Run CK job on jenkins

Jenkinsfile

@@ -89,6 +89,8 @@ def rocmnodename(name) {
         node_name = "${rocmtest_name} && vega";
     } else if(name == "navi21") {
         node_name = "${rocmtest_name} && navi21";
+    } else if(name == "mi100+") {
+        node_name = "${rocmtest_name} && (gfx908 || gfx90a)";
     } else if(name == "anygpu") {
         node_name = "${rocmtest_name} && (gfx908 || gfx90a || vega)";
     } else if(name == "nogpu") {
@@ -120,7 +122,7 @@ rocmtest clang_debug: rocmnode('vega') { cmake_build ->
     }
 }, hiprtc_gpu_debug: rocmnode('vega') { cmake_build ->
     stage('HipRTC GPU Debug') {
-        cmake_build(flags: "-DCMAKE_BUILD_TYPE=release -DMIGRAPHX_USE_HIPRTC=On", gpu_debug: true, hiprtc_workarounds:  true)
+        cmake_build(flags: "-DCMAKE_BUILD_TYPE=release -DMIGRAPHX_USE_HIPRTC=On", gpu_debug: true, hiprtc_workarounds: true)
     }
 }, all_targets_debug : rocmnode('vega') { cmake_build ->
     stage('All targets Release') {
@@ -134,6 +136,12 @@ rocmtest clang_debug: rocmnode('vega') { cmake_build ->
             cmake_build(flags: "-DCMAKE_BUILD_TYPE=debug -DMIGRAPHX_ENABLE_PYTHON=Off -DMIGRAPHX_ENABLE_MLIR=On -DCMAKE_CXX_FLAGS_DEBUG='${debug_flags}' -DCMAKE_C_FLAGS_DEBUG='${debug_flags}'")
         }
     }
+}, ck_release: rocmnode('mi100+') { cmake_build ->
+    stage('CK Release') {
+        withEnv(['MIGRAPHX_ENABLE_CK=1', 'MIGRAPHX_TUNE_CK=1']) {
+            cmake_build(flags: "-DCMAKE_BUILD_TYPE=release")
+        }
+    }
 }, clang_asan: rocmnode('nogpu') { cmake_build ->
     stage('Clang ASAN') {
         def sanitizers = "undefined,address"

src/targets/gpu/compile_ops.cpp

@@ -113,7 +113,7 @@ struct compile_plan
     instruction_ref ins;
     optional<tuning_config> config       = nullopt;
     std::vector<compiled_result> results = {};
-    void update_config() { config = get_tuning_config(*ctx, ins, preop); }
+    void update_config(bool exhaustive) { config = get_tuning_config(*ctx, ins, preop, exhaustive); }
     template <class Vector>
     void add_compiles(Vector& compiles, problem_cache& pc)
     {
@@ -202,9 +202,7 @@ struct compile_manager
 
     void update_configs()
     {
-        if(not exhaustive)
-            return;
-        par_compile(cps.size(), [&](auto i) { cps[i].update_config(); });
+        par_compile(cps.size(), [&](auto i) { cps[i].update_config(exhaustive); });
     }
 
     void compile(module& m)

src/targets/gpu/compiler.cpp

@@ -63,9 +63,9 @@ compile_op(const std::string& name, context& ctx, const std::vector<shape>& inpu
     return compiler_map().at(name).compile_op(ctx, inputs, v);
 }
 
-optional<tuning_config> get_tuning_config(context& ctx, instruction_ref ins, const operation& op)
+optional<tuning_config> get_tuning_config(context& ctx, instruction_ref ins, const operation& op, bool exhaustive)
 {
-    return compiler_map().at(op.name()).get_tuning_config(ctx, ins, op);
+    return compiler_map().at(op.name()).get_tuning_config(ctx, ins, op, exhaustive);
 }
 
 } // namespace gpu

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

@@ -79,7 +79,7 @@ using compiler_compile =
 using compiler_compile_op =
     std::function<operation(context&, const std::vector<shape>& inputs, const value&)>;
 using compiler_tuning_config =
-    std::function<optional<tuning_config>(context&, instruction_ref, const operation&)>;
+    std::function<optional<tuning_config>(context&, instruction_ref, const operation&, bool)>;
 
 void register_compiler(const std::string& name,
                        compiler_compile c,
@@ -91,7 +91,7 @@ compiler_replace
 compile(context& ctx, instruction_ref ins, const operation& op, const value& solution);
 operation
 compile_op(const std::string& name, context& ctx, const std::vector<shape>& inputs, const value& v);
-optional<tuning_config> get_tuning_config(context& ctx, instruction_ref ins, const operation& op);
+optional<tuning_config> get_tuning_config(context& ctx, instruction_ref ins, const operation& op, bool exhaustive);
 
 template <class T>
 void register_compiler()
@@ -125,7 +125,7 @@ template <class Derived>
 struct compiler : auto_register_compiler<Derived>
 {
     const Derived& derived() const { return static_cast<const Derived&>(*this); }
-    optional<tuning_config> get_tuning_config(context&, instruction_ref, const operation&) const
+    optional<tuning_config> get_tuning_config(context&, instruction_ref, const operation&, bool) const
     {
         return nullopt;
     }

src/targets/gpu/jit/ck_gemm.cpp

@@ -50,6 +50,7 @@ MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_LOG_CK_GEMM);
 MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_CK_TUNING);
 MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_CK_TUNING_VALUE);
 MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_CK_DEBUG);
+MIGRAPHX_DECLARE_ENV_VAR(MIGRAPHX_TUNE_CK);
 
 // NOLINTNEXTLINE
 static const char* const ck_gemm_kernel = R"__migraphx__(
@@ -436,8 +437,10 @@ struct ck_gemm_compiler : compiler<ck_gemm_compiler>
     }
 
     optional<tuning_config>
-    get_tuning_config(context& ctx, instruction_ref ins, const operation& op) const
+    get_tuning_config(context& ctx, instruction_ref ins, const operation& op, bool exhaustive) const
     {
+        if (not exhaustive and not enabled(MIGRAPHX_TUNE_CK{}))
+            return nullopt;
         tuning_config tc;
         auto shapes    = to_shapes(ins->inputs());
         auto problem   = create_problem(shapes, create_settings(ins, op));