Add flag to enable cpu backend (#680)

* Add flag to enable cpu backend

* Make buffers shared

* Enable optimizations

* Formatting

* Enable cpu backend for gcc builds

CMakeLists.txt

@@ -43,6 +43,9 @@ else()
     set(MIGRAPHX_ENABLE_GPU Off CACHE BOOL "")
 endif()
 
+# Disable cpu backend by default
+set(MIGRAPHX_ENABLE_CPU Off CACHE BOOL "")
+
 set(CMAKE_CXX_STANDARD_DEFAULT "")
 add_compile_options(-std=c++14)
 

Jenkinsfile

@@ -156,7 +156,7 @@ rocmtest tidy: rocmnode('rocmtest') { cmake_build ->
         // TODO: Add bounds-strict
         def sanitizers = "undefined,address"
         def debug_flags = "-g -fprofile-arcs -ftest-coverage -fno-omit-frame-pointer -fsanitize-address-use-after-scope -fsanitize=${sanitizers} -fno-sanitize-recover=${sanitizers}"
-        cmake_build("g++-7", "-DCMAKE_BUILD_TYPE=debug -DMIGRAPHX_ENABLE_PYTHON=Off ${cmake_linker_flags} -DCMAKE_CXX_FLAGS_DEBUG='${debug_flags}'")
+        cmake_build("g++-7", "-DCMAKE_BUILD_TYPE=debug -DMIGRAPHX_ENABLE_CPU=On -DMIGRAPHX_ENABLE_PYTHON=Off ${cmake_linker_flags} -DCMAKE_CXX_FLAGS_DEBUG='${debug_flags}'")
 
     }
     stage('Codecov') {

src/CMakeLists.txt

@@ -177,6 +177,11 @@ add_subdirectory(tf)
 
 add_subdirectory(py)
 add_subdirectory(targets/ref)
+if(MIGRAPHX_ENABLE_CPU)
+add_subdirectory(targets/cpu)
+target_link_libraries(migraphx_all_targets INTERFACE migraphx_cpu)
+target_compile_definitions(migraphx_all_targets INTERFACE -DHAVE_CPU)
+endif()
 if(MIGRAPHX_ENABLE_GPU)
 list(APPEND PACKAGE_DEPENDS PACKAGE MIOpen PACKAGE rocblas)
 add_subdirectory(targets/gpu)

src/include/migraphx/argument.hpp

@@ -23,9 +23,8 @@ struct argument : raw_data<argument>
 
     argument(const shape& s) : m_shape(s)
     {
-        std::vector<char> buffer(s.bytes());
-        // TODO: Move vector
-        data = [=]() mutable { return buffer.data(); };
+        auto buffer = std::make_shared<std::vector<char>>(s.bytes());
+        data        = [=]() mutable { return buffer->data(); };
     }
 
     template <class F, MIGRAPHX_REQUIRES(std::is_pointer<decltype(std::declval<F>()())>{})>

src/targets/cpu/gemm.cpp

@@ -39,10 +39,10 @@ struct is_fast_gemm_type : std::false_type
 {
 };
 
-template <>
-struct is_fast_gemm_type<float> : std::true_type
-{
-};
+// template <>
+// struct is_fast_gemm_type<float> : std::true_type
+// {
+// };
 
 template <class T, class F>
 void migemm_impl(

src/targets/cpu/target.cpp

 
+#include <migraphx/auto_contiguous.hpp>
+#include <migraphx/check_context.hpp>
+#include <migraphx/dead_code_elimination.hpp>
+#include <migraphx/decompose.hpp>
+#include <migraphx/eliminate_allocation.hpp>
+#include <migraphx/eliminate_common_subexpression.hpp>
+#include <migraphx/eliminate_concat.hpp>
+#include <migraphx/eliminate_contiguous.hpp>
+#include <migraphx/eliminate_identity.hpp>
+#include <migraphx/eliminate_pad.hpp>
+#include <migraphx/memory_coloring.hpp>
+#include <migraphx/propagate_constant.hpp>
+#include <migraphx/register_target.hpp>
+#include <migraphx/remap.hpp>
+#include <migraphx/rewrite_batchnorm.hpp>
+#include <migraphx/rewrite_pooling.hpp>
+#include <migraphx/rewrite_rnn.hpp>
+#include <migraphx/schedule.hpp>
+#include <migraphx/simplify_algebra.hpp>
+#include <migraphx/simplify_reshapes.hpp>
 #include <migraphx/cpu/target.hpp>
 #include <migraphx/cpu/lowering.hpp>
-#include <migraphx/register_target.hpp>
 #include <migraphx/pass.hpp>
-#include <migraphx/auto_contiguous.hpp>
-#include <migraphx/rewrite_rnn.hpp>
-#include <migraphx/dead_code_elimination.hpp>
 #include <migraphx/generate.hpp>
 
