Use row/col layout

src/simplify_algebra.cpp

@@ -326,7 +326,7 @@ struct find_double_add_lit_broadcast
 
 struct find_inner_broadcast
 {
-    auto matcher() const { return pointwise(match::all_of[match::inputs()](match::broadcast())); }
+    auto matcher() const { return pointwise(match::all_of[match::inputs()](match::broadcast()))(match::none_of(match::name("layout"))); }
 
     void apply(module& m, const match::matcher_result& r) const
     {

src/targets/gpu/CMakeLists.txt

@@ -108,6 +108,7 @@ add_library(migraphx_gpu
     perfdb.cpp
     pooling.cpp
     reverse.cpp
+    rewrite_ops.cpp
     rnn_variable_seq_lens.cpp
     rocblas.cpp
     scatter.cpp

src/targets/gpu/target.cpp

@@ -43,6 +43,7 @@
 #include <migraphx/register_target.hpp>
 #include <migraphx/replace_allocate.hpp>
 #include <migraphx/rewrite_gelu.hpp>
+#include <migraphx/gpu/rewrite_ops.hpp>
 #include <migraphx/rewrite_pooling.hpp>
 #include <migraphx/rewrite_quantization.hpp>
 #include <migraphx/rewrite_rnn.hpp>
@@ -129,6 +130,8 @@ std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_opti
         simplify_algebra{},
         prefuse_ops{},
         dead_code_elimination{},
+        rewrite_ops{},
+        dead_code_elimination{},
         auto_contiguous{},
         simplify_reshapes{},
         propagate_constant{},