diff --git a/src/tools/Emulator.cpp b/src/tools/Emulator.cpp
index aed720b0..0e3ded82 100644
--- a/src/tools/Emulator.cpp
+++ b/src/tools/Emulator.cpp
@@ -2844,8 +2844,7 @@ EmulationResult tools::emulate(const EmulationData& data)
     if(!data.memoryDump.empty())
         dumpMemory(mem, data.memoryDump, uniformAddress, false);
 
-    EmulationResult result{data};
-    result.executionSuccessful = status;
+    EmulationResult result{data, status, {}};
 
     result.results.reserve(data.parameter.size());
     for(std::size_t i = 0; i < data.parameter.size(); ++i)
diff --git a/testing/boost-compute/test_function.cl b/testing/boost-compute/test_function.cl
index f4423fee..730baff7 100644
--- a/testing/boost-compute/test_function.cl
+++ b/testing/boost-compute/test_function.cl
@@ -30,3 +30,25 @@ for(uint i = 0; i < n; i++){
 
 }
 
+inline bool compare_second(_pair_int_float_t a, _pair_int_float_t b){ return a.second < b.second; }
+
+
+__kernel void serial_insertion_sort_second(__local _pair_int_float_t* data, uint n, __global _pair_int_float_t* _buf0)
+{
+for(uint i = 0; i < n; i++){
+    data[i] = _buf0[i];
+}
+for(uint i = 1; i < n; i++){
+    const _pair_int_float_t value = data[i];
+    uint pos = i;
+    while(pos > 0 && compare_second(value, data[pos-1])){
+        data[pos] = data[pos-1];
+        pos--;
+    }
+    data[pos] = value;
+}
+for(uint i = 0; i < n; i++){
+    _buf0[i] = data[i];
+}
+
+}
diff --git a/testing/test_shuffle.cl b/testing/test_shuffle.cl
index 87f35032..a24b582f 100644
--- a/testing/test_shuffle.cl
+++ b/testing/test_shuffle.cl
@@ -292,3 +292,107 @@ __kernel void sample_test_char4(__global char* source, __global char4* dest)
     tmp.s1 = shuffle_fn(source[31]);
     dest[31] = tmp;
 }
+
+__kernel void sample_test_float3(__global float* source, __global float* dest)
+{
+    if(get_global_id(0) != 0)
+        return;
+    // float src1 /*, src2*/;
+    float3 tmp;
+    tmp = (float) ((float) 0);
+    tmp.s1 = source[0];
+    vstore3(tmp, 0, dest);
+    tmp = (float) ((float) 0);
+    tmp.S0 = source[1];
+    vstore3(tmp, 1, dest);
+    tmp = (float) ((float) 0);
+    tmp.S1 = source[2];
+    vstore3(tmp, 2, dest);
+    tmp = (float) ((float) 0);
+    tmp.s2 = source[3];
+    vstore3(tmp, 3, dest);
+    tmp = (float) ((float) 0);
+    tmp.s0 = source[4];
+    vstore3(tmp, 4, dest);
+    tmp = (float) ((float) 0);
+    tmp.S0 = source[5];
+    vstore3(tmp, 5, dest);
+    tmp = (float) ((float) 0);
+    tmp.s2 = source[6];
+    vstore3(tmp, 6, dest);
+    tmp = (float) ((float) 0);
+    tmp.s2 = source[7];
+    vstore3(tmp, 7, dest);
+    tmp = (float) ((float) 0);
+    tmp.S2 = source[8];
+    vstore3(tmp, 8, dest);
+    tmp = (float) ((float) 0);
+    tmp.s2 = source[9];
+    vstore3(tmp, 9, dest);
+    tmp = (float) ((float) 0);
+    tmp.s2 = source[10];
+    vstore3(tmp, 10, dest);
+    tmp = (float) ((float) 0);
+    tmp.S0 = source[11];
+    vstore3(tmp, 11, dest);
+    tmp = (float) ((float) 0);
+    tmp.s0 = source[12];
+    vstore3(tmp, 12, dest);
+    tmp = (float) ((float) 0);
+    tmp.s1 = source[13];
+    vstore3(tmp, 13, dest);
+    tmp = (float) ((float) 0);
+    tmp.S0 = source[14];
+    vstore3(tmp, 14, dest);
+    tmp = (float) ((float) 0);
+    tmp.s1 = source[15];
+    vstore3(tmp, 15, dest);
+    tmp = (float) ((float) 0);
+    tmp.S2 = source[16];
+    vstore3(tmp, 16, dest);
+    tmp = (float) ((float) 0);
+    tmp.S2 = source[17];
+    vstore3(tmp, 17, dest);
+    tmp = (float) ((float) 0);
+    tmp.s2 = source[18];
+    vstore3(tmp, 18, dest);
+    tmp = (float) ((float) 0);
+    tmp.S2 = source[19];
+    vstore3(tmp, 19, dest);
+    tmp = (float) ((float) 0);
+    tmp.s2 = source[20];
+    vstore3(tmp, 20, dest);
+    tmp = (float) ((float) 0);
+    tmp.S2 = source[21];
+    vstore3(tmp, 21, dest);
+    tmp = (float) ((float) 0);
+    tmp.S1 = source[22];
+    vstore3(tmp, 22, dest);
+    tmp = (float) ((float) 0);
+    tmp.s0 = source[23];
+    vstore3(tmp, 23, dest);
+    tmp = (float) ((float) 0);
+    tmp.S2 = source[24];
+    vstore3(tmp, 24, dest);
+    tmp = (float) ((float) 0);
+    tmp.S1 = source[25];
+    vstore3(tmp, 25, dest);
+    tmp = (float) ((float) 0);
+    tmp.s2 = source[26];
+    vstore3(tmp, 26, dest);
+    tmp = (float) ((float) 0);
+    tmp.s1 = source[27];
+    vstore3(tmp, 27, dest);
+    tmp = (float) ((float) 0);
+    tmp.s2 = source[28];
+    vstore3(tmp, 28, dest);
+    tmp = (float) ((float) 0);
+    tmp.S2 = source[29];
+    vstore3(tmp, 29, dest);
+    tmp = (float) ((float) 0);
+    tmp.S2 = source[30];
+    vstore3(tmp, 30, dest);
+    tmp = (float) ((float) 0);
+    tmp.s1 = source[31];
+    vstore3(tmp, 31, dest);
+}