OpenCL 3.1 CTS

[bashbaug]

This PR adds tests for OpenCL 3.1.

[franz]

I've been testing with PoCL, and there is one place that still needs patching:

diff --git a/test_conformance/basic/test_enqueued_local_size.cpp b/test_conformance/basic/test_enqueued_local_size.cpp
index 685982d1..2563fb47 100644
--- a/test_conformance/basic/test_enqueued_local_size.cpp
+++ b/test_conformance/basic/test_enqueued_local_size.cpp
@@ -102,7 +102,7 @@ REGISTER_TEST_VERSION(enqueued_local_size, Version(2, 0))
     test_error(err, "clCreateBuffer failed.");
 
     std::string cl_std = "-cl-std=CL";
-    cl_std += (get_device_cl_version(device) == Version(3, 0)) ? "3.0" : "2.0";
+    cl_std += (get_device_cl_version(device) >= Version(3, 0)) ? "3.0" : "2.0";
     err = create_single_kernel_helper(
         context, &program[0], &kernel[0], 1, &enqueued_local_size_1d_code,
         "test_enqueued_local_size_1d", cl_std.c_str());

Also, I've noticed that the test for event-complete-is-sync-point is missing - is that intentional ?

[bashbaug]

Also, I've noticed that the test for event-complete-is-sync-point is missing - is that intentional ?

Yes, this is intentional, see: KhronosGroup/OpenCL-Docs#1558

[bashbaug]

and there is one place that still needs patching

Thanks! Fixed by ea26359.

[karolherbst]

@bashbaug brought that up in the past that __OPENCL_C_VERSION__ wasn't checked and I confirmed that the CTS actually does not verify it's set to the proper version. So when compiling against -cl-std=3.1, the CTS does not verify that __OPENCL_C_VERSION__ is set to 310 and happily accepts 300.

I also noticed that CL_VERSION_3_1 is not checked as well and has the wrong number inside the OpenCL C docs:

CL_VERSION_3_1
Substitutes the integer 301 reflecting the OpenCL 3.1 version. Requires support for OpenCL C 3.1 or newer.

[bashbaug]

__OPENCL_C_VERSION__ wasn't checked

Yeah, I filed #2670 to track this.

has the wrong number inside the OpenCL C docs

Eek. Yes, this should obviously be 310, not 301. Probably best to file an OpenCL-Docs issue to track this.

[karolherbst]

has the wrong number inside the OpenCL C docs

Eek. Yes, this should obviously be 310, not 301. Probably best to file an OpenCL-Docs issue to track this.

KhronosGroup/OpenCL-Docs#1568