diff --git a/Makefile b/Makefile
index 63ed3dd..c736c7a 100644
--- a/Makefile
+++ b/Makefile
@@ -1,12 +1,10 @@
 all: main.elf
 
 include common.mk
+include example/example.mk
 
 MAIN_OBJ = \
-	start.o \
 	main.o \
-	load.o \
-	cache.o \
 	vga.o \
 	rgb.o \
 	holly/background.o \
@@ -16,14 +14,13 @@ MAIN_OBJ = \
 	maple/maple.o \
 	scene.o \
 	macaw.data.o \
-	wink.data.o \
-	$(LIBGCC)
+	wink.data.o
 
-serial.elf: start.o serial_main.o load.o cache.o
-	$(LD) $(LDFLAGS) -T $(LIB)/alt.lds $^ -o $@
+serial.elf: LDSCRIPT = $(LIB)/alt.lds
+serial.elf: $(START_OBJ) serial_main.o load.o
 
-main.elf: $(MAIN_OBJ)
-	$(LD) $(LDFLAGS) -T $(LIB)/main.lds $^ -o $@
+main.elf: LDSCRIPT = $(LIB)/main.lds
+main.elf: $(START_OBJ) $(MAIN_OBJ)
 
-test.elf: $(MAIN_OBJ)
-	$(LD) $(LDFLAGS) -T $(LIB)/alt.lds $^ -o $@
+test.elf: LDSCRIPT = $(LIB)/alt.lds
+test.elf: $(START_OBJ) $(MAIN_OBJ)
diff --git a/aica.h b/aica.hpp
similarity index 100%
rename from aica.h
rename to aica.hpp
diff --git a/align.hpp b/align.hpp
new file mode 100644
index 0000000..fae4dbe
--- /dev/null
+++ b/align.hpp
@@ -0,0 +1,9 @@
+#pragma once
+
+#include <cstdint>
+
+template <typename T>
+inline T * align_32byte(T * mem)
+{
+  return reinterpret_cast<T *>((((reinterpret_cast<uint32_t>(mem) + 31) & ~31)));
+}
diff --git a/cache.cpp b/cache.cpp
index 9f4b401..8dea9e6 100644
--- a/cache.cpp
+++ b/cache.cpp
@@ -1,13 +1,15 @@
-#include "type.h"
-#include "sh7091.h"
-#include "sh7091_bits.h"
+#include "type.hpp"
+#include "sh7091.hpp"
+#include "sh7091_bits.hpp"
 
-#include "cache.h"
+#include "cache.hpp"
 
 extern volatile reg32 sh7091_ic_a[256][(1 << 5) / 4] __asm("sh7091_ic_a");
 extern volatile reg32 sh7091_oc_a[512][(1 << 5) / 4] __asm("sh7091_oc_a");
 
-void cache_init()
+namespace cache {
+
+void init()
 {
   for (int i = 0; i < 256; i++) {
     sh7091_ic_a[i][0] = 0;
@@ -28,3 +30,5 @@ void cache_init()
 
   asm volatile ("nop;nop;nop;nop;nop;nop;nop;nop;");
 }
+
+}
diff --git a/cache.h b/cache.h
deleted file mode 100644
index 91397b9..0000000
--- a/cache.h
+++ /dev/null
@@ -1,3 +0,0 @@
-#pragma once
-
-void cache_init() __attribute__ ((section (".p2ram.cache_init")));
diff --git a/cache.hpp b/cache.hpp
new file mode 100644
index 0000000..192c228
--- /dev/null
+++ b/cache.hpp
@@ -0,0 +1,7 @@
+#pragma once
+
+namespace cache {
+
+void init() __attribute__ ((section (".p2ram.cache_init")));
+
+}
diff --git a/common.mk b/common.mk
index 6db30bc..58f10fd 100644
--- a/common.mk
+++ b/common.mk
@@ -1,5 +1,8 @@
+MAKEFILE_PATH := $(abspath $(lastword $(MAKEFILE_LIST)))
+DIR := $(dir $(MAKEFILE_PATH))
+
 LIB ?= .
-OPT ?= -Og
+OPT ?= -Os
 DEBUG ?= -g -gdwarf-4
 GENERATED ?=
 
@@ -11,6 +14,7 @@ CFLAGS += -falign-functions=4 -ffunction-sections -fdata-sections -fshort-enums
 CFLAGS += -Wall -Werror -Wfatal-errors
 CFLAGS += -Wno-error=narrowing -Wno-error=unused-variable
 CFLAGS += -mfsca -funsafe-math-optimizations
+CFLAGS += -I$(dir $(MAKEFILE_PATH))
 DEPFLAGS = -MMD -E
 # --print-gc-sections
 LDFLAGS = --gc-sections --no-warn-rwx-segment --print-memory-usage --entry=_start --orphan-handling=error
@@ -48,6 +52,11 @@ IP_OBJ = \
 	sg/sg_ini.o \
 	sg/aip.o
 
+START_OBJ = \
+	start.o \
+	runtime.o \
+	cache.o
+
 %.bin.o: %.bin
 	$(BUILD_BINARY_O)
 
@@ -71,6 +80,9 @@ IP_OBJ = \
 %.o: %.cpp %.cpp.d
 	$(CXX) $(CARCH) $(CFLAGS) $(CXXFLAGS) $(OPT) $(DEBUG) -c $< -o $@
 
+%.elf:
+	$(LD) $(LDFLAGS) -T $(LDSCRIPT) $^ -o $@
+
 %.bin: %.elf
 	$(OBJCOPY) -O binary $< $@
 
