OpenCL: Self-test with default LWS at device maximum

Also with a GWS of 2x that LWS.  The new figures are better at triggering
bugs. If a kernel needs a lower LWS than device max. our code already
handles that.

We previously had it as LWS=7 GWS=49 for speed (and for checking
heuristics that could bug out on non-log2 values) but that was
introduced before our autotune was sped up with orders of magnitude
and those heuristics has been stable for many years.

Like before, the self-test will obey any give lws/gws options or
environment variables.

Closes #5822

Author: magnumripper

src/opencl_autotune.h

@@ -112,11 +112,9 @@ static void autotune_run_extra(struct fmt_main *self, unsigned int rounds,
 	if (options.flags & FLG_SHOW_CHK)
 		return;
 
-	// FIXME add optional test-same-sizes
-	if (self_test_running) {
-		local_work_size = 7;
-		global_work_size = 49;
-	}
+	/* We could do with 2x get_device_warp_size() but that isn't faster */
+	if (self_test_running)
+		global_work_size = local_work_size = get_device_max_lws(gpu_id);
 
 	ocl_autotune_running = 1;
 

src/opencl_common.c

@@ -1099,22 +1099,20 @@ void opencl_get_user_preferences(const char *format)
 
 void opencl_get_sane_lws_gws_values()
 {
-	if (self_test_running) {
-		local_work_size = 7;
-		global_work_size = 49;
-	}
+	if (self_test_running)
+		global_work_size = local_work_size = get_device_max_lws(gpu_id);
 
 	if (!local_work_size) {
 		if (cpu(device_info[gpu_id]))
 			local_work_size =
 				get_platform_vendor_id(platform_id) == DEV_INTEL ?
 			8 : 1;
 		else
-			local_work_size = 64;
+			local_work_size = 2 * get_device_warp_size(gpu_id);
 	}
 
 	if (!global_work_size)
-		global_work_size = 768;
+		global_work_size = 12 * local_work_size;
 }
 
 char* get_device_name_(int sequential_id)