halide · aekul · Apr 4, 2023 · Apr 5, 2022 · Jul 28, 2022 · Jul 28, 2022
diff --git a/apps/cuda_mat_mul/mat_mul_generator.cpp b/apps/cuda_mat_mul/mat_mul_generator.cpp
@@ -34,28 +34,38 @@ class MatMul : public Halide::Generator<MatMul> {
         Var xi, yi, xio, xii, yii, xo, yo, x_pair, xiio, ty;
         RVar rxo, rxi;
 
-        out.bound(x, 0, size)
-            .bound(y, 0, size)
-            .tile(x, y, xi, yi, 64, 16)
-            .tile(xi, yi, xii, yii, 4, 8)
-            .gpu_blocks(x, y)
-            .gpu_threads(xi, yi)
-            .unroll(xii)
-            .unroll(yii);
-        prod.compute_at(out, xi)
-            .vectorize(x)
-            .unroll(y)
-            .update()
-            .reorder(x, y, r)
-            .vectorize(x)
-            .unroll(y)
-            .unroll(r, 8);
-        A.in().compute_at(prod, r).vectorize(_0).unroll(_1);
-        B.in().compute_at(prod, r).vectorize(_0).unroll(_1);
+        if (!using_autoscheduler()) {
+            out.bound(x, 0, size)
+                .bound(y, 0, size)
+                .tile(x, y, xi, yi, 64, 16)
+                .tile(xi, yi, xii, yii, 4, 8)
+                .gpu_blocks(x, y)
+                .gpu_threads(xi, yi)
+                .unroll(xii)
+                .unroll(yii);
+            prod.compute_at(out, xi)
+                .vectorize(x)
+                .unroll(y)
+                .update()
+                .reorder(x, y, r)
+                .vectorize(x)
+                .unroll(y)
+                .unroll(r, 8);
+            A.in().compute_at(prod, r).vectorize(_0).unroll(_1);
+            B.in().compute_at(prod, r).vectorize(_0).unroll(_1);
 
-        set_alignment_and_bounds(A, size);
-        set_alignment_and_bounds(B, size);
-        set_alignment_and_bounds(out, size);
+            set_alignment_and_bounds(A, size);
+            set_alignment_and_bounds(B, size);
+            set_alignment_and_bounds(out, size);
+        } else {
+            A.dim(0).set_estimate(0, size).dim(1).set_estimate(0, size);
+            B.dim(0).set_estimate(0, size).dim(1).set_estimate(0, size);
+        }
+
+        // Always specify bounds for outputs, whether autoscheduled or not
+        out
+            .bound(x, 0, size)
+            .bound(y, 0, size);
     }
 };
 

diff --git a/apps/images/low_res_in.png b/apps/images/low_res_in.png
diff --git a/apps/images/matrix_3200.mat b/apps/images/matrix_3200.mat
diff --git a/apps/images/matrix_7000.mat b/apps/images/matrix_7000.mat
diff --git a/src/Generator.h b/src/Generator.h
@@ -387,7 +387,9 @@ template<typename First, typename... Rest>
 struct select_type : std::conditional<First::value, typename First::type, typename select_type<Rest...>::type> {};
 
 template<typename First>
-struct select_type<First> { using type = typename std::conditional<First::value, typename First::type, void>::type; };
+struct select_type<First> {
+    using type = typename std::conditional<First::value, typename First::type, void>::type;
+};
 
 class GeneratorParamInfo;
 
@@ -2155,7 +2157,9 @@ class GeneratorInput_Arithmetic : public GeneratorInput_Scalar<T> {
 };
 
 template<typename>
-struct type_sink { typedef void type; };
+struct type_sink {
+    typedef void type;
+};
 
 template<typename T2, typename = void>
 struct has_static_halide_type_method : std::false_type {};
@@ -3770,7 +3774,9 @@ class Generator : public Internal::GeneratorBase {
     // std::is_member_function_pointer will fail if there is no member of that name,
     // so we use a little SFINAE to detect if there are method-shaped members.
     template<typename>
-    struct type_sink { typedef void type; };
+    struct type_sink {
+        typedef void type;
+    };
 
     template<typename T2, typename = void>
     struct has_configure_method : std::false_type {};

diff --git a/src/autoschedulers/CMakeLists.txt b/src/autoschedulers/CMakeLists.txt
@@ -27,3 +27,4 @@ add_subdirectory(common)
 add_subdirectory(adams2019)
 add_subdirectory(li2018)
 add_subdirectory(mullapudi2016)
+add_subdirectory(anderson2021)
diff --git a/src/autoschedulers/adams2019/CMakeLists.txt b/src/autoschedulers/adams2019/CMakeLists.txt
@@ -2,6 +2,8 @@
 # Build rules for the Adams2019 autoscheduler library
 ##
 