diff --git a/example/cube.cpp b/example/cube.cpp
new file mode 100644
index 0000000..879df23
--- /dev/null
+++ b/example/cube.cpp
@@ -0,0 +1,12 @@
+#include <cstdint>
+
+#include "vga.hpp"
+#include "memorymap.hpp"
+#include "rgb.hpp"
+
+void main()
+{
+  vga();
+
+  while (1);
+}
diff --git a/example/example.mk b/example/example.mk
new file mode 100644
index 0000000..2434323
--- /dev/null
+++ b/example/example.mk
@@ -0,0 +1,17 @@
+CUBE_OBJ = \
+	example/cube.o \
+	vga.o
+
+example/cube.elf: LDSCRIPT = $(LIB)/alt.lds
+example/cube.elf: $(START_OBJ) $(CUBE_OBJ)
+
+MAPLE_WINK_OBJ = \
+	example/maple_wink.o \
+	vga.o \
+	rgb.o \
+	serial.o \
+	maple/maple.o \
+	wink.data.o
+
+example/maple_wink.elf: LDSCRIPT = $(LIB)/alt.lds
+example/maple_wink.elf: $(START_OBJ) $(MAPLE_WINK_OBJ)
diff --git a/example/maple_wink.cpp b/example/maple_wink.cpp
new file mode 100644
index 0000000..6e105e0
--- /dev/null
+++ b/example/maple_wink.cpp
@@ -0,0 +1,52 @@
+#include <cstdint>
+
+#include "maple/maple.hpp"
+#include "vga.hpp"
+#include "align.hpp"
+#include "serial.hpp"
+
+extern uint32_t _binary_wink_data_start __asm("_binary_wink_data_start");
+
+void make_wink(uint32_t * buf)
+{
+  const uint8_t * src = reinterpret_cast<const uint8_t *>(&_binary_wink_data_start);
+  uint8_t * dst = reinterpret_cast<uint8_t *>(buf);
+
+  uint32_t ix = 0;
+  dst[ix] = 0;
+  for (int i = 0; i < 48 * 32; i++) {
+    dst[ix] |= ((src[i] & 1) << (7 - (i % 8)));
+
+    if (i % 8 == 7) {
+      ix++;
+      dst[ix] = 0;
+    }
+  }
+}
+
+void main()
+{
+  constexpr int width = 48;
+  constexpr int height = 32;
+  constexpr int pixels_per_byte = 8;
+
+  uint32_t wink_buf[width * height / pixels_per_byte];
+  make_wink(wink_buf);
+
+  uint32_t _command_buf[128 / 4];
+  uint32_t _receive_buf[128 / 4];
+  uint32_t * command_buf = align_32byte(_command_buf);
+  uint32_t * receive_buf = align_32byte(_receive_buf);
+
+  maple::init_block_write(command_buf, receive_buf, wink_buf);
+  maple::dma_start(command_buf);
+
+  for (int i = 0; i < 32; i++) {
+    serial::integer<uint32_t>(receive_buf[i]);
+  }
+
+  vga();
+  v_sync_in();
+  vga_fill_framebuffer();
+  while(1);
+}
diff --git a/serial_main.cpp b/example/serial_transfer.cpp
similarity index 83%
rename from serial_main.cpp
rename to example/serial_transfer.cpp
index e3d0d4a..f5e89f9 100644
--- a/serial_main.cpp
+++ b/example/serial_transfer.cpp
@@ -1,8 +1,10 @@
-#include "sh7091.h"
-#include "sh7091_bits.h"
+#include <cstdint>
 
-#include "cache.h"
-#include "load.h"
+#include "sh7091.hpp"
+#include "sh7091_bits.hpp"
+
+#include "cache.hpp"
+#include "load.hpp"
 
 extern uint32_t __bss_link_start __asm("__bss_link_start");
 extern uint32_t __bss_link_end __asm("__bss_link_end");
diff --git a/float_uint32.h b/float_uint32.hpp
similarity index 100%
rename from float_uint32.h
rename to float_uint32.hpp
diff --git a/holly.h b/holly.hpp
similarity index 99%
rename from holly.h
rename to holly.hpp
index 7287787..6458a44 100644
--- a/holly.h
+++ b/holly.hpp
@@ -1,7 +1,7 @@
-#include <stdint.h>
-#include <stddef.h>
+#include <cstdint>
+#include <cstddef>
 
-#include "type.h"
+#include "type.hpp"
 
 struct holly_reg {
   reg32 ID;                /* Device ID */
diff --git a/holly/background.cpp b/holly/background.cpp
index a4da067..5fd72a9 100644
--- a/holly/background.cpp
+++ b/holly/background.cpp
@@ -1,6 +1,6 @@
 #include <cstdint>
 
-#include "isp_tsp.h"
+#include "isp_tsp.hpp"
 
 struct vertex_parameter {
   float x;
diff --git a/holly/background.h b/holly/background.hpp
similarity index 100%
rename from holly/background.h
rename to holly/background.hpp
diff --git a/holly/core.cpp b/holly/core.cpp
index 8aa531e..246cc41 100644
--- a/holly/core.cpp
+++ b/holly/core.cpp
@@ -1,13 +1,13 @@
-#include "../float_uint32.h"
-#include "core_bits.h"
-#include "../holly.h"
-#include "../memorymap.h"
+#include "../float_uint32.hpp"
+#include "core_bits.hpp"
+#include "../holly.hpp"
+#include "../memorymap.hpp"
 
-#include "texture_memory_alloc.h"
+#include "texture_memory_alloc.hpp"
 
-#include "core.h"
-#include "background.h"
-#include "region_array.h"
+#include "core.hpp"
+#include "background.hpp"
+#include "region_array.hpp"
 
 void core_init()
 {
diff --git a/holly/core.h b/holly/core.hpp
similarity index 100%
rename from holly/core.h
rename to holly/core.hpp
diff --git a/holly/core_bits.h b/holly/core_bits.hpp
similarity index 99%
rename from holly/core_bits.h
rename to holly/core_bits.hpp
index 772da55..3b25c11 100644
--- a/holly/core_bits.h
+++ b/holly/core_bits.hpp
@@ -1,6 +1,6 @@
 #include <cstdint>
 
-#include "../float_uint32.h"
+#include "../float_uint32.hpp"
 
 namespace id {
   constexpr uint32_t device_id(uint32_t reg) { return (reg >> 16) & 0xffff; }
diff --git a/holly/isp_tsp.h b/holly/isp_tsp.hpp
similarity index 100%
rename from holly/isp_tsp.h
rename to holly/isp_tsp.hpp
diff --git a/holly/region_array.cpp b/holly/region_array.cpp
index 64e139a..f09057c 100644
--- a/holly/region_array.cpp
+++ b/holly/region_array.cpp
@@ -1,4 +1,6 @@
-#include "region_array.h"
+#include <cstdint>
+
+#include "region_array.hpp"
 
 #define REGION_ARRAY__LAST_REGION (1 << 31)
 #define REGION_ARRAY__Z_CLEAR (1 << 30)
diff --git a/holly/region_array.h b/holly/region_array.hpp
similarity index 100%
rename from holly/region_array.h
rename to holly/region_array.hpp
diff --git a/holly/ta_bits.h b/holly/ta_bits.hpp
similarity index 99%
rename from holly/ta_bits.h
rename to holly/ta_bits.hpp
index ce28ee6..1bf9ee6 100644
--- a/holly/ta_bits.h
+++ b/holly/ta_bits.hpp
@@ -1,6 +1,6 @@
 #include <cstdint>
 
