ahmed/OpenCL-CTS
1
0
Fork 0
mirror of https://github.com/KhronosGroup/OpenCL-CTS.git synced 2026-03-19 06:09:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

New subgroups - full changes set (#1074)

Browse Source 
* 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
This commit is contained in:
Grzegorz Wawiorko
2021-04-06 18:25:48 +02:00
committed by GitHub
parent c5e4ca6c91
commit 71bef8563e
18 changed files with 5075 additions and 957 deletions
Download Patch File Download Diff File Expand all files Collapse all files

46
test_conformance/subgroups/test_barrier.cpp
View File

@@ -59,10 +59,15 @@ static const char *gbar_source =
// barrier test functions
template <int Which> struct BAR
{
static void gen(cl_int *x, cl_int *t, cl_int *m, int ns, int nw, int ng)
static void gen(cl_int *x, cl_int *t, cl_int *m,
const WorkGroupParams &test_params)
{
int i, ii, j, k, n;
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;
int e;
ii = 0;
@@ -79,8 +84,7 @@ template <int Which> struct BAR
// Now map into work group using map from device
for (j = 0; j < nw; ++j)
{
i = m[2 * j + 1] * ns + m[2 * j];
x[j] = t[i];
x[j] = t[j];
}
x += nw;
@@ -89,10 +93,14 @@ template <int Which> struct BAR
}
static int chk(cl_int *x, cl_int *y, cl_int *mx, cl_int *my, cl_int *m,
int ns, int nw, int ng)
const WorkGroupParams &test_params)
{
int ii, i, j, k, n;
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;
cl_int tr, rr;
if (Which == 0)
@@ -105,9 +113,8 @@ template <int Which> struct BAR
// Map to array indexed to array indexed by local ID and sub group
for (j = 0; j < nw; ++j)
{
i = m[2 * j + 1] * ns + m[2 * j];
mx[i] = x[j];
my[i] = y[j];
mx[j] = x[j];
my[j] = y[j];
}
for (j = 0; j < nj; ++j)
@@ -123,8 +130,9 @@ template <int Which> struct BAR
if (tr != rr)
{
log_error("ERROR: sub_group_barrier mismatch for local "
"id %d in sub group %d in group %d\n",
i, j, k);
"id %d in sub group %d in group %d expected "
"%d got %d\n",
i, j, k, tr, rr);
return -1;
}
}
@@ -144,18 +152,18 @@ int test_barrier_functions(cl_device_id device, cl_context context,
cl_command_queue queue, int num_elements,
bool useCoreSubgroups)
{
int error;
int error = TEST_PASS;
// Adjust these individually below if desired/needed
#define G 2000
#define L 200
error = test<cl_int, BAR<0>, G, L>::run(device, context, queue,
num_elements, "test_lbar",
lbar_source, 0, useCoreSubgroups);
error = test<cl_int, BAR<1>, G, L, G>::run(
device, context, queue, num_elements, "test_gbar", gbar_source, 0,
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;
error = test<cl_int, BAR<0>>::run(device, context, queue, num_elements,
"test_lbar", lbar_source, test_params);
error |= test<cl_int, BAR<1>, global_work_size>::run(
device, context, queue, num_elements, "test_gbar", gbar_source,
test_params);
return error;
}