diff --git a/example/clipping.cpp b/example/clipping.cpp
index 59645f5..40e3335 100644
--- a/example/clipping.cpp
+++ b/example/clipping.cpp
@@ -50,6 +50,7 @@ void do_get_condition()
   }
   maple::dma_start(send_buf, writer.send_offset,
                    recv_buf, writer.recv_offset);
+  maple::dma_wait_complete();
 
   for (uint8_t port = 0; port < 4; port++) {
     auto& bus_data = host_response[port].bus_data;
diff --git a/example/clipping2.cpp b/example/clipping2.cpp
index 62ce420..051f76b 100644
--- a/example/clipping2.cpp
+++ b/example/clipping2.cpp
@@ -53,6 +53,7 @@ void do_get_condition()
   }
   maple::dma_start(send_buf, writer.send_offset,
                    recv_buf, writer.recv_offset);
+  maple::dma_wait_complete();
 
   for (uint8_t port = 0; port < 4; port++) {
     auto& bus_data = host_response[port].bus_data;
diff --git a/example/clipping_textured.cpp b/example/clipping_textured.cpp
index 614c071..586e476 100644
--- a/example/clipping_textured.cpp
+++ b/example/clipping_textured.cpp
@@ -55,6 +55,7 @@ void do_get_condition()
   }
   maple::dma_start(send_buf, writer.send_offset,
                    recv_buf, writer.recv_offset);
+  maple::dma_wait_complete();
 
   for (uint8_t port = 0; port < 4; port++) {
     auto& bus_data = host_response[port].bus_data;
diff --git a/example/example.mk b/example/example.mk
index 8bc1801..ee5b394 100644
--- a/example/example.mk
+++ b/example/example.mk
@@ -43,11 +43,47 @@ MACAW_TWIDDLE_OBJ = \
 	holly/region_array.o \
 	holly/background.o \
 	holly/ta_fifo_polygon_converter.o \
-	macaw.data.o
+	texture/macaw/macaw.data.o
 
 example/macaw_twiddle.elf: LDSCRIPT = $(LIB)/main.lds
 example/macaw_twiddle.elf: $(START_OBJ) $(MACAW_TWIDDLE_OBJ)
 
+MACAW_CUBE_OBJ = \
+	example/macaw_cube.o \
+	holly/video_output.o \
+	holly/core.o \
+	holly/region_array.o \
+	holly/background.o \
+	holly/ta_fifo_polygon_converter.o \
+	texture/macaw/macaw.data.o
+
+example/macaw_cube.elf: LDSCRIPT = $(LIB)/main.lds
+example/macaw_cube.elf: $(START_OBJ) $(MACAW_CUBE_OBJ)
+
+MACAW_CUBE_RENDER_TO_TEXTURE_OBJ = \
+	example/macaw_cube_render_to_texture.o \
+	holly/video_output.o \
+	holly/core.o \
+	holly/region_array.o \
+	holly/background.o \
+	holly/ta_fifo_polygon_converter.o \
+	texture/macaw/macaw.data.o
+
+example/macaw_cube_render_to_texture.elf: LDSCRIPT = $(LIB)/main.lds
+example/macaw_cube_render_to_texture.elf: $(START_OBJ) $(MACAW_CUBE_RENDER_TO_TEXTURE_OBJ)
+
+MACAW_MULTIPASS_OBJ = \
+	example/macaw_multipass.o \
+	holly/video_output.o \
+	holly/core.o \
+	holly/region_array.o \
+	holly/background.o \
+	holly/ta_fifo_polygon_converter.o \
+	texture/macaw/macaw.data.o
+
+example/macaw_multipass.elf: LDSCRIPT = $(LIB)/main.lds
+example/macaw_multipass.elf: $(START_OBJ) $(MACAW_MULTIPASS_OBJ)
+
 FONT_BITMAP_OBJ = \
 	example/font_bitmap.o \
 	holly/video_output.o \
@@ -67,7 +103,7 @@ FONT_OUTLINE_OBJ = \
 	holly/region_array.o \
 	holly/background.o \
 	holly/ta_fifo_polygon_converter.o \
-	dejavusansmono.data.o
+	font/dejavusansmono/dejavusansmono.data.o
 
 example/font_outline.elf: LDSCRIPT = $(LIB)/main.lds
 example/font_outline.elf: $(START_OBJ) $(FONT_OUTLINE_OBJ)
@@ -79,23 +115,11 @@ FONT_OUTLINE_PUNCH_THROUGH_OBJ = \
 	holly/region_array.o \
 	holly/background.o \
 	holly/ta_fifo_polygon_converter.o \
-	dejavusansmono_mono.data.o
+	font/dejavusansmono/dejavusansmono_mono.data.o
 
 example/font_outline_punch_through.elf: LDSCRIPT = $(LIB)/main.lds
 example/font_outline_punch_through.elf: $(START_OBJ) $(FONT_OUTLINE_PUNCH_THROUGH_OBJ)
 