-#include "../float_uint32.h"
+#include "../float_uint32.hpp"
 
 namespace ta_ol_base {
   constexpr uint32_t base_address(uint32_t num) { return (num & 0xffffe0) << 0; }
diff --git a/holly/ta_fifo_polygon_converter.cpp b/holly/ta_fifo_polygon_converter.cpp
index fe43d3a..da05fbc 100644
--- a/holly/ta_fifo_polygon_converter.cpp
+++ b/holly/ta_fifo_polygon_converter.cpp
@@ -1,16 +1,16 @@
 #include <cstdint>
 
-#include "core_bits.h"
-#include "ta_bits.h"
-#include "../holly.h"
-#include "../systembus.h"
-#include "../systembus_bits.h"
-#include "../sh7091.h"
-#include "../sh7091_bits.h"
+#include "core_bits.hpp"
+#include "ta_bits.hpp"
+#include "../holly.hpp"
+#include "../systembus.hpp"
+#include "../systembus_bits.hpp"
+#include "../sh7091.hpp"
+#include "../sh7091_bits.hpp"
 
-#include "texture_memory_alloc.h"
+#include "texture_memory_alloc.hpp"
 
-#include "ta_fifo_polygon_converter.h"
+#include "ta_fifo_polygon_converter.hpp"
 
 void ta_polygon_converter_init()
 {
diff --git a/holly/ta_fifo_polygon_converter.h b/holly/ta_fifo_polygon_converter.hpp
similarity index 100%
rename from holly/ta_fifo_polygon_converter.h
rename to holly/ta_fifo_polygon_converter.hpp
diff --git a/holly/ta_parameter.h b/holly/ta_parameter.hpp
similarity index 99%
rename from holly/ta_parameter.h
rename to holly/ta_parameter.hpp
index 963c33e..fdf679b 100644
--- a/holly/ta_parameter.h
+++ b/holly/ta_parameter.hpp
@@ -1,8 +1,8 @@
 #include <cstdint>
 #include <cstddef>
 
-#include "../float_uint32.h"
-#include "isp_tsp.h"
+#include "../float_uint32.hpp"
+#include "isp_tsp.hpp"
 
 namespace para_control {
   namespace para_type {
diff --git a/holly/texture_memory_alloc.h b/holly/texture_memory_alloc.hpp
similarity index 100%
rename from holly/texture_memory_alloc.h
rename to holly/texture_memory_alloc.hpp
diff --git a/imask.h b/imask.hpp
similarity index 100%
rename from imask.h
rename to imask.hpp
diff --git a/load.cpp b/load.cpp
index 49b1b7d..9e12807 100644
--- a/load.cpp
+++ b/load.cpp
@@ -1,8 +1,8 @@
-#include <stdint.h>
+#include <cstdint>
 
-#include "sh7091.h"
-#include "sh7091_bits.h"
-#include "holly.h"
+#include "sh7091.hpp"
+#include "sh7091_bits.hpp"
+#include "holly.hpp"
 
 enum load_command {
   CMD_NONE,
diff --git a/load.h b/load.hpp
similarity index 100%
rename from load.h
rename to load.hpp
diff --git a/macaw.h b/macaw.hpp
similarity index 100%
rename from macaw.h
rename to macaw.hpp
diff --git a/main.cpp b/main.cpp
index 6ae5739..07f2e96 100644
--- a/main.cpp
+++ b/main.cpp
@@ -1,67 +1,29 @@
-#include <stdint.h>
+#include <cstdint>
 
-#include "memorymap.h"
+#include "memorymap.hpp"
 
-#include "sh7091.h"
-#include "sh7091_bits.h"
-#include "holly.h"
-#include "holly/core.h"
-#include "holly/core_bits.h"
-#include "holly/ta_fifo_polygon_converter.h"
-#include "systembus.h"
-#include "maple/maple.h"
-#include "maple/maple_bits.h"
-#include "maple/maple_bus_commands.h"
-#include "maple/maple_bus_ft0.h"
+#include "sh7091.hpp"
+#include "sh7091_bits.hpp"
+#include "holly.hpp"
+#include "holly/core.hpp"
+#include "holly/core_bits.hpp"
+#include "holly/ta_fifo_polygon_converter.hpp"
+#include "systembus.hpp"
+#include "maple/maple.hpp"
+#include "maple/maple_bits.hpp"
+#include "maple/maple_bus_commands.hpp"
+#include "maple/maple_bus_ft0.hpp"
 
-#include "holly/texture_memory_alloc.h"
+#include "holly/texture_memory_alloc.hpp"
 
-#include "cache.h"
-#include "load.h"
-#include "vga.h"
-#include "rgb.h"
-#include "string.h"
-#include "scene.h"
+#include "cache.hpp"
+#include "load.hpp"
+#include "vga.hpp"
+#include "rgb.hpp"
+#include "string.hpp"
+#include "scene.hpp"
 
-#include "macaw.h"
-
-extern uint32_t __bss_link_start __asm("__bss_link_start");
-extern uint32_t __bss_link_end __asm("__bss_link_end");
-
-void serial()
-{
-  sh7091.SCIF.SCSCR2 = 0;
-  sh7091.SCIF.SCSMR2 = 0;
-  sh7091.SCIF.SCBRR2 = 1; // 520833.3
-
-  sh7091.SCIF.SCFCR2 = SCFCR2__TFRST | SCFCR2__RFRST;
-  // tx/rx trigger on 1 byte
-  sh7091.SCIF.SCFCR2 = 0;
-
-  sh7091.SCIF.SCSPTR2 = 0;
-  sh7091.SCIF.SCLSR2 = 0;
-
-  sh7091.SCIF.SCSCR2 = SCSCR2__TE | SCSCR2__RE;
-}
-
-inline void serial_char(const char c)
-{
-  // wait for transmit fifo to become empty
-  while ((sh7091.SCIF.SCFSR2 & SCFSR2__TDFE) == 0);
-
-  for (int i = 0; i < 100000; i++) {
-    asm volatile ("nop;");
-  }
-
-  sh7091.SCIF.SCFTDR2 = static_cast<uint8_t>(c);
-}
-
-void serial_string(const char * s)
-{
-  while (*s != '\0') {
-    serial_char(*s++);
-  }
-}
+#include "macaw.hpp"
 
 /* must be aligned to 32-bytes for DMA transfer */
 // the aligned(32) attribute does not actually align to 32 bytes; gcc is the best compiler.
@@ -69,54 +31,9 @@ void serial_string(const char * s)
 // __attribute__((aligned(32)))
 uint32_t _scene[((32 * 6) + 32) / 4];
 
