Fix static_assert in large reduction (#1604)

src/targets/gpu/jit/reduce.cpp

@@ -127,7 +127,7 @@ struct reduce_compiler : compiler<reduce_compiler>
                 vec = vectorize::elements(ctx, faxis, options.virtual_inputs);
             auto relements  = get_reduce_elements(options.virtual_inputs) / vec.size;
             auto block_size = compute_block_size(relements, 256);
-            if(relements > block_size * 256)
+            if(relements >= block_size * 256)
                 algo = "block_large";
             options.set_launch_params(
                 v, compute_global_for(ctx, nelements * block_size, 256), block_size);

test/verify/test_reduce_op_large.cpp

@@ -77,7 +77,7 @@ struct test_reduce_mean_2 : verify_program<test_reduce_mean_2>
     };
 };
 
-struct test_large_reduce_mean : verify_program<test_large_reduce_mean>
+struct test_large_reduce_mean1 : verify_program<test_large_reduce_mean1>
 {
     migraphx::program create_program() const
     {
@@ -89,3 +89,16 @@ struct test_large_reduce_mean : verify_program<test_large_reduce_mean>
         return p;
     };
 };
+
+struct test_large_reduce_mean2 : verify_program<test_large_reduce_mean2>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        auto* mm = p.get_main_module();
+        migraphx::shape s{migraphx::shape::float_type, {1, 32, 262144}};
+        auto x = mm->add_parameter("x", s);
+        mm->add_instruction(migraphx::op::reduce_mean{{2}}, x);
+        return p;
+    };
+};