-MACAW_MULTIPASS_OBJ = \
-	example/macaw_multipass.o \
-	holly/video_output.o \
-	holly/core.o \
-	holly/region_array.o \
-	holly/background.o \
-	holly/ta_fifo_polygon_converter.o \
-	macaw.data.o
-
-example/macaw_multipass.elf: LDSCRIPT = $(LIB)/main.lds
-example/macaw_multipass.elf: $(START_OBJ) $(MACAW_MULTIPASS_OBJ)
-
 TRANSLUCENCY_OBJ = \
 	example/translucency.o \
 	holly/video_output.o \
@@ -103,7 +127,7 @@ TRANSLUCENCY_OBJ = \
 	holly/region_array.o \
 	holly/background.o \
 	holly/ta_fifo_polygon_converter.o \
-	macaw.data.o
+	texture/macaw/macaw.data.o
 
 example/translucency.elf: LDSCRIPT = $(LIB)/main.lds
 example/translucency.elf: $(START_OBJ) $(TRANSLUCENCY_OBJ)
@@ -174,8 +198,8 @@ MODIFIER_VOLUME_WITH_TWO_VOLUMES_OBJ = \
 	holly/region_array.o \
 	holly/background.o \
 	holly/ta_fifo_polygon_converter.o \
-	wolf.data.o \
-	macaw.data.o \
+	texture/wolf/wolf.data.o \
+	texture/macaw/macaw.data.o \
 	maple/maple.o \
 	$(LIBGCC)
 
@@ -205,30 +229,6 @@ VIEWING_SYSTEM_OBJ = \
 example/viewing_system.elf: LDSCRIPT = $(LIB)/main.lds
 example/viewing_system.elf: $(START_OBJ) $(VIEWING_SYSTEM_OBJ)
 
-MACAW_CUBE_OBJ = \
-	example/macaw_cube.o \
-	holly/video_output.o \
-	holly/core.o \
-	holly/region_array.o \
-	holly/background.o \
-	holly/ta_fifo_polygon_converter.o \
-	macaw.data.o
-
-example/macaw_cube.elf: LDSCRIPT = $(LIB)/main.lds
-example/macaw_cube.elf: $(START_OBJ) $(MACAW_CUBE_OBJ)
-
-MACAW_CUBE_RENDER_TO_TEXTURE_OBJ = \
-	example/macaw_cube_render_to_texture.o \
-	holly/video_output.o \
-	holly/core.o \
-	holly/region_array.o \
-	holly/background.o \
-	holly/ta_fifo_polygon_converter.o \
-	macaw.data.o
-
-example/macaw_cube_render_to_texture.elf: LDSCRIPT = $(LIB)/main.lds
-example/macaw_cube_render_to_texture.elf: $(START_OBJ) $(MACAW_CUBE_RENDER_TO_TEXTURE_OBJ)
-
 CLIPPING_OBJ = \
 	example/clipping.o \
 	holly/video_output.o \
diff --git a/example/font_outline.cpp b/example/font_outline.cpp
index d01ebf0..b630559 100644
--- a/example/font_outline.cpp
+++ b/example/font_outline.cpp
@@ -18,6 +18,7 @@
 #include "holly/region_array.hpp"
 #include "twiddle.hpp"
 #include "palette.hpp"
+#include "sh7091/serial.hpp"
 
 #include "font/font.hpp"
 #include "font/dejavusansmono/dejavusansmono.data.h"
@@ -72,7 +73,7 @@ uint32_t transform(ta_parameter_writer& parameter,
                                         | tsp_instruction_word::texture_v_size::from_int(texture_height);
 
     const uint32_t texture_address = texture_memory_alloc::texture.start;
-    const uint32_t texture_control_word = texture_control_word::pixel_format::_4bpp_palette
+    const uint32_t texture_control_word = texture_control_word::pixel_format::_8bpp_palette
                                         | texture_control_word::scan_order::twiddled
                                         | texture_control_word::texture_address(texture_address / 8);
 