-template <typename T>
-T * align_32byte(T * mem)
-{
-  return reinterpret_cast<T *>((((reinterpret_cast<uint32_t>(mem) + 31) & ~31)));
-}
-
-void serial_int32(const uint32_t n)
-{
-  char num_buf[9];
-  string::hex<char>(num_buf, 8, n);
-  num_buf[8] = 0;
-  serial_string("0x");
-  serial_string(num_buf);
-  serial_string("\n");
-}
-
-void serial_int8(const uint8_t n)
-{
-  char num_buf[3];
-  string::hex<char>(num_buf, 2, n);
-  num_buf[2] = 0;
-  serial_string("0x");
-  serial_string(num_buf);
-  serial_string("\n");
-}
-
 uint32_t _receive_address[(256 + 32) / 4] = {0};
 uint32_t _command_buf[(256 + 32) / 4] = {0};
 
-extern uint32_t _binary_wink_data_start __asm("_binary_wink_data_start");
-
-void make_wink(uint32_t * buf)
-{
-  const uint8_t * src = reinterpret_cast<const uint8_t *>(&_binary_wink_data_start);
-  uint8_t * dst = reinterpret_cast<uint8_t *>(buf);
-
-  uint32_t ix = 0;
-  dst[ix] = 0;
-  for (int i = 0; i < 48 * 32; i++) {
-    dst[ix] |= ((src[i] & 1) << (7 - (i % 8)));
-
-    if (i % 8 == 7) {
-      ix++;
-      dst[ix] = 0;
-    }
-  }
-}
-
 bool maple_test()
 {
   v_sync_out();
@@ -128,9 +45,6 @@ bool maple_test()
 
   //maple_init_device_request(command_buf, receive_address);
   //maple_init_get_condition(command_buf, receive_address);
-  uint32_t wink_buf[192];
-  make_wink(wink_buf);
-  maple_init_block_write(command_buf, receive_address, wink_buf);
 
   serial_int32(command_buf[0]);
   serial_char('\n');
@@ -154,18 +68,8 @@ bool maple_test()
   return !(data_format->data.digital_button & ft0::data_transfer::digital_button::a);
 }
 
-extern "C"
 void main()
 {
-  cache_init();
-
-  // clear BSS
-  uint32_t * start = &__bss_link_start;
-  uint32_t * end = &__bss_link_end;
-  while (start < end) {
-    *start++ = 0;
-  }
-
   //serial();
 
   vga();
diff --git a/maple/maple.cpp b/maple/maple.cpp
index a0762aa..9a7e5c3 100644
--- a/maple/maple.cpp
+++ b/maple/maple.cpp
@@ -1,27 +1,29 @@
 #include <cstdint>
 #include <bit>
 
-#include "../sh7091.h"
-#include "../sh7091_bits.h"
-#include "../systembus.h"
-#include "../systembus_bits.h"
+#include "../sh7091.hpp"
+#include "../sh7091_bits.hpp"
+#include "../systembus.hpp"
+#include "../systembus_bits.hpp"
 
-#include "maple_bits.h"
-#include "maple_host_bits.h"
-#include "maple_bus_commands.h"
-#include "maple.h"
+#include "maple_bits.hpp"
+#include "maple_bus_bits.hpp"
+#include "maple_bus_commands.hpp"
+#include "maple.hpp"
 
-void maple_init_host_command(uint32_t * buf, uint32_t * receive_address, uint8_t command_code, uint8_t data_size)
+namespace maple {
+
+void init_host_command(uint32_t * buf, uint32_t * receive_buf, uint8_t command_code, uint8_t data_size)
 {
   // this function does not care about the template instantiation of
-  // maple_host_command--data_fields is not manipulated here.
-  auto host_command = reinterpret_cast<maple_host_command<uint32_t> *>(buf);
+  // host_command--data_fields is not manipulated here.
+  auto host_command = reinterpret_cast<struct host_command<uint32_t> *>(buf);
 
   host_command->host_instruction = host_instruction::end_flag
-    | host_instruction::port_select::a
-    | host_instruction::transfer_length((data_size / 4));
+				 | host_instruction::port_select::a
+				 | host_instruction::transfer_length((data_size / 4));
 
-  host_command->receive_data_storage_address = reinterpret_cast<uint32_t>(receive_address) & 0x1fff'ffff;
+  host_command->receive_data_storage_address = reinterpret_cast<uint32_t>(receive_buf) & 0x1fff'ffff;
 
   host_command->bus_data.command_code = command_code;
   host_command->bus_data.destination_ap = ap::de::expansion_device | ap::port_select::a | ap::lm_bus::_0;
@@ -29,54 +31,47 @@ void maple_init_host_command(uint32_t * buf, uint32_t * receive_address, uint8_t
   host_command->bus_data.data_size = data_size / 4;
 }
 
-void maple_init_device_request(uint32_t * buf, uint32_t * receive_address)
+void init_device_request(uint32_t * buf, uint32_t * receive_buf)
 {
-  maple_init_host_command(buf, receive_address, device_request::command_code, (sizeof (struct device_request::data_fields)));
+  init_host_command(buf, receive_buf, device_request::command_code, (sizeof (struct device_request::data_fields)));
 }
 
-void maple_init_get_condition(uint32_t * buf, uint32_t * receive_address)
+void init_get_condition(uint32_t * buf, uint32_t * receive_buf)
 {
-  maple_init_host_command(buf, receive_address, get_condition::command_code, (sizeof (struct get_condition::data_fields)));
+  init_host_command(buf, receive_buf, get_condition::command_code, (sizeof (struct get_condition::data_fields)));
 
-  auto host_command = reinterpret_cast<maple_host_command<get_condition::data_fields> *>(buf);
+  auto host_command = reinterpret_cast<struct host_command<get_condition::data_fields> *>(buf);
 
-  auto& function_type = host_command->bus_data.data_fields.function_type;
+  auto& fields = host_command->bus_data.data_fields;
   // controller function type
-  function_type[0] = 0x00;
-  function_type[1] = 0x00;
-  function_type[2] = 0x00;
-  function_type[3] = 0x01;
+  fields.function_type = std::byteswap(function_type::controller);
 }
 
-void maple_init_block_write(uint32_t * buf, uint32_t * receive_address, uint32_t * data)
+void init_block_write(uint32_t * buf, uint32_t * receive_buf, uint32_t * data)
 {
-  maple_init_host_command(buf, receive_address, block_write::command_code, (sizeof (struct block_write::data_fields<uint32_t[192 / 4]>)));
+  init_host_command(buf, receive_buf, block_write::command_code, (sizeof (struct block_write::data_fields<uint32_t[192 / 4]>)));
 
