check for divisor encodable or not, fallback if needed (#906)

* check for divisor encodable or not, fallback if needed

* verify test for retinaface case

src/targets/gpu/device/include/migraphx/gpu/device/nary.hpp

@@ -352,7 +352,8 @@ bool broadcastable(bool& divisible_by_4,
         auto b_len          = result.get_shape().lens()[b_idx];
         auto b_stride       = result.get_shape().strides()[b_idx];
         assert(bshape.lens()[b_idx] == b_len);
-        if(b_len <= max_size and std::none_of(std::next(b_it), strides.end(), not_zero))
+        if(b_len <= max_size and std::none_of(std::next(b_it), strides.end(), not_zero) and
+           is_divisor_encodable(b_stride * b_len))
         {
 
             divisible_by_4 = (b_len % 4 == 0) and (b_stride % 4 == 0) and

test/verify/test_add_broadcast6.cpp

+
+#include "verify_program.hpp"
+#include <migraphx/program.hpp>
+#include <migraphx/generate.hpp>
+#include <migraphx/make_op.hpp>
+
+#include <migraphx/instruction.hpp>
+
+struct test_add_broadcast6 : verify_program<test_add_broadcast6>
+{
+    migraphx::program create_program() const
+    {
+        migraphx::program p;
+        auto* mm = p.get_main_module();
+        auto x   = mm->add_parameter("x", {migraphx::shape::float_type, {1, 64, 568, 1328}});
+        auto y   = mm->add_parameter("y", {migraphx::shape::float_type, {64}});
+        auto by  = mm->add_instruction(
+            migraphx::make_op("broadcast", {{"axis", 1}, {"dims", {1, 64, 568, 1328}}}), y);
+        mm->add_instruction(migraphx::make_op("add"), x, by);
+        return p;
+    }
+};