add experimental matcher

src/targets/gpu/fuse_mlir.cpp

@@ -911,6 +911,46 @@ struct find_mlir_attention_fused_ops : public find_mlir_standalone_attention_op
     }
 };
 
+struct find_mlir_reshape_pointwise
+{
+    mlir_mode conv_mode = mlir_mode::none;
+    mlir_mode dot_mode  = mlir_mode::none;
+    auto matcher() const
+    {
+        auto reshapes = reshaper_names();
+        // Dont want to move ops before slice
+        reshapes.erase("slice");
+        auto rsp_dot_or_conv =
+            match::name(reshapes)(
+                match::any_of[match::inputs()](is_mlir_dot(dot_mode), is_mlir_conv(conv_mode))
+                    .bind("mlir_op"))
+                .bind("rsp");
+        return mlir_pointwise()(match::any_of[match::inputs()](rsp_dot_or_conv));
+    }
+
+    void apply(module_pass_manager& mpm, const match::matcher_result& r) const
+    {
+        auto ins = r.result;
+        auto rsp = r.instructions["rsp"];
+
+        if(not std::all_of(ins->inputs().begin(), ins->inputs().end(), [&](auto i) {
+               return i->get_operator() == rsp->get_operator();
+           }))
+            return;
+
+        std::vector<instruction_ref> new_inps;
+        std::transform(ins->inputs().begin(),
+                       ins->inputs().end(),
+                       std::back_inserter(new_inps),
+                       [&](auto i) { return i->inputs().front(); });
+
+        auto new_pw = mpm.get_module().insert_instruction(
+            ins, ins->get_operator(), new_inps, ins->module_inputs());
+
+        mpm.get_module().replace_instruction(ins, rsp->get_operator(), {new_pw});
+    }
+};
+
 struct find_pointwise_mlir
 {
     auto supported_pointwise() const { return mlir_input_pointwise(match::used_once()); }
@@ -1061,6 +1101,12 @@ void fuse_mlir::apply(module_pass_manager& mpm) const
         return std::max(m1, m2);
     };
 
+    match::find_matches(
+        mpm,
+        find_mlir_reshape_pointwise{.conv_mode = get_mode("fused_convolution", mlir_mode::fast),
+                                    .dot_mode  = get_mode("fused_dot", mlir_mode::fast)});
+    mpm.run_pass(dead_code_elimination{});
+
     // Attention offloads; default disabled
     if(mlir_attention_enabled(ctx) or enable_extra)
     {