-  auto host_command = reinterpret_cast<maple_host_command<block_write::data_fields<uint32_t[192 / 4]>> *>(buf);
+  auto host_command = reinterpret_cast<struct host_command<block_write::data_fields<uint32_t[192 / 4]>> *>(buf);
 
   auto& fields = host_command->bus_data.data_fields;
   // BW LCD function type
-  fields.function_type[0] = 0x00;
-  fields.function_type[1] = 0x00;
-  fields.function_type[2] = 0x00;
-  fields.function_type[3] = 0x04;
+  fields.function_type = std::byteswap(function_type::bw_lcd);
 
   // lcd number 0 (1 total lcd)
-  fields.pt[0] = 0;
+  fields.pt = 0;
 
   // phase 0 (from 0 to 3)
-  fields.phase[0] = 0;
+  fields.phase = 0;
 
   // plane 0 (2 total levels of gradation)
-  fields.block_no[0] = 0x00;
-  fields.block_no[1] = 0x00;
+  fields.block_no = std::byteswap(0x0000);
 
   for (uint32_t i = 0; i < (192 / 4); i++) {
     fields.written_data[i] = data[i];
   }
 }
 
-void maple_dma_start(uint32_t * command_buf)
+void dma_start(uint32_t * command_buf)
 {
   sh7091.DMAC.DMAOR = DMAOR__DDT                 /* on-demand data transfer mode */
     | DMAOR__PR__CH2_CH0_CH1_CH3 /* priority mode; CH2 > CH0 > CH1 > CH3 */
@@ -113,3 +108,5 @@ void maple_dma_start(uint32_t * command_buf)
   while ((system.ISTNRM & ISTNRM__END_OF_DMA_MAPLE_DMA) == 0);
   system.ISTNRM = ISTNRM__END_OF_DMA_MAPLE_DMA;
 }
+
+}
diff --git a/maple/maple.h b/maple/maple.h
deleted file mode 100644
index 621542e..0000000
--- a/maple/maple.h
+++ /dev/null
@@ -1,22 +0,0 @@
-#pragma once
-
-#include <cstdint>
-
-template <typename T>
-struct maple_host_command {
-  uint32_t host_instruction;
-  uint32_t receive_data_storage_address;
-  struct bus_data {
-    uint8_t command_code;
-    uint8_t destination_ap;
-    uint8_t source_ap;
-    uint8_t data_size;
-    T data_fields;
-  } bus_data;
-};
-
-void maple_init_host_command(uint32_t * buf, uint32_t * receive_address);
-void maple_init_device_request(uint32_t * buf, uint32_t * receive_address);
-void maple_init_get_condition(uint32_t * buf, uint32_t * receive_address);
-void maple_init_block_write(uint32_t * buf, uint32_t * receive_address, uint32_t * data);
-void maple_dma_start(uint32_t * command_buf);
diff --git a/maple/maple.hpp b/maple/maple.hpp
new file mode 100644
index 0000000..9450fd3
--- /dev/null
+++ b/maple/maple.hpp
@@ -0,0 +1,27 @@
+#pragma once
+
+#include <cstdint>
+
+namespace maple {
+
+template <typename T>
+struct host_command {
+  uint32_t host_instruction;
+  uint32_t receive_data_storage_address;
+  struct bus_data {
+    uint8_t command_code;
+    uint8_t destination_ap;
+    uint8_t source_ap;
+    uint8_t data_size;
+    T data_fields;
+  } bus_data;
+};
+
+void init_host_command(uint32_t * command_buf, uint32_t * receive_buf);
+void init_device_request(uint32_t * command_buf, uint32_t * receive_buf);
+void init_get_condition(uint32_t * command_buf, uint32_t * receive_buf);
+void init_block_write(uint32_t * command_buf, uint32_t * receive_buf, uint32_t * data);
+void dma_start(uint32_t * command_buf);
+
+}
+
diff --git a/maple/maple_bits.h b/maple/maple_bits.hpp
similarity index 98%
rename from maple/maple_bits.h
rename to maple/maple_bits.hpp
index 860786a..95b7fa4 100644
--- a/maple/maple_bits.h
+++ b/maple/maple_bits.hpp
@@ -1,6 +1,6 @@
 #include <cstdint>
 
-#include "../float_uint32.h"
+#include "../float_uint32.hpp"
 
 namespace mdstar {
   constexpr uint32_t table_address(uint32_t num) { return (num & 0xfffffe0) << 0; }
diff --git a/maple/maple_host_bits.h b/maple/maple_bus_bits.hpp
similarity index 69%
rename from maple/maple_host_bits.h
rename to maple/maple_bus_bits.hpp
index 1949632..85c2980 100644
--- a/maple/maple_host_bits.h
+++ b/maple/maple_bus_bits.hpp
@@ -1,6 +1,6 @@
 #include <cstdint>
 
-#include "../float_uint32.h"
+#include "../float_uint32.hpp"
 
 namespace host_instruction {
   constexpr uint32_t end_flag = 1 << 31;
@@ -46,3 +46,18 @@ namespace ap {
   }
 }
 
+namespace function_type {
+  constexpr uint32_t camera = 1 << 11;
+  constexpr uint32_t exchange_media = 1 << 10;
+  constexpr uint32_t pointing = 1 << 9;
+  constexpr uint32_t vibration = 1 << 8;
+  constexpr uint32_t light_gun = 1 << 7;
+  constexpr uint32_t keyboard = 1 << 6;
+  constexpr uint32_t ar_gun = 1 << 5;
+  constexpr uint32_t audio_input = 1 << 4;
+  constexpr uint32_t timer = 1 << 3;
+  constexpr uint32_t bw_lcd = 1 << 2;
+  constexpr uint32_t storage = 1 << 1;
+  constexpr uint32_t controller = 1 << 0;
+}
+
diff --git a/maple/maple_bus_commands.h b/maple/maple_bus_commands.hpp
similarity index 79%
rename from maple/maple_bus_commands.h
rename to maple/maple_bus_commands.hpp
index 68b90fc..560291e 100644
--- a/maple/maple_bus_commands.h
+++ b/maple/maple_bus_commands.hpp
@@ -33,12 +33,12 @@ namespace device_status {
   
   struct data_fields {
     uint8_t device_id[16];
-    uint8_t destination_code[1];
-    uint8_t connection_direction[1];
+    uint8_t destination_code;
+    uint8_t connection_direction;
     uint8_t product_name[30];
     uint8_t license[60];
-    uint8_t low_consumption_standby_current[2];
-    uint8_t maximum_current_consumption[2];
+    uint16_t low_consumption_standby_current;
+    uint16_t maximum_current_consumption;
   };
   
   static_assert((sizeof (struct data_fields)) == 112);
@@ -50,12 +50,12 @@ namespace device_all_status {
   template <typename T>
   struct data_fields {
     uint8_t device_id[16];
-    uint8_t destination_code[1];
-    uint8_t connection_direction[1];
+    uint8_t destination_code;
+    uint8_t connection_direction;
     uint8_t product_name[30];
     uint8_t license[60];
-    uint8_t low_consumption_standby_current[2];
-    uint8_t maximum_current_consumption[2];
+    uint16_t low_consumption_standby_current;
+    uint16_t maximum_current_consumption;
     T free_device_status;
   };
   
