New subgroups - full changes set (#1074)

* Extended subgroups - extended types types

* Extended subgroups - non uniform vote tests

* Extended subgroups - non uniform arithmetic tests

* Extended subgroups - ballot tests

* Extended subgroups - clustered reduce tests

* Extended subgroups - shuffle tests

* Extended subgroups - formating issues

* Extended subgroups - review fixes

* Extended subgroups - review fixes

Fixed: removed additional brakes, kernel_sstr

* Extended subgroups - fix macos build error

* Extended subgroups - review fixes

Fixed: mac os build error

* Extended subgroups - data type verification example

* Extended subgroups - error unification

* Extended subgroups - fix header years

* Extended subgroups - use is_half_nan

* Extended subgroups - compare half as float

* Review fixes mostly for ballot functions.

- Modify kernels for better handling active/inactive workitems
- Modify gen/chk functions for handling non uniform workgroup sizes
- Introduce new variables naming convention
- minor fixes

* Extended subgroups - simplification data generation for ballot lsb/msb functions

* Extended subgroups - minor fixes

* Extended subgroups - move common code to function

* Extended subgroups - formatting errors fix

* Extended subgroups - fix build error

* Extended subgroups - sub_group_elect more sophisticated

Define mask which is 4bytes pattern where bit 1 means work item is active.
If workitem in subgroup matches pattern then run sub_group_elect()

* Extended subgroups - fix Ubuntu build error

* Extended subgroups - voting function review fixes

* adjust all function for using masks
* remove calculate templates
* merge code to one common template
* check results only in active workitems
* normalize values on host side
* minor fixes

* Extended subgroups - fix typos

* Set of fixes and improvements after review

* define WorkGroupParams to stop extended parameters list in function
* better workitems mask handing (WorkGroupParams)
* narrow values of data input generation to avoid overflows (arithmetic func)
* implement work item masks for arithmetic functions
* enable half type testing for reduction/scan/broadcast
* minor fixes

* Extended subgroups - fix Linux issues

* Extended subgroups - fix sub_group_local_id data type

* Extended subgroups - use vector instead of array.

* Extended subgroups - change names to subgroup

* Extended subgroups - uncomment code, fix build

* Extended subgroups - build fix, use cl_half_from_float func

* Extended subgroups - remove is_half_nan

* Extended subgroups - do no use undef min/max

* Extended subgroups - use parenthesis, fix formatting

test_conformance/subgroups/test_ifp.cpp

@@ -46,7 +46,7 @@ static const char *ifp_source =
     "#define INST_COUNT 0x3\n"
     "\n"
     "__kernel void\n"
-    "test_ifp(const __global int *in, __global int2 *xy, __global int *out)\n"
+    "test_ifp(const __global int *in, __global int4 *xy, __global int *out)\n"
     "{\n"
     "    __local atomic_int loc[NUM_LOC];\n"
     "\n"
@@ -225,10 +225,15 @@ void run_insts(cl_int *x, cl_int *p, int n)
 
 struct IFP
 {
-    static void gen(cl_int *x, cl_int *t, cl_int *, int ns, int nw, int ng)
+    static void gen(cl_int *x, cl_int *t, cl_int *,
+                    const WorkGroupParams &test_params)
     {
         int k;
+        int nw = test_params.local_workgroup_size;
+        int ns = test_params.subgroup_size;
+        int ng = test_params.global_workgroup_size;
         int nj = (nw + ns - 1) / ns;
+        ng = ng / nw;
 
         // We need at least 2 sub groups per group for this test
         if (nj == 1) return;
@@ -240,11 +245,15 @@ struct IFP
         }
     }
 
-    static int chk(cl_int *x, cl_int *y, cl_int *t, cl_int *, cl_int *, int ns,
-                   int nw, int ng)
+    static int chk(cl_int *x, cl_int *y, cl_int *t, cl_int *, cl_int *,
+                   const WorkGroupParams &test_params)
     {
         int i, k;
+        int nw = test_params.local_workgroup_size;
+        int ns = test_params.subgroup_size;
+        int ng = test_params.global_workgroup_size;
         int nj = (nw + ns - 1) / ns;
+        ng = ng / nw;
 
         // We need at least 2 sub groups per group for this tes
         if (nj == 1) return 0;
@@ -275,14 +284,17 @@ struct IFP
 int test_ifp(cl_device_id device, cl_context context, cl_command_queue queue,
              int num_elements, bool useCoreSubgroups)
 {
-    int error;
+    int error = TEST_PASS;
 
+    // Global/local work group sizes
     // Adjust these individually below if desired/needed
-#define G 2000
-#define L 200
-    error = test<cl_int, IFP, G, L>::run(device, context, queue, num_elements,
-                                         "test_ifp", ifp_source, NUM_LOC + 1,
-                                         useCoreSubgroups);
+    constexpr size_t global_work_size = 2000;
+    constexpr size_t local_work_size = 200;
+    WorkGroupParams test_params(global_work_size, local_work_size);
+    test_params.use_core_subgroups = useCoreSubgroups;
+    test_params.dynsc = NUM_LOC + 1;
+    error = test<cl_int, IFP>::run(device, context, queue, num_elements,
+                                   "test_ifp", ifp_source, test_params);
     return error;
 }