+set(COMMON_DIR "${Halide_SOURCE_DIR}/src/autoschedulers/common")
+
 # =================================================================
 # weights
 set(WF_CPP baseline.cpp)
@@ -75,9 +77,10 @@ endif ()
 if (WITH_UTILS)
     add_executable(adams2019_retrain_cost_model
                    DefaultCostModel.cpp
-                   Weights.cpp
+                   ${COMMON_DIR}/Weights.cpp
                    retrain_cost_model.cpp
                    $<TARGET_OBJECTS:adams2019_weights_obj>)
+    target_include_directories(adams2019_retrain_cost_model PRIVATE "${Halide_SOURCE_DIR}/src/autoschedulers/adams2019")
     target_link_libraries(adams2019_retrain_cost_model PRIVATE ASLog adams2019_cost_model adams2019_train_cost_model Halide::Halide Halide::Plugin)
 endif ()
 
@@ -95,23 +98,25 @@ add_autoscheduler(
     FunctionDAG.cpp
     LoopNest.cpp
     State.cpp
-    Weights.cpp
+    ${COMMON_DIR}/Weights.cpp
     $<TARGET_OBJECTS:adams2019_weights_obj>
 )
 
+target_include_directories(Halide_Adams2019 PRIVATE "${Halide_SOURCE_DIR}/src/autoschedulers/adams2019")
 target_link_libraries(Halide_Adams2019 PRIVATE ASLog ParamParser adams2019_cost_model adams2019_train_cost_model)
 
 # ====================================================
 # Auto-tuning support utilities.
 # TODO(#4053): implement auto-tuning support in CMake?
 
 if (WITH_UTILS)
-    add_executable(adams2019_featurization_to_sample featurization_to_sample.cpp)
+    add_executable(adams2019_featurization_to_sample ${COMMON_DIR}/featurization_to_sample.cpp)
 
-    add_executable(adams2019_get_host_target get_host_target.cpp)
+    add_executable(adams2019_get_host_target ${COMMON_DIR}/get_host_target.cpp)
     target_link_libraries(adams2019_get_host_target PRIVATE Halide::Halide)
 
-    add_executable(adams2019_weightsdir_to_weightsfile weightsdir_to_weightsfile.cpp Weights.cpp)
+    add_executable(adams2019_weightsdir_to_weightsfile ${COMMON_DIR}/weightsdir_to_weightsfile.cpp ${COMMON_DIR}/Weights.cpp)
+    target_include_directories(adams2019_weightsdir_to_weightsfile PRIVATE "${Halide_SOURCE_DIR}/src/autoschedulers/adams2019" ${COMMON_DIR})
     target_link_libraries(adams2019_weightsdir_to_weightsfile PRIVATE Halide::Runtime)
 endif ()
 
@@ -121,11 +126,12 @@ endif ()
 
 if (WITH_TESTS)
 
-    add_executable(adams2019_test_perfect_hash_map test_perfect_hash_map.cpp)
+    add_executable(adams2019_test_perfect_hash_map ${COMMON_DIR}/test_perfect_hash_map.cpp)
     add_test(NAME adams2019_test_perfect_hash_map COMMAND adams2019_test_perfect_hash_map)
     set_tests_properties(adams2019_test_perfect_hash_map PROPERTIES LABELS "adams2019;autoschedulers;auto_schedule")
 
-    add_executable(adams2019_test_function_dag test_function_dag.cpp FunctionDAG.cpp)
+    add_executable(adams2019_test_function_dag ${COMMON_DIR}/test_function_dag.cpp FunctionDAG.cpp)
+    target_include_directories(adams2019_test_function_dag PRIVATE "${Halide_SOURCE_DIR}/src/autoschedulers/adams2019" ${COMMON_DIR})
     target_link_libraries(adams2019_test_function_dag PRIVATE ASLog Halide::Halide Halide::Tools Halide::Plugin)
     add_test(NAME adams2019_test_function_dag COMMAND adams2019_test_function_dag)
     set_tests_properties(adams2019_test_function_dag PROPERTIES LABELS "adams2019;autoschedulers;auto_schedule")

diff --git a/src/autoschedulers/adams2019/Makefile b/src/autoschedulers/adams2019/Makefile
@@ -68,8 +68,8 @@ $(BIN)/libautoschedule_adams2019.$(PLUGIN_EXT): \
 				$(SRC)/Cache.cpp \
 				$(SRC)/DefaultCostModel.h \
 				$(SRC)/DefaultCostModel.cpp \
-				$(SRC)/Weights.h \
-				$(SRC)/Weights.cpp \
+				$(COMMON_DIR)/Weights.h \
+				$(COMMON_DIR)/Weights.cpp \
 				$(SRC)/FunctionDAG.h \
 				$(SRC)/FunctionDAG.cpp \
 				$(SRC)/LoopNest.h \
@@ -79,38 +79,38 @@ $(BIN)/libautoschedule_adams2019.$(PLUGIN_EXT): \
 				$(SRC)/State.h \
 				$(SRC)/State.cpp \
 				$(SRC)/Timer.h \
-				$(SRC)/PerfectHashMap.h \
+				$(COMMON_DIR)/PerfectHashMap.h \
 				$(AUTOSCHED_WEIGHT_OBJECTS) \
 				$(AUTOSCHED_COST_MODEL_LIBS) \
 				$(BIN)/auto_schedule_runtime.a \
 				| $(LIB_HALIDE)
 	@mkdir -p $(@D)
-	$(CXX) -shared $(USE_EXPORT_DYNAMIC) -fPIC -fvisibility=hidden -fvisibility-inlines-hidden $(CXXFLAGS) $(OPTIMIZE) -I $(BIN)/cost_model $(filter-out %.h $(LIBHALIDE_LDFLAGS),$^) -o $@ $(HALIDE_SYSTEM_LIBS) $(HALIDE_RPATH_FOR_LIB)
+	$(CXX) -shared $(USE_EXPORT_DYNAMIC) -fPIC -fvisibility=hidden -fvisibility-inlines-hidden $(CXXFLAGS) $(OPTIMIZE) -I $(BIN)/cost_model $(filter-out %.h $(LIBHALIDE_LDFLAGS),$^) -o $@ $(HALIDE_SYSTEM_LIBS) $(HALIDE_RPATH_FOR_LIB) -I $(SRC)
 
 $(BIN)/retrain_cost_model: $(SRC)/retrain_cost_model.cpp \
 				$(COMMON_DIR)/ASLog.cpp \
 				$(SRC)/DefaultCostModel.h \
 				$(SRC)/DefaultCostModel.cpp \
-				$(SRC)/Weights.h \
-				$(SRC)/Weights.cpp \
+				$(COMMON_DIR)/Weights.h \
+				$(COMMON_DIR)/Weights.cpp \
 				$(SRC)/CostModel.h \
 				$(SRC)/NetworkSize.h \
 				$(AUTOSCHED_COST_MODEL_LIBS) \
 				$(AUTOSCHED_WEIGHT_OBJECTS) \
 				$(BIN)/auto_schedule_runtime.a
 	@mkdir -p $(@D)
-	$(CXX) $(CXXFLAGS) -frtti -Wall -I ../support -I $(BIN)/cost_model $(OPTIMIZE) $(filter-out %.h,$^) -o $@ $(LIBHALIDE_LDFLAGS) $(USE_OPEN_MP) $(HALIDE_RPATH_FOR_BIN)
+	$(CXX) $(CXXFLAGS) -frtti -Wall -I ../support -I $(BIN)/cost_model $(OPTIMIZE) $(filter-out %.h,$^) -o $@ $(LIBHALIDE_LDFLAGS) $(USE_OPEN_MP) $(HALIDE_RPATH_FOR_BIN) -I $(SRC)
 
-$(BIN)/featurization_to_sample: $(SRC)/featurization_to_sample.cpp
+$(BIN)/featurization_to_sample: $(COMMON_DIR)/featurization_to_sample.cpp
 	@mkdir -p $(@D)
 	$(CXX) $(CXXFLAGS) $< $(OPTIMIZE) -o $@ 
 
-$(BIN)/get_host_target: $(SRC)/get_host_target.cpp $(LIB_HALIDE) $(HALIDE_DISTRIB_PATH)/include/Halide.h
+$(BIN)/get_host_target: $(COMMON_DIR)/get_host_target.cpp $(LIB_HALIDE) $(HALIDE_DISTRIB_PATH)/include/Halide.h
 	@mkdir -p $(@D)
 	$(CXX) $(CXXFLAGS) $(filter %.cpp,$^) $(LIBHALIDE_LDFLAGS) $(OPTIMIZE) -o $@ $(HALIDE_RPATH_FOR_BIN)
-$(BIN)/weightsdir_to_weightsfile: $(SRC)/weightsdir_to_weightsfile.cpp $(SRC)/Weights.cpp
+$(BIN)/weightsdir_to_weightsfile: $(COMMON_DIR)/weightsdir_to_weightsfile.cpp $(COMMON_DIR)/Weights.cpp
 	@mkdir -p $(@D)
-	$(CXX) $(CXXFLAGS) $^ $(OPTIMIZE) -o $@
+	$(CXX) $(CXXFLAGS) $^ $(OPTIMIZE) -o $@ -I $(SRC)
 
 .PHONY: clean