@@ -74,7 +74,7 @@ namespace data_transfer {
   
   template <typename T>
   struct data_fields {
-    uint8_t function_type[4];
+    uint32_t function_type;
     T data;
   };
   
@@ -85,7 +85,7 @@ namespace get_condition {
   constexpr uint32_t command_code = 0x9;
   
   struct data_fields {
-    uint8_t function_type[4];
+    uint32_t function_type;
   };
   
   static_assert((sizeof (struct data_fields)) == 4);
@@ -95,8 +95,8 @@ namespace get_media_info {
   constexpr uint32_t command_code = 0xa;
   
   struct data_fields {
-    uint8_t function_type[4];
-    uint8_t pt[4];
+    uint32_t function_type;
+    uint32_t pt;
   };
   
   static_assert((sizeof (struct data_fields)) == 8);
@@ -106,10 +106,10 @@ namespace block_read {
   constexpr uint32_t command_code = 0xb;
   
   struct data_fields {
-    uint8_t function_type[4];
-    uint8_t pt[1];
-    uint8_t phase[1];
-    uint8_t block_no[2];
+    uint32_t function_type;
+    uint8_t pt;
+    uint8_t phase;
+    uint16_t block_no;
   };
   
   static_assert((sizeof (struct data_fields)) == 8);
@@ -120,10 +120,10 @@ namespace block_write {
   
   template <typename T>
   struct data_fields {
-    uint8_t function_type[4];
-    uint8_t pt[1];
-    uint8_t phase[1];
-    uint8_t block_no[2];
+    uint32_t function_type;
+    uint8_t pt;
+    uint8_t phase;
+    uint16_t block_no;
     T written_data;
   };
   
@@ -134,10 +134,10 @@ namespace get_last_error {
   constexpr uint32_t command_code = 0xd;
   
   struct data_fields {
-    uint8_t function_type[4];
-    uint8_t pt[1];
-    uint8_t phase[1];
-    uint8_t block_no[2];
+    uint32_t function_type;
+    uint8_t pt;
+    uint8_t phase;
+    uint16_t block_no;
   };
   
   static_assert((sizeof (struct data_fields)) == 8);
@@ -148,7 +148,7 @@ namespace set_condition {
   
   template <typename T>
   struct data_fields {
-    uint8_t function_type[4];
+    uint32_t function_type;
     T write_in_data;
   };
   
@@ -160,7 +160,7 @@ namespace ft4_control {
   
   template <typename T>
   struct data_fields {
-    uint8_t function_type[4];
+    uint32_t function_type;
     T ft4_data;
   };
   
@@ -172,7 +172,7 @@ namespace ar_control {
   
   template <typename T>
   struct data_fields {
-    uint8_t function_type[4];
+    uint32_t function_type;
     T data;
   };
   
@@ -204,7 +204,7 @@ namespace file_error {
   constexpr uint32_t command_code = 0xfb;
   
   struct data_fields {
-    uint8_t function_error_code[4];
+    uint32_t function_error_code;
   };
   
   static_assert((sizeof (struct data_fields)) == 4);
@@ -214,7 +214,7 @@ namespace lcd_error {
   constexpr uint32_t command_code = 0xfa;
   
   struct data_fields {
-    uint8_t function_error_code[4];
+    uint32_t function_error_code;
   };
   
   static_assert((sizeof (struct data_fields)) == 4);
@@ -224,7 +224,7 @@ namespace ar_error {
   constexpr uint32_t command_code = 0xf9;
   
   struct data_fields {
-    uint8_t function_error_code[4];
+    uint32_t function_error_code;
   };
   
   static_assert((sizeof (struct data_fields)) == 4);
diff --git a/maple/maple_bus_ft0.h b/maple/maple_bus_ft0.hpp
similarity index 100%
rename from maple/maple_bus_ft0.h
rename to maple/maple_bus_ft0.hpp
diff --git a/memorymap.h b/memorymap.hpp
similarity index 97%
rename from memorymap.h
rename to memorymap.hpp
index fed8642..80ce5eb 100644
--- a/memorymap.h
+++ b/memorymap.hpp
@@ -1,4 +1,4 @@
-#include <stdint.h>
+#include <cstdint>
 
 extern volatile uint32_t system_boot_rom[0x200000] __asm("system_boot_rom");
 extern volatile uint32_t aica_wave_memory[0x200000] __asm("aica_wave_memory");
diff --git a/regs/gen/core_bits.py b/regs/gen/core_bits.py
index 4d06746..aca89c1 100644
--- a/regs/gen/core_bits.py
+++ b/regs/gen/core_bits.py
@@ -36,7 +36,7 @@ def aggregate_enums(aggregated_rows):
     def assert_unique_ordered(bits):
         nonlocal all_bits
         assert all(bit not in all_bits for bit in bits), bits
-        assert max(all_bits, default=32) > max(bits)
+        assert max(all_bits, default=32) > max(bits), (all_bits, bits)
         all_bits |= bits
 
     for row in aggregated_rows:
@@ -206,7 +206,7 @@ def render_registers(registers):
 def header():
     yield "#include <cstdint>"
     yield ""
-    yield '#include "../float_uint32.h"'
+    yield '#include "../float_uint32.hpp"'
     yield ""
 
 if __name__ == "__main__":
diff --git a/regs/gen/maple_commands.py b/regs/gen/maple_commands.py
index 6429bb1..e31e634 100644
--- a/regs/gen/maple_commands.py
+++ b/regs/gen/maple_commands.py
@@ -34,7 +34,11 @@ def command_namespace(namespace: CommandNamespace,
         for field_name, field_size in data_fields:
             const, var = field_size
             if var is None:
-                yield f"uint8_t {field_name}[{const}];"
+                if const in {1, 2, 4}:
+                    bits = const * 8
+                    yield f"uint{bits}_t {field_name};"
+                else:
+                    yield f"uint8_t {field_name}[{const}];"
             elif const == 0:
                 assert var == "n"
                 yield f"T {field_name};"
diff --git a/regs/gen/memorymap.py b/regs/gen/memorymap.py
index 8540835..ec88a6f 100644
--- a/regs/gen/memorymap.py
+++ b/regs/gen/memorymap.py
@@ -5,7 +5,7 @@ from sh7091 import read_input
 from generate import renderer
 
