Add fuse_mlir pass

src/targets/gpu/CMakeLists.txt

@@ -140,6 +140,7 @@ add_library(migraphx_gpu
     device_name.cpp
     eliminate_workspace.cpp
     elu.cpp
+    fuse_mlir.cpp
     fuse_ops.cpp
     gather.cpp
     gemm_impl.cpp

src/targets/gpu/include/migraphx/gpu/fuse_mlir.hpp

+#ifndef MIGRAPHX_GUARD_GPU_FUSE_MLIR_HPP
+#define MIGRAPHX_GUARD_GPU_FUSE_MLIR_HPP
+
+#include <migraphx/config.hpp>
+#include <migraphx/gpu/context.hpp>
+
+namespace migraphx {
+inline namespace MIGRAPHX_INLINE_NS {
+
+struct module;
+
+namespace gpu {
+
+struct fuse_mlir
+{
+    context* ctx   = nullptr;
+    std::string name() const { return "gpu::fuse_mlir"; }
+    void apply(module& m) const;
+};
+
+} // namespace gpu
+
+} // namespace MIGRAPHX_INLINE_NS
+} // namespace migraphx
+#endif // MIGRAPHX_GUARD_GPU_FUSE_MLIR_HPP

src/targets/gpu/mlir.cpp

@@ -509,8 +509,11 @@ std::string dump_mlir(const module& m)
 
 code_object_op compile_mlir(const module& m)
 {
+    std::cout << m << std::endl;
     mlir_program mp;
     mp.parse(m);
+    auto mod_op = mlirModuleGetOperation(mp.mmodule.get());
+    std::cout << mlir_print(&mlirOperationPrint, mod_op) << std::endl;
     return mp.compile();
 }
 

src/targets/gpu/target.cpp

@@ -30,6 +30,7 @@
 #include <migraphx/gpu/concat_gpu_opt.hpp>
 #include <migraphx/gpu/context.hpp>
 #include <migraphx/gpu/eliminate_workspace.hpp>
+#include <migraphx/gpu/fuse_mlir.hpp>
 #include <migraphx/gpu/fuse_ops.hpp>
 #include <migraphx/gpu/lowering.hpp>
 #include <migraphx/gpu/mlir_conv.hpp>
@@ -102,7 +103,8 @@ std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_opti
         dead_code_elimination{},
         fuse_pointwise{},
         dead_code_elimination{},
-        mlir_conv{&ctx},
+        fuse_mlir{&ctx},
+        dead_code_elimination{},
         lowering{&ctx, options.offload_copy},
         eliminate_contiguous{"gpu::contiguous"},
         dead_code_elimination{},
@@ -112,8 +114,8 @@ std::vector<pass> target::get_passes(migraphx::context& gctx, const compile_opti
         dead_code_elimination{},
         adjust_allocation{gpu_allocation_model{}},
         dead_code_elimination{},
-        fuse_ops{&ctx, options.fast_math},
-        dead_code_elimination{},
+        // fuse_ops{&ctx, options.fast_math},
+        // dead_code_elimination{},
         compile_ops{&ctx},
         dead_code_elimination{},
         write_literals{&ctx},