID = sycl::get_kernel_id<TestKernel>();
-
-  sycl::kernel_bundle KernelBundle =
-      sycl::get_kernel_bundle<sycl::bundle_state::executable>(Q.get_context(),
-                                                              {TestKernelID});
-
-  Q.submit([&](sycl::handler &CGH) {
-    CGH.use_kernel_bundle(KernelBundle);
-    sycl::accessor Acc{Buf, CGH, sycl::write_only};
-    CGH.parallel_for<TestKernel>(
-        sycl::nd_range<1>(sycl::range{32}, sycl::range{32}),
-        [=](sycl::nd_item<1> item) {
-          auto SG = item.get_sub_group();
-          Acc[0][SG.get_group_linear_id()][SG.get_local_linear_id()] =
-              SG.leader();
-          Acc[1][SG.get_group_linear_id()][SG.get_local_linear_id()] =
-              SG.get_group_linear_range();
-          Acc[2][SG.get_group_linear_id()][SG.get_local_linear_id()] =
-              SG.get_local_linear_range();
-        });
-  });
-
-  sycl::host_accessor Acc{Buf, sycl::read_only};
-
-  sycl::kernel Kernel = KernelBundle.get_kernel(TestKernelID);
-
-  const size_t SubgroupSize =
-      Kernel.get_info<sycl::info::kernel_device_specific::max_sub_group_size>(
-          Q.get_device(), sycl::range{32, 1, 1});
+  size_t SubgroupSize = 0;
+
+  sycl::accessor WriteAcc{Buf, sycl::write_only};
+  const auto KernelFunc = [=](sycl::nd_item<1> item) {
+    auto SG = item.get_sub_group();
+    WriteAcc[0][SG.get_group_linear_id()][SG.get_local_linear_id()] =
+        SG.leader();
+    WriteAcc[1][SG.get_group_linear_id()][SG.get_local_linear_id()] =
+        SG.get_group_linear_range();
+    WriteAcc[2][SG.get_group_linear_id()][SG.get_local_linear_id()] =
+        SG.get_local_linear_range();
+  };
+
+  if (Q.get_backend() != sycl::backend::cuda) {
+    sycl::kernel_id TestKernelID = sycl::get_kernel_id<TestKernel>();
+    sycl::kernel_bundle KernelBundle =
+        sycl::get_kernel_bundle<sycl::bundle_state::executable>(Q.get_context(),
+                                                                {TestKernelID});
+
+    Q.submit([&](sycl::handler &CGH) {
+      CGH.use_kernel_bundle(KernelBundle);
+      CGH.require(WriteAcc);
+      CGH.parallel_for<TestKernel>(
+          sycl::nd_range<1>(sycl::range{32}, sycl::range{32}), KernelFunc);
+    });
+
+    sycl::kernel Kernel = KernelBundle.get_kernel(TestKernelID);
+    SubgroupSize =
+        Kernel.get_info<sycl::info::kernel_device_specific::max_sub_group_size>(
+            Q.get_device(), sycl::range{32, 1, 1});
+  } else {
+    // CUDA sub-group size is 32 by default (size of a warp) so the kernel
+    // bundle is not strictly needed to do this test for the CUDA backend.
+    // TODO: Remove this special CUDA path once the CUDA backend supports kernel
+    // bundles.
+    SubgroupSize = 32;
+    Q.submit([&](sycl::handler &CGH) {
+      CGH.require(WriteAcc);
+      CGH.parallel_for<TestKernelCUDA>(
+          sycl::nd_range<1>(sycl::range{32}, sycl::range{32}), KernelFunc);
+    });
+  }
+
+  sycl::host_accessor HostAcc{Buf, sycl::read_only};
+
   const size_t MaxNumSubgroups = 32 / SubgroupSize;
 
   for (size_t SGNo = 0; SGNo < MaxNumSubgroups; SGNo++) {
     for (size_t WINo = 0; WINo < SubgroupSize; WINo++) {
       const int Leader = WINo == 0 ? 1 : 0;
-      assert(Acc[0][SGNo][WINo] == Leader);
-      assert(Acc[1][SGNo][WINo] == MaxNumSubgroups);
-      assert(Acc[2][SGNo][WINo] == SubgroupSize);
+      assert(HostAcc[0][SGNo][WINo] == Leader);
+      assert(HostAcc[1][SGNo][WINo] == MaxNumSubgroups);
+      assert(HostAcc[2][SGNo][WINo] == SubgroupSize);
     }
   }