@@ -140,7 +141,7 @@ uint32_t transform2(ta_parameter_writer& parameter,
                                       | tsp_instruction_word::texture_v_size::from_int(texture_height);
 
   const uint32_t texture_address = texture_memory_alloc::texture.start;
-  const uint32_t texture_control_word = texture_control_word::pixel_format::_4bpp_palette
+  const uint32_t texture_control_word = texture_control_word::pixel_format::_8bpp_palette
                                       | texture_control_word::scan_order::twiddled
                                       | texture_control_word::texture_address(texture_address / 8);
 
@@ -189,15 +190,24 @@ void init_texture_memory(const struct opb_size& opb_size)
   background_parameter(0xff0000ff);
 }
 
-void inflate_font(const uint32_t * src,
-                  const uint32_t stride,
-                  const uint32_t curve_end_ix)
+void inflate_font(const uint8_t * src,
+                  const uint32_t width,
+                  const uint32_t height)
 {
-  auto texture = reinterpret_cast<volatile uint16_t *>(&texture_memory64[texture_memory_alloc::texture.start / 4]);
+  auto texture = reinterpret_cast<volatile uint32_t *>(&texture_memory64[texture_memory_alloc::texture.start / 4]);
 
-  twiddle::texture3<4, 8>(texture, reinterpret_cast<const uint8_t *>(src),
-                          stride,
-                          curve_end_ix);
+  const uint32_t size = width * height;
+
+  uint8_t temp[size] __attribute__((aligned(4)));
+
+  twiddle::texture(temp,
+		   src,
+		   width,
+		   height);
+
+  for (uint32_t i = 0; i < size / 4; i++) {
+    texture[i] = reinterpret_cast<uint32_t *>(temp)[i];
+  }
 }
 
 uint32_t _ta_parameter_buf[((32 * 10 * 17) + 32) / 4];
@@ -206,14 +216,13 @@ void main()
 {
   video_output::set_mode_vga();
 
-  auto font = reinterpret_cast<const struct font *>(&_binary_dejavusansmono_data_start);
+  auto font = reinterpret_cast<const struct font *>(&_binary_font_dejavusansmono_dejavusansmono_data_start);
   auto glyphs = reinterpret_cast<const struct glyph *>(&font[1]);
-  auto texture = reinterpret_cast<const uint32_t *>(&glyphs[font->glyph_count]);
+  auto texture = reinterpret_cast<const uint8_t *>(&glyphs[font->glyph_count]);
 
   /*
   serial::integer<uint32_t>(font->first_char_code);
   serial::integer<uint32_t>(font->glyph_count);
-  serial::integer<uint32_t>(font->glyph_height);
   serial::integer<uint32_t>(font->texture_stride);
   serial::integer<uint32_t>(font->texture_width);
   serial::integer<uint32_t>(font->texture_height);
@@ -223,9 +232,9 @@ void main()
   */
 
   inflate_font(texture,
-               font->texture_stride,
-               font->max_z_curve_ix);
-  palette_data<16>();
+               font->texture_width,
+	       font->texture_height);
+  palette_data<256>();
 
   // The address of `ta_parameter_buf` must be a multiple of 32 bytes.
   // This is mandatory for ch2-dma to the ta fifo polygon converter.
@@ -272,14 +281,14 @@ void main()
               font->first_char_code,
               glyphs,
               ana, 17,
-              font->glyph_height * 0);
+              font->face_metrics.height * 0);
 
     transform(parameter,
               font->texture_width, font->texture_height,
               font->first_char_code,
               glyphs,
               cabal, 26,
-              font->glyph_height * 1);
+              font->face_metrics.height * 1);
 
     parameter.append<ta_global_parameter::end_of_list>() = ta_global_parameter::end_of_list(para_control::para_type::end_of_list);
 
diff --git a/example/font_outline_punch_through.cpp b/example/font_outline_punch_through.cpp
index a14dcb6..62d073a 100644
--- a/example/font_outline_punch_through.cpp
+++ b/example/font_outline_punch_through.cpp
@@ -20,7 +20,7 @@
 #include "palette.hpp"
 
 #include "font/font.hpp"
-#include "dejavusansmono_mono.hpp"
+#include "font/dejavusansmono/dejavusansmono_mono.data.h"
 
 struct vertex {
   float x;
@@ -71,7 +71,7 @@ uint32_t transform(ta_parameter_writer& parameter,
                                         | tsp_instruction_word::texture_v_size::from_int(texture_height);
 
     const uint32_t texture_address = texture_memory_alloc::texture.start;
-    const uint32_t texture_control_word = texture_control_word::pixel_format::_4bpp_palette
+    const uint32_t texture_control_word = texture_control_word::pixel_format::_8bpp_palette
                                         | texture_control_word::scan_order::twiddled
                                         | texture_control_word::texture_address(texture_address / 8);
 