 def includes():
-    yield "#include <stdint.h>"
+    yield "#include <cstdint>"
     yield ""
 
 def process_rows(rows):
diff --git a/regs/gen/sh7091.py b/regs/gen/sh7091.py
index c416f61..63f88a6 100644
--- a/regs/gen/sh7091.py
+++ b/regs/gen/sh7091.py
@@ -145,10 +145,10 @@ def blocks(rows):
     yield 'extern struct sh7091_reg sh7091 __asm("sh7091");'
 
 def headers():
-    yield "#include <stdint.h>"
-    yield "#include <stddef.h>"
+    yield "#include <cstdint>"
+    yield "#include <cstddef>"
     yield ""
-    yield '#include "type.h"'
+    yield '#include "type.hpp"'
     yield ""
 
 if __name__ == "__main__":
diff --git a/regs/maple_bus_bits.csv b/regs/maple_bus_bits.csv
new file mode 100644
index 0000000..029a65e
--- /dev/null
+++ b/regs/maple_bus_bits.csv
@@ -0,0 +1,43 @@
+"register_name","enum_name","bits","bit_name","value","mask","description"
+"host_instruction",,31,"end_flag",1,,
+,,,,,,
+"host_instruction","port_select","17-16","a",0,,
+"host_instruction","port_select","17-16","b",1,,
+"host_instruction","port_select","17-16","c",2,,
+"host_instruction","port_select","17-16","d",3,,
+,,,,,,
+"host_instruction","pattern","10-8","normal","0b000",,
+"host_instruction","pattern","10-8","light_gun_mode","0b010",,
+"host_instruction","pattern","10-8","reset","0b011",,
+"host_instruction","pattern","10-8","return_from_light_gun_mode","0b100",,
+"host_instruction","pattern","10-8","nop","0b111",,
+,,,,,,
+"host_instruction",,"7-0","transfer_length",,"0xff",
+,,,,,,
+"ap","port_select","7-6","a","0b00",,
+"ap","port_select","7-6","b","0b01",,
+"ap","port_select","7-6","c","0b10",,
+"ap","port_select","7-6","d","0b11",,
+,,,,,,
+"ap","de",5,"device",1,,
+"ap","de",5,"expansion_device",0,,
+"ap","de",5,"port",0,,
+,,,,,,
+"ap","lm_bus","4-0","_4","0b10000",,
+"ap","lm_bus","4-0","_3","0b01000",,
+"ap","lm_bus","4-0","_2","0b00100",,
+"ap","lm_bus","4-0","_1","0b00010",,
+"ap","lm_bus","4-0","_0","0b00001",,
+,,,,,,
+"function_type",,11,"camera",1,,
+"function_type",,10,"exchange_media",1,,
+"function_type",,9,"pointing",1,,
+"function_type",,8,"vibration",1,,
+"function_type",,7,"light_gun",1,,
+"function_type",,6,"keyboard",1,,
+"function_type",,5,"ar_gun",1,,
+"function_type",,4,"audio_input",1,,
+"function_type",,3,"timer",1,,
+"function_type",,2,"bw_lcd",1,,
+"function_type",,1,"storage",1,,
+"function_type",,0,"controller",1,,
diff --git a/regs/maple_bus_bits.ods b/regs/maple_bus_bits.ods
new file mode 100644
index 0000000..16322fc
Binary files /dev/null and b/regs/maple_bus_bits.ods differ
diff --git a/rgb.cpp b/rgb.cpp
index e7f34d4..accf545 100644
--- a/rgb.cpp
+++ b/rgb.cpp
@@ -1,4 +1,4 @@
-#include "rgb.h"
+#include "rgb.hpp"
 
 struct rgb hsv_to_rgb(struct hsv hsv)
 {
diff --git a/rgb.h b/rgb.hpp
similarity index 100%
rename from rgb.h
rename to rgb.hpp
diff --git a/runtime.cpp b/runtime.cpp
new file mode 100644
index 0000000..bdd69a5
--- /dev/null
+++ b/runtime.cpp
@@ -0,0 +1,27 @@
+#include <cstdint>
+
+#include "cache.hpp"
+
+extern uint32_t __bss_link_start __asm("__bss_link_start");
+extern uint32_t __bss_link_end __asm("__bss_link_end");
+
+extern void main();
+
+extern "C"
+void runtime_init()
+__attribute__((section(".text.startup.runtime_init")));
+
+extern "C"
+void runtime_init()
+{
+  // clear BSS
+  uint32_t * start = &__bss_link_start;
+  uint32_t * end = &__bss_link_end;
+  while (start < end) {
+    *start++ = 0;
+  }
+
+  cache::init();
+
+  main();
+}
diff --git a/scene.cpp b/scene.cpp
index 4896bad..a9a5d6d 100644
--- a/scene.cpp
+++ b/scene.cpp
@@ -1,9 +1,9 @@
 #include <cstdint>
 #include <cstddef>
 
-#include "holly/ta_parameter.h"
+#include "holly/ta_parameter.hpp"
 
-#include "holly/texture_memory_alloc.h"
+#include "holly/texture_memory_alloc.hpp"
 
 /*
   -0.5,-0.5      0.5,-0.5
@@ -80,8 +80,6 @@ union ta_parameter {
   struct global_end_of_list global_end_of_list;
 };
 
