[Beignet] [PATCH 5/5 V2] Add utest for workgroup_broadcast.

[junyan.he at inbox.com]

From: Junyan He <junyan.he at linux.intel.com>

Signed-off-by: Junyan He <junyan.he at linux.intel.com>
---
 kernels/compiler_workgroup_broadcast.cl |    9 ++++++
 utests/CMakeLists.txt                   |    1 +
 utests/compiler_workgroup_broadcast.cpp |   47 +++++++++++++++++++++++++++++++
 3 files changed, 57 insertions(+)
 create mode 100644 kernels/compiler_workgroup_broadcast.cl
 create mode 100644 utests/compiler_workgroup_broadcast.cpp

diff --git a/kernels/compiler_workgroup_broadcast.cl b/kernels/compiler_workgroup_broadcast.cl
new file mode 100644
index 0000000..4df74e3
--- /dev/null
+++ b/kernels/compiler_workgroup_broadcast.cl
@@ -0,0 +1,9 @@
+kernel void compiler_workgroup_broadcast(global uint *src, global uint *dst) {
+    uint val = src[get_group_id(0)*(get_local_size(1) * get_local_size(0))
+	+ get_group_id(1)*(get_local_size(1) * get_local_size(0) * get_num_groups(0))
+	+ get_local_id(1)* get_local_size(0) + get_local_id(0)];
+    uint bv = work_group_broadcast(val, 8, 3);
+    dst[get_group_id(0)*(get_local_size(1) * get_local_size(0))
+	+ get_group_id(1)*(get_local_size(1) * get_local_size(0) * get_num_groups(0))
+	+ get_local_id(1)* get_local_size(0) + get_local_id(0)] = bv;
+}
diff --git a/utests/CMakeLists.txt b/utests/CMakeLists.txt
index b3a051d..b4f8407 100644
--- a/utests/CMakeLists.txt
+++ b/utests/CMakeLists.txt
@@ -119,6 +119,7 @@ set (utests_sources
   compiler_math.cpp
   compiler_atomic_functions.cpp
   compiler_async_copy.cpp
+  compiler_workgroup_broadcast.cpp
   compiler_async_stride_copy.cpp
   compiler_insn_selection_min.cpp
   compiler_insn_selection_max.cpp
diff --git a/utests/compiler_workgroup_broadcast.cpp b/utests/compiler_workgroup_broadcast.cpp
new file mode 100644
index 0000000..d45e5d8
--- /dev/null
+++ b/utests/compiler_workgroup_broadcast.cpp
@@ -0,0 +1,47 @@
+#include <cstdint>
+#include <cstring>
+#include <iostream>
+#include "utest_helper.hpp"
+
+void compiler_workgroup_broadcast(void)
+{
+  const size_t n0 = 32;
+  const size_t n1 = 16;
+  const size_t n = n0 * n1;
+  uint32_t src[n];
+
+  // Setup kernel and buffers
+  OCL_CREATE_KERNEL("compiler_workgroup_broadcast");
+  OCL_CREATE_BUFFER(buf[0], 0, n * sizeof(uint32_t), NULL);
+  OCL_CREATE_BUFFER(buf[1], 0, n * sizeof(uint32_t), NULL);
+  OCL_SET_ARG(0, sizeof(cl_mem), &buf[0]);
+  OCL_SET_ARG(1, sizeof(cl_mem), &buf[1]);
+  globals[0] = n0;
+  globals[1] = n1;
+  locals[0] = 16;
+  locals[1] = 16;
+
+  for (int32_t i = 0; i < (int32_t) n; ++i) {
+    src[i] = i;
+  }
+  OCL_MAP_BUFFER(0);
+  memcpy(buf_data[0], src, sizeof(src));
+  OCL_UNMAP_BUFFER(0);
+
+  // Run the kernel on GPU
+  OCL_NDRANGE(2);
+
+  // Compare
+  OCL_MAP_BUFFER(1);
+  for (int32_t i = 0; i < (int32_t) n/2; ++i) {
+//    printf("%u ", ((uint32_t *)buf_data[1])[i]);
+    OCL_ASSERT(((uint32_t *)buf_data[1])[i] == 56);
+  }
+  for (int32_t i = n/2; i < (int32_t) n; ++i) {
+    //	printf("%u ", ((uint32_t *)buf_data[1])[i]);
+    OCL_ASSERT(((uint32_t *)buf_data[1])[i] == 312);
+  }
+  OCL_UNMAP_BUFFER(1);
+}
+
+MAKE_UTEST_FROM_FUNCTION(compiler_workgroup_broadcast);
-- 
1.7.9.5