Utest: Add test case for sub_group_shuffle_down/up/xor

V2:Add subgroups extension check.
Signed-off-by: Pan Xiuli <xiuli.pan@intel.com>
Reviewed-by: Yang Rong <rong.r.yang@intel.com>

1 files changed, 48 insertions, 0 deletions

diff --git a/utests/compiler_sub_group_shuffle_xor.cpp b/utests/compiler_sub_group_shuffle_xor.cpp
new file mode 100644
index 00000000..967ec3ed
--- /dev/null
+++ b/utests/compiler_sub_group_shuffle_xor.cpp
@@ -0,0 +1,48 @@
+#include "utest_helper.hpp"
+
+void compiler_sub_group_shuffle_xor(void)
+{
+  if(!cl_check_subgroups())
+    return;
+  const size_t n = 32;
+  const int32_t buf_size = 4 * n + 1;
+
+  // Setup kernel and buffers
+  OCL_CREATE_KERNEL("compiler_sub_group_shuffle_xor");
+  OCL_CREATE_BUFFER(buf[0], 0, buf_size * sizeof(int), NULL);
+  OCL_SET_ARG(0, sizeof(cl_mem), &buf[0]);
+
+  int c = 3;
+  OCL_SET_ARG(1, sizeof(int), &c);
+
+  globals[0] = n;
+  locals[0] = 16;
+
+  OCL_MAP_BUFFER(0);
+  for (int32_t i = 0; i < buf_size; ++i)
+    ((int*)buf_data[0])[i] = -1;
+  OCL_UNMAP_BUFFER(0);
+
+  // Run the kernel on GPU
+  OCL_NDRANGE(1);
+
+  // Compare
+  OCL_MAP_BUFFER(0);
+  int* dst = (int *)buf_data[0];
+  int suggroupsize = dst[0];
+  OCL_ASSERT(suggroupsize == 8 || suggroupsize == 16);
+
+  dst++;
+  for (int32_t i = 0; i < (int32_t) n; ++i){
+    int round = i / suggroupsize;
+    int index = i % suggroupsize;
+    OCL_ASSERT(index == dst[4*i]);
+    //printf("%d %d %d %d\n", i, dst[4*i+1], dst[4*i+2], dst[4*i+3]);
+    OCL_ASSERT((round * suggroupsize + (c ^ index)) == dst[4*i+1]);
+    OCL_ASSERT((round * suggroupsize + (index ^ (suggroupsize - index -1))) == dst[4*i+2]);
+    OCL_ASSERT((round * suggroupsize + (index ^ (index + 1) % suggroupsize)) == dst[4*i+3]);
+  }
+  OCL_UNMAP_BUFFER(0);
+}
+
+MAKE_UTEST_FROM_FUNCTION(compiler_sub_group_shuffle_xor);