Use minimum block size of 64 threads (#1427)

Local Threads of multiples 32 were introduced in #1348
But LocalThreads that are not multiple of 64 are causing correctness issues.

src/targets/gpu/compile_hip_code_object.cpp

@@ -144,9 +144,8 @@ compute_global_for(context& ctx, std::size_t n, std::size_t over)
 
 std::size_t compute_block_size(std::size_t n, std::size_t max_block_size)
 {
-    const std::size_t min_block_size  = 64;
-    const std::size_t base_block_size = 32;
-    auto block_size                   = (((n - 1) / base_block_size + 1)) * base_block_size;
+    const std::size_t min_block_size = 64;
+    auto block_size                  = (((n - 1) / min_block_size + 1)) * min_block_size;
     return std::min(std::max(min_block_size, block_size), max_block_size);
 }
 

test/verify/test_reduce_op_large.cpp

@@ -51,7 +51,7 @@ template struct test_reduce_op_large<migraphx::op::reduce_min, 1, migraphx::shap
 template struct test_reduce_op_large<migraphx::op::reduce_prod, 2, migraphx::shape::float_type>;
 template struct test_reduce_op_large<migraphx::op::reduce_sum, 1, migraphx::shape::float_type>;
 
-struct test_reduce_mean : verify_program<test_reduce_mean>
+struct test_reduce_mean_1 : verify_program<test_reduce_mean_1>
 {
     migraphx::program create_program() const
     {
@@ -63,3 +63,16 @@ struct test_reduce_mean : verify_program<test_reduce_mean>
         return p;
     };
 };
+
+struct test_reduce_mean_2 : verify_program<test_reduce_mean_2>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        auto* mm = p.get_main_module();
+        migraphx::shape s{migraphx::shape::float_type, {336, 400}};
+        auto x = mm->add_parameter("x", s);
+        mm->add_instruction(migraphx::op::reduce_mean{{1}}, x);
+        return p;
+    };
+};