-extern void serial_string(const char * s);
-
 uint32_t scene_transform(uint32_t * _scene)
 {
   ta_parameter * scene = reinterpret_cast<ta_parameter *>(&_scene[0]);
diff --git a/scene.h b/scene.hpp
similarity index 100%
rename from scene.h
rename to scene.hpp
diff --git a/serial.cpp b/serial.cpp
new file mode 100644
index 0000000..ba028bc
--- /dev/null
+++ b/serial.cpp
@@ -0,0 +1,61 @@
+#include <cstdint>
+
+#include "sh7091.hpp"
+#include "sh7091_bits.hpp"
+
+#include "string.hpp"
+
+namespace serial {
+
+void init()
+{
+  sh7091.SCIF.SCSCR2 = 0;
+  sh7091.SCIF.SCSMR2 = 0;
+  sh7091.SCIF.SCBRR2 = 1; // 520833.3
+
+  sh7091.SCIF.SCFCR2 = SCFCR2__TFRST | SCFCR2__RFRST;
+  // tx/rx trigger on 1 byte
+  sh7091.SCIF.SCFCR2 = 0;
+
+  sh7091.SCIF.SCSPTR2 = 0;
+  sh7091.SCIF.SCLSR2 = 0;
+
+  sh7091.SCIF.SCSCR2 = SCSCR2__TE | SCSCR2__RE;
+}
+
+void character(const char c)
+{
+  // wait for transmit fifo to become empty
+  while ((sh7091.SCIF.SCFSR2 & SCFSR2__TDFE) == 0);
+
+  for (int i = 0; i < 100000; i++) {
+    asm volatile ("nop;");
+  }
+
+  sh7091.SCIF.SCFTDR2 = static_cast<uint8_t>(c);
+}
+
+void string(const char * s)
+{
+  while (*s != '\0') {
+    character(*s++);
+  }
+}
+
+template <typename T>
+void integer(const T n)
+{
+  constexpr uint32_t length = (sizeof (T)) * 2;
+  char num_buf[length + 1];
+  string::hex<char>(num_buf, length, n);
+  num_buf[length] = 0;
+  string("0x");
+  string(num_buf);
+  string("\n");
+}
+
+template void integer<uint32_t>(uint32_t param);
+template void integer<uint16_t>(uint16_t param);
+template void integer<uint8_t>(uint8_t param);
+
+}
diff --git a/serial.hpp b/serial.hpp
new file mode 100644
index 0000000..add72ea
--- /dev/null
+++ b/serial.hpp
@@ -0,0 +1,12 @@
+namespace serial {
+
+void init();
+
+void character(const char c);
+
+void string(const char * s);
+
+template <typename T>
+void integer(const T n);
+
+}
diff --git a/sh7091.h b/sh7091.hpp
similarity index 99%
rename from sh7091.h
rename to sh7091.hpp
index 039e71a..0eb68ba 100644
--- a/sh7091.h
+++ b/sh7091.hpp
@@ -1,7 +1,7 @@
-#include <stdint.h>
-#include <stddef.h>
+#include <cstdint>
+#include <cstddef>
 
-#include "type.h"
+#include "type.hpp"
 
 struct ccn_reg {
   reg32 PTEH;              /* Page table entry high register */
diff --git a/sh7091_bits.h b/sh7091_bits.hpp
similarity index 100%
rename from sh7091_bits.h
rename to sh7091_bits.hpp
diff --git a/start.s b/start.s
index b4e5b23..2268c9b 100644
--- a/start.s
+++ b/start.s
@@ -10,8 +10,8 @@ _start:
         or      r1,r0
         ldc     r0,sr
 
-        /* jump to main */
-        mov.l main_ptr,r0
+        /* jump to runtime_init */
+        mov.l runtime_init_ptr,r0
         jmp @r0
         nop
 
@@ -20,5 +20,5 @@ p1ram_end_ptr:
         .long __p1ram_end
 imask_all:
         .long 0xf0
-main_ptr:
-        .long _main
+runtime_init_ptr:
+        .long _runtime_init
diff --git a/storequeue.cpp b/storequeue.cpp
index e3d05ed..a5cfdda 100644
--- a/storequeue.cpp
+++ b/storequeue.cpp
@@ -1,5 +1,5 @@
-#include "sh7091.h"
-#include "memorymap.h"
+#include "sh7091.hpp"
+#include "memorymap.hpp"
 
 void sq_transfer_32byte(volatile void * dst)
 {
diff --git a/storequeue.h b/storequeue.hpp
similarity index 100%
rename from storequeue.h
rename to storequeue.hpp
diff --git a/string.h b/string.hpp
similarity index 100%
rename from string.h
rename to string.hpp
diff --git a/systembus.h b/systembus.hpp
similarity index 99%
rename from systembus.h
rename to systembus.hpp
index 7f30197..f67dcda 100644
--- a/systembus.h
+++ b/systembus.hpp
@@ -1,7 +1,7 @@
-#include <stdint.h>
-#include <stddef.h>
+#include <cstdint>
+#include <cstddef>
 
-#include "type.h"
+#include "type.hpp"
 
 struct system_reg {
   reg32 C2DSTAT;           /* CH2-DMA destination address */
diff --git a/systembus_bits.h b/systembus_bits.hpp
similarity index 100%
rename from systembus_bits.h
rename to systembus_bits.hpp
diff --git a/test.c b/test.c
index 96b70a2..30c7e68 100644
--- a/test.c
+++ b/test.c
@@ -1,9 +1,9 @@
-#include "type.h"
+#include "type.hpp"
 
-#include "rgb.h"
-#include "vga.h"
-#include "systembus.h"
-#include "holly.h"
+#include "rgb.hpp"
+#include "vga.hpp"
+#include "systembus.hpp"
+#include "holly.hpp"
 
 void *memcpy(void *restrict dest, const void *restrict src, size_t n)
 {
diff --git a/type.h b/type.hpp
similarity index 85%
rename from type.h
rename to type.hpp
index bc79def..fad568b 100644
--- a/type.h
+++ b/type.hpp
@@ -1,5 +1,7 @@
-#include <stddef.h>
-#include <stdint.h>
+#pragma once
+
+#include <cstddef>
+#include <cstdint>
 
 #ifndef __cplusplus
 #ifndef static_assert
diff --git a/vga.cpp b/vga.cpp
index aaaa3e7..efb8254 100644
--- a/vga.cpp
+++ b/vga.cpp
@@ -1,14 +1,14 @@
-#include <stdint.h>
+#include <cstdint>
 
-#include "sh7091.h"
-#include "sh7091_bits.h"
-#include "holly.h"
-#include "holly/core_bits.h"
-#include "aica.h"
-#include "memorymap.h"
+#include "sh7091.hpp"
+#include "sh7091_bits.hpp"
+#include "holly.hpp"
+#include "holly/core_bits.hpp"
+#include "aica.hpp"
+#include "memorymap.hpp"
 
-#include "vga.h"
-#include "rgb.h"
+#include "vga.hpp"
+#include "rgb.hpp"
 
 uint32_t get_cable_type()
 {
@@ -144,7 +144,7 @@ void vga()
   holly.SOFTRESET = 0;
 }
 
-void fill_framebuffer()
+void vga_fill_framebuffer()
 {
   volatile uint16_t * vram = reinterpret_cast<volatile uint16_t *>(texture_memory);
   for (int y = 0; y < 480; y++) {
@@ -154,6 +154,4 @@ void fill_framebuffer()
       vram[y * 640 + x] = ((rgb.r >> 3) << 11) | ((rgb.g >> 2) << 5) | ((rgb.b >> 3) << 0);
     }
   }
-  vram[0] = 0xf0ff;
-  vram[10] = 0xf0ff;
 }
diff --git a/vga.h b/vga.hpp
similarity index 68%
rename from vga.h
rename to vga.hpp
index 3555293..41061b7 100644
--- a/vga.h
+++ b/vga.hpp
@@ -3,3 +3,4 @@
 void vga();
 void v_sync_in();
 void v_sync_out();
+void vga_fill_framebuffer();