summaryrefslogtreecommitdiff
path: root/utests/compiler_sub_group_shuffle.cpp
diff options
context:
space:
mode:
authorPan Xiuli <xiuli.pan@intel.com>2016-10-19 14:37:21 +0800
committerYang Rong <rong.r.yang@intel.com>2016-11-03 12:24:06 +0800
commitac9f80a8f93d4fe258cd3286b4b8b97cec7207b2 (patch)
tree70a6a9ef6b65e781f4f804ed5e04b098ca4619db /utests/compiler_sub_group_shuffle.cpp
parentc997d76b510065ee22615faa1a9a4062021096d8 (diff)
downloadbeignet-ac9f80a8f93d4fe258cd3286b4b8b97cec7207b2.tar.gz
Utest: Add test case for short type sub group shuffle
Signed-off-by: Pan Xiuli <xiuli.pan@intel.com> Reviewed-by: Yang Rong <rong.r.yang@intel.com>
Diffstat (limited to 'utests/compiler_sub_group_shuffle.cpp')
-rw-r--r--utests/compiler_sub_group_shuffle.cpp52
1 files changed, 49 insertions, 3 deletions
diff --git a/utests/compiler_sub_group_shuffle.cpp b/utests/compiler_sub_group_shuffle.cpp
index f33e9de5..2aadfeda 100644
--- a/utests/compiler_sub_group_shuffle.cpp
+++ b/utests/compiler_sub_group_shuffle.cpp
@@ -1,6 +1,6 @@
#include "utest_helper.hpp"
-void compiler_sub_group_shuffle(void)
+void compiler_sub_group_shuffle_int(void)
{
if(!cl_check_subgroups())
return;
@@ -8,7 +8,8 @@ void compiler_sub_group_shuffle(void)
const int32_t buf_size = 4 * n + 1;
// Setup kernel and buffers
- OCL_CREATE_KERNEL("compiler_sub_group_shuffle");
+ OCL_CREATE_KERNEL_FROM_FILE("compiler_sub_group_shuffle",
+ "compiler_sub_group_shuffle_int");
OCL_CREATE_BUFFER(buf[0], 0, buf_size * sizeof(int), NULL);
OCL_SET_ARG(0, sizeof(cl_mem), &buf[0]);
@@ -43,5 +44,50 @@ void compiler_sub_group_shuffle(void)
}
OCL_UNMAP_BUFFER(0);
}
+MAKE_UTEST_FROM_FUNCTION(compiler_sub_group_shuffle_int);
+void compiler_sub_group_shuffle_short(void)
+{
+ if(!cl_check_subgroups_short())
+ return;
+ const size_t n = 32;
+ const int32_t buf_size = 4 * n + 1;
+
+ // Setup kernel and buffers
+ OCL_CALL(cl_kernel_init, "compiler_sub_group_shuffle.cl",
+ "compiler_sub_group_shuffle_short",
+ SOURCE, "-DSHORT");
+ OCL_CREATE_BUFFER(buf[0], 0, buf_size * sizeof(short), 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)
+ ((short*)buf_data[0])[i] = -1;
+ OCL_UNMAP_BUFFER(0);
+
+ // Run the kernel on GPU
+ OCL_NDRANGE(1);
-MAKE_UTEST_FROM_FUNCTION(compiler_sub_group_shuffle);
+ // Compare
+ OCL_MAP_BUFFER(0);
+ short* dst = (short*)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]);
+ OCL_ASSERT((round * suggroupsize + c) == dst[4*i+1]);
+ OCL_ASSERT((round * suggroupsize + 5) == dst[4*i+2]);
+ OCL_ASSERT((round * suggroupsize + (suggroupsize - index - 1)) == dst[4*i+3]);
+ }
+ OCL_UNMAP_BUFFER(0);
+}
+MAKE_UTEST_FROM_FUNCTION(compiler_sub_group_shuffle_short);