summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
-rw-r--r--kernels/builtin_shuffle2.cl13
-rw-r--r--utests/CMakeLists.txt1
-rw-r--r--utests/builtin_shuffle2.cpp45
3 files changed, 59 insertions, 0 deletions
diff --git a/kernels/builtin_shuffle2.cl b/kernels/builtin_shuffle2.cl
new file mode 100644
index 00000000..1a122d4a
--- /dev/null
+++ b/kernels/builtin_shuffle2.cl
@@ -0,0 +1,13 @@
+kernel void builtin_shuffle2(global float *src1, global float *src2, global float *dst1, global float *dst2) {
+ int i = get_global_id(0);
+ float2 x = (float2)(src1[i], src2[i]);
+ float2 y = (float2)(1234, 5678);
+ uint4 mask = (uint4)(1, 0, 0, 0);
+ float4 v1 = shuffle2(x, y, mask);
+ float16 x2 = 0;
+ float16 y2 = (float16)(src1[i], src2[i], 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0);
+ uint16 mask2 = (uint16)(17, 16, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0);
+ float16 v2 = shuffle2(x2, y2, mask2);
+ dst1[i] = v1.s0 + v2.s0;
+ dst2[i] = v1.s1 + v2.s1;
+}
diff --git a/utests/CMakeLists.txt b/utests/CMakeLists.txt
index 5cd20c3f..4cdb449a 100644
--- a/utests/CMakeLists.txt
+++ b/utests/CMakeLists.txt
@@ -107,6 +107,7 @@ set (utests_sources
builtin_nextafter.cpp
builtin_remquo.cpp
builtin_shuffle.cpp
+ builtin_shuffle2.cpp
builtin_sign.cpp
buildin_work_dim.cpp
builtin_global_size.cpp
diff --git a/utests/builtin_shuffle2.cpp b/utests/builtin_shuffle2.cpp
new file mode 100644
index 00000000..7a9ebd15
--- /dev/null
+++ b/utests/builtin_shuffle2.cpp
@@ -0,0 +1,45 @@
+#include "utest_helper.hpp"
+
+void builtin_shuffle2(void)
+{
+ const int n = 32;
+
+ // Setup kernel and buffers
+ OCL_CREATE_KERNEL("builtin_shuffle2");
+ OCL_CREATE_BUFFER(buf[0], 0, n * sizeof(float), NULL);
+ OCL_CREATE_BUFFER(buf[1], 0, n * sizeof(float), NULL);
+ OCL_CREATE_BUFFER(buf[2], 0, n * sizeof(float), NULL);
+ OCL_CREATE_BUFFER(buf[3], 0, n * sizeof(float), NULL);
+ OCL_SET_ARG(0, sizeof(cl_mem), &buf[0]);
+ OCL_SET_ARG(1, sizeof(cl_mem), &buf[1]);
+ OCL_SET_ARG(2, sizeof(cl_mem), &buf[2]);
+ OCL_SET_ARG(3, sizeof(cl_mem), &buf[3]);
+ globals[0] = n;
+ locals[0] = 16;
+
+ OCL_MAP_BUFFER(0);
+ OCL_MAP_BUFFER(1);
+ for (int i = 0; i < n; i ++) {
+ ((float *)(buf_data[0]))[i] = (rand() & 15) * 0.1f;
+ ((float *)(buf_data[1]))[i] = (rand() & 15) * 0.1f;
+ }
+ OCL_UNMAP_BUFFER(0);
+ OCL_UNMAP_BUFFER(1);
+
+ OCL_NDRANGE(1);
+
+ OCL_MAP_BUFFER(0);
+ OCL_MAP_BUFFER(1);
+ OCL_MAP_BUFFER(2);
+ OCL_MAP_BUFFER(3);
+ for (int i = 0; i < n; i ++) {
+ OCL_ASSERT(2 * ((float *)(buf_data[0]))[i] == ((float *)(buf_data[3]))[i]);
+ OCL_ASSERT(2 * ((float *)(buf_data[1]))[i] == ((float *)(buf_data[2]))[i]);
+ }
+ OCL_UNMAP_BUFFER(0);
+ OCL_UNMAP_BUFFER(1);
+ OCL_UNMAP_BUFFER(2);
+ OCL_UNMAP_BUFFER(3);
+}
+
+MAKE_UTEST_FROM_FUNCTION(builtin_shuffle2);