 namespace migraphx {
@@ -16,9 +32,26 @@ std::string target::name() const { return "cpu"; }
 
 std::vector<pass> target::get_passes(migraphx::context&, const compile_options&) const
 {
-    return {rewrite_rnn{},
+    return {decompose{},
+            dead_code_elimination{},
+            simplify_reshapes{},
+            eliminate_identity{},
+            eliminate_pad{},
+            dead_code_elimination{},
+            rewrite_batchnorm{},
+            dead_code_elimination{},
+            rewrite_rnn{},
+            dead_code_elimination{},
+            rewrite_pooling{},
+            dead_code_elimination{},
+            eliminate_common_subexpression{},
             dead_code_elimination{},
+            simplify_algebra{},
+            simplify_reshapes{},
+            simplify_algebra{},
             auto_contiguous{},
+            simplify_reshapes{},
+            propagate_constant{},
             dead_code_elimination{},
             lowering{},
             dead_code_elimination{}};

src/targets/gpu/target.cpp

-#include <migraphx/gpu/target.hpp>
-#include <migraphx/register_target.hpp>
-#include <migraphx/gpu/lowering.hpp>
-#include <migraphx/memory_coloring.hpp>
-#include <migraphx/gpu/write_literals.hpp>
-#include <migraphx/gpu/context.hpp>
-#include <migraphx/gpu/eliminate_workspace.hpp>
-#include <migraphx/eliminate_allocation.hpp>
-#include <migraphx/gpu/fuse_ops.hpp>
-#include <migraphx/check_context.hpp>
 #include <migraphx/auto_contiguous.hpp>
+#include <migraphx/check_context.hpp>
 #include <migraphx/dead_code_elimination.hpp>
-#include <migraphx/simplify_reshapes.hpp>
-#include <migraphx/simplify_algebra.hpp>
-#include <migraphx/propagate_constant.hpp>
-#include <migraphx/eliminate_contiguous.hpp>
+#include <migraphx/decompose.hpp>
+#include <migraphx/eliminate_allocation.hpp>
 #include <migraphx/eliminate_common_subexpression.hpp>
-#include <migraphx/rewrite_batchnorm.hpp>
-#include <migraphx/rewrite_rnn.hpp>
-#include <migraphx/rewrite_pooling.hpp>
 #include <migraphx/eliminate_concat.hpp>
+#include <migraphx/eliminate_contiguous.hpp>
 #include <migraphx/eliminate_identity.hpp>
-#include <migraphx/gpu/concat_gpu_opt.hpp>
-#include <migraphx/gpu/schedule_model.hpp>
-#include <migraphx/gpu/adjust_allocation.hpp>
-#include <migraphx/gpu/preallocate_param.hpp>
-#include <migraphx/gpu/pack_int8_args.hpp>
-#include <migraphx/gpu/sync_device.hpp>
 #include <migraphx/eliminate_pad.hpp>
-#include <migraphx/decompose.hpp>
+#include <migraphx/memory_coloring.hpp>
+#include <migraphx/propagate_constant.hpp>
+#include <migraphx/register_target.hpp>
 #include <migraphx/remap.hpp>
+#include <migraphx/rewrite_batchnorm.hpp>
+#include <migraphx/rewrite_pooling.hpp>
+#include <migraphx/rewrite_rnn.hpp>
 #include <migraphx/schedule.hpp>
+#include <migraphx/simplify_algebra.hpp>
+#include <migraphx/simplify_reshapes.hpp>
+#include <migraphx/gpu/adjust_allocation.hpp>
+#include <migraphx/gpu/concat_gpu_opt.hpp>
+#include <migraphx/gpu/context.hpp>
+#include <migraphx/gpu/eliminate_workspace.hpp>
+#include <migraphx/gpu/fuse_ops.hpp>
+#include <migraphx/gpu/lowering.hpp>
+#include <migraphx/gpu/pack_int8_args.hpp>
+#include <migraphx/gpu/preallocate_param.hpp>
+#include <migraphx/gpu/schedule_model.hpp>
+#include <migraphx/gpu/sync_device.hpp>
+#include <migraphx/gpu/target.hpp>
+#include <migraphx/gpu/write_literals.hpp>
 
 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {