Add 120 star tas benchmark

Makefile

@@ -5,14 +5,43 @@ include util.mk
 # Default target
 default: all
 
+# Preprocessor definitions
+DEFINES :=
+
 # Use Libdragon IPL3
 # WARNING: This CAN and WILL break certain (most) emulators.
 # Use if you care about console or ares boot times.
 
-LIBDRAGON_IPL3 := 0
+LIBDRAGON_IPL3 ?= 0
 
-# Preprocessor definitions
-DEFINES :=
+# Build types
+# debug   - Debug build
+# general - General release build
+# final   - Final release build with no crash screen, game will reset upon exception.
+
+RELEASE ?= debug
+
+# Benchmark build
+# Builds a benchmark build that runs a 120 star TAS for automated verification.
+# This setting is also in include/cfg/benchmark.h, but provided here as well for automation purposes.
+
+BENCHMARK ?= 0
+
+ifeq ($(BENCHMARK), 1)
+	DEFINES += CFG_BENCHMARK=1
+  RELEASE = general
+endif
+
+ifeq ($(RELEASE), debug)
+  OPT_FLAGS := -Og -ggdb3
+	DEFINES +=  _DEBUG=1
+else ifeq ($(RELEASE), general)
+  OPT_FLAGS := -Os -ggdb3
+else ifeq ($(RELEASE), final)
+  OPT_FLAGS := -Os
+else
+  $(error Invalid build release setting.)
+endif
 
 ifeq ($(LIBDRAGON_IPL3), 1)
   DEFINES += LIBDRAGON_IPL3=1
@@ -87,17 +116,9 @@ else ifeq ($(GRUCODE),f3dzex) # Fast3DZEX (2.0J / Animal Forest - Dōbutsu no Mo
   DEFINES += F3DZEX_GBI_2=1 F3DEX_GBI_2=1 F3DEX_GBI_SHARED=1
 endif
 
-NON_MATCHING := 1
 MIPSISET     := -mips3
-OPT_FLAGS    := -Os
-
-
-# NON_MATCHING - whether to build a matching, identical copy of the ROM
-#   1 - enable some alternate, more portable code that does not produce a matching ROM
-#   0 - build a matching ROM
-NON_MATCHING ?= 0
-$(eval $(call validate-option,NON_MATCHING,0 1))
 
+NON_MATCHING := 1
 ifeq ($(TARGET_N64),0)
   NON_MATCHING := 1
 endif
@@ -140,6 +161,7 @@ ifeq ($(filter clean distclean,$(MAKECMDGOALS)),)
   else
     $(info IPL:            Nintendo IPL3)
   endif
+  $(info Build type:     $(RELEASE))
   $(info =======================)
 endif
 
@@ -306,7 +328,7 @@ endif
 
 # C compiler options
 CFLAGS = -G 0 $(TARGET_CFLAGS) $(DEF_INC_CFLAGS) $(foreach i,$(INCLUDE_DIRS),--embed-dir=$(i))
-CFLAGS += -std=gnu23 -fno-inline-functions -Wno-unused-variable -mno-shared -march=vr4300 -mfix4300 -mabi=32 -mhard-float -mdivide-breaks -fno-stack-protector -fno-common -fno-zero-initialized-in-bss -fno-PIC -mno-abicalls -fno-strict-aliasing -ffreestanding -fwrapv -Wall -Wextra
+CFLAGS += -std=gnu23 -fno-inline -Wno-unused-variable -mno-shared -march=vr4300 -mfix4300 -mabi=32 -mhard-float -mdivide-breaks -fno-unsafe-math-optimizations -fno-stack-protector -fno-common -fno-zero-initialized-in-bss -fno-PIC -mno-abicalls -fno-strict-aliasing -ffreestanding -fwrapv -Wall -Wextra
 CFLAGS += -Wno-missing-braces -Wno-maybe-uninitialized
 
 ASFLAGS     := -march=vr4300 -mabi=32 $(foreach i,$(INCLUDE_DIRS),-I$(i)) $(foreach d,$(DEFINES),--defsym $(d))

include/cfg/benchmark.h

@@ -0,0 +1,8 @@
+#pragma once
+
+/*
+ * Enables a basic performance benchmark over the vanilla game by replaying the 2012
+ * 120 star TAS.
+ */
+
+// #define CFG_BENCHMARK

lib/hackerlibultra/Makefile

@@ -41,7 +41,7 @@ CC := $(CROSS)gcc
 
 WARNINGS := -Wall -Wextra -Wno-format-security -Wno-unused-function -Wno-unused-parameter -Wno-unused-variable -Wno-builtin-declaration-mismatch
 WARNINGS += -Wno-int-conversion -Wno-incompatible-pointer-types -Wno-implicit-function-declaration # TODO: Try adjusting code to remove these
-CFLAGS := -std=gnu23 -G 0 -c -nostdinc -march=vr4300 -mfix4300 -mabi=32 -mno-abicalls -mdivide-breaks -fno-PIC -fno-common -ffreestanding -fbuiltin -fno-builtin-sinf -fno-builtin-cosf -funsigned-char $(WARNINGS)
+CFLAGS := -std=gnu23 -G 0 -c -fno-inline -nostdinc -march=vr4300 -mfix4300 -mabi=32 -mno-abicalls -mdivide-breaks -fno-PIC -fno-common -ffreestanding -fbuiltin -fno-builtin-sinf -fno-builtin-cosf -funsigned-char $(WARNINGS)
 CFLAGS += -fno-strict-aliasing # TODO: Try adjusting code to remove this
 ASFLAGS := -w -nostdinc -c -G 0 -march=vr4300 -mabi=32 -mgp32 -mfp32 -DMIPSEB -D_LANGUAGE_ASSEMBLY -D_MIPS_SIM=1 -D_ULTRA64
 CPPFLAGS = -DMODERN_CC -D_MIPS_SZLONG=32 -D__USE_ISOC99 $(GBIDEFINE) $(VERSION_DEFINE) $(DEBUGFLAG)

src/game/benchmark.c

@@ -0,0 +1,48 @@
+#include <ultra64.h>
+
+#include "cfg/benchmark.h"
+
+#include "game_init.h"
+#include "n64-stdio.h"
+
+#ifdef CFG_BENCHMARK
+
+#define REPLAY_FILE "data/benchmark_replay.m64"
+#define REPLAY_HEADER_SIZE 0x400
+#define REPLAY_FRAME_SIZE sizeof(ReplayContPad)
+
+typedef struct {
+	u16     button;
+	s8      stickX;
+	s8      stickY;
+} ReplayContPad;
+
+static alignas(32) u8 replayData[] = {
+    #embed REPLAY_FILE
+};
+
+static u8 *replayDataPtr = &replayData[0] + REPLAY_HEADER_SIZE;
+static u8 *replayDataEnd = &replayData[0] + (sizeof(replayData));
+static u32 benchmarkFrame = 0;
+
+void replay_contpad(OSContPad *pad) {
+    ReplayContPad *input = (ReplayContPad *) replayDataPtr;
+
+    if (replayDataPtr >= replayDataEnd) {
+        return;
+    }
+
+    pad->button =  input->button;
+    pad->stick_x = input->stickX;
+    pad->stick_y = input->stickY;
+
+    replayDataPtr += REPLAY_FRAME_SIZE;
+}
+
+void replay_print_stats() {
+    OSTime finalTime = osGetTime();
+    n64_printf("BENCHMARK END\n");
+    n64_printf("Time to complete: %llu us\n", OS_CYCLES_TO_USEC(finalTime));
+    n64_printf("Average FPS: %.2f", (f32)gGlobalTimer * 1000000.0f / OS_CYCLES_TO_USEC(finalTime));
+}
+#endif

src/game/benchmark.h

@@ -0,0 +1,6 @@
+#pragma once
+
+#define REPLAY_CONTPAD_BITS 0xF0
+
+void replay_contpad(OSContPad *pad);
+void replay_print_stats();

src/game/game_init.c

@@ -1,5 +1,7 @@
 #include <ultra64.h>
 
+#include "cfg/benchmark.h"
+
 #include "sm64.h"
 #include "gfx_dimensions.h"
 #include "audio/external.h"
@@ -20,6 +22,8 @@
 #include "segment_symbols.h"
 #include "rumble_init.h"
 
+#include "benchmark.h"
+
 // First 3 controller slots
 struct Controller gControllers[3];
 
@@ -525,15 +529,23 @@ void run_demo_inputs(void) {
  */
 void read_controller_inputs(void) {
     s32 i;
-
+#ifndef CFG_BENCHMARK
     // If any controllers are plugged in, update the controller information.
     if (gControllerBits) {
         osRecvMesg(&gSIEventMesgQueue, &gMainReceivedMesg, OS_MESG_BLOCK);
         osContGetReadData(&gControllerPads[0]);
-#if ENABLE_RUMBLE
+    #if ENABLE_RUMBLE
         release_rumble_pak_control();
-#endif
+    #endif
     }
+#else
+    osRecvMesg(&gSIEventMesgQueue, &gMainReceivedMesg, OS_MESG_BLOCK);
+    replay_contpad(&gControllerPads[0]);
+    #if ENABLE_RUMBLE
+        release_rumble_pak_control();
+    #endif
+#endif
+
     run_demo_inputs();
 
     for (i = 0; i < 2; i++) {
@@ -579,10 +591,15 @@ void init_controllers(void) {
 
     // Set controller 1 to point to the set of status/pads for input 1 and
     // init the controllers.
+#ifndef CFG_BENCHMARK
     gControllers[0].statusData = &gControllerStatuses[0];
     gControllers[0].controllerData = &gControllerPads[0];
     osContInit(&gSIEventMesgQueue, &gControllerBits, &gControllerStatuses[0]);
-
+#else
+    gControllers[0].statusData = &gControllerStatuses[0];
+    gControllers[0].controllerData = &gControllerPads[0];
+    gControllerBits = REPLAY_CONTPAD_BITS;
+#endif
     // Strangely enough, the EEPROM probe for save data is done in this function.
     // Save Pak detection?
     gEepromProbe = osEepromProbe(&gSIEventMesgQueue);

src/game/geo_misc.c

@@ -1,5 +1,7 @@
 #include <PR/ultratypes.h>
 
+#include "cfg/benchmark.h"
+
 #include "sm64.h"
 #include "geo_misc.h"
 
@@ -14,6 +16,7 @@
 #include "init/memory.h"
 #include "object_list_processor.h"
 #include "rendering_graph_node.h"
+#include "benchmark.h"
 #include "save_file.h"
 #include "segment2.h"
 
@@ -199,6 +202,12 @@ Gfx *geo_exec_cake_end_screen(s32 callContext, struct GraphNode *node, UNUSED f3
     Gfx *displayList = NULL;
     Gfx *displayListHead = NULL;
 
+#ifdef CFG_BENCHMARK
+    if (callContext == GEO_CONTEXT_AREA_LOAD) {
+        replay_print_stats();
+    }
+#endif
+
     if (callContext == GEO_CONTEXT_RENDER) {
         displayList = alloc_display_list(3 * sizeof(*displayList));
         displayListHead = displayList;