@@ -139,7 +139,7 @@ uint32_t transform2(ta_parameter_writer& parameter,
                                       | tsp_instruction_word::texture_v_size::from_int(texture_height);
 
   const uint32_t texture_address = texture_memory_alloc::texture.start;
-  const uint32_t texture_control_word = texture_control_word::pixel_format::_4bpp_palette
+  const uint32_t texture_control_word = texture_control_word::pixel_format::_8bpp_palette
                                       | texture_control_word::scan_order::twiddled
                                       | texture_control_word::texture_address(texture_address / 8);
 
@@ -188,15 +188,39 @@ void init_texture_memory(const struct opb_size& opb_size)
   background_parameter(0xff0000ff);
 }
 
-void inflate_font(const uint32_t * src,
-                  const uint32_t stride,
-                  const uint32_t curve_end_ix)
+void inflate_font(const uint8_t * src,
+                  const uint32_t width,
+                  const uint32_t height)
 {
-  auto texture = reinterpret_cast<volatile uint16_t *>(&texture_memory64[texture_memory_alloc::texture.start / 4]);
+  auto texture = reinterpret_cast<volatile uint32_t *>(&texture_memory64[texture_memory_alloc::texture.start / 4]);
 
-  twiddle::texture3<4, 1>(texture, reinterpret_cast<const uint8_t *>(src),
-                          stride,
-                          curve_end_ix);
+  const uint32_t bits_per_pixel = 1;
+  const uint32_t pixels_per_byte = 8;
+  const uint32_t stride = width / pixels_per_byte;
+  const uint32_t mask = (1 << bits_per_pixel) - 1;
+
+  const uint32_t size = width * height;
+
+  uint8_t temp1[size];
+  for (uint32_t y = 0; y < height; y++) {
+    for (uint32_t x = 0; x < width; x++) {
+      const uint32_t texture_ix = y * stride + x / pixels_per_byte;
+      const uint32_t texture_ix_mod = x % pixels_per_byte;
+      const uint32_t value = (src[texture_ix] >> (bits_per_pixel * texture_ix_mod)) & mask;
+      temp1[y * width + x] = value;
+    }
+  }
+
+  uint8_t temp[size] __attribute__((aligned(4)));
+
+  twiddle::texture(temp,
+		   temp1,
+		   width,
+		   height);
+
+  for (uint32_t i = 0; i < size / 4; i++) {
+    texture[i] = reinterpret_cast<uint32_t *>(temp)[i];
+  }
 }
 
 uint32_t _ta_parameter_buf[((32 * 10 * 17) + 32) / 4];
@@ -205,9 +229,9 @@ void main()
 {
   video_output::set_mode_vga();
 
-  auto font = reinterpret_cast<const struct font *>(&_binary_dejavusansmono_mono_data_start);
+  auto font = reinterpret_cast<const struct font *>(&_binary_font_dejavusansmono_dejavusansmono_mono_data_start);
   auto glyphs = reinterpret_cast<const struct glyph *>(&font[1]);
-  auto texture = reinterpret_cast<const uint32_t *>(&glyphs[font->glyph_count]);
+  auto texture = reinterpret_cast<const uint8_t *>(&glyphs[font->glyph_count]);
 
   /*
   serial::integer<uint32_t>(font->first_char_code);
@@ -222,8 +246,8 @@ void main()
   */
 
   inflate_font(texture,
-               font->texture_stride,
-               font->max_z_curve_ix);
+               font->texture_width,
+	       font->texture_height);
   palette_data<2>();
 
   // The address of `ta_parameter_buf` must be a multiple of 32 bytes.
@@ -274,14 +298,14 @@ void main()
               font->first_char_code,
               glyphs,
               ana, 17,
-              font->glyph_height * 0);
+              font->face_metrics.height * 0);
 
     transform(parameter,
               font->texture_width, font->texture_height,
               font->first_char_code,
               glyphs,
               cabal, 26,
-              font->glyph_height * 1);
+              font->face_metrics.height * 1);
 
     parameter.append<ta_global_parameter::end_of_list>() = ta_global_parameter::end_of_list(para_control::para_type::end_of_list);
 
diff --git a/example/heart.cpp b/example/heart.cpp
index 0a24f37..1d1dc7a 100644
--- a/example/heart.cpp
+++ b/example/heart.cpp
@@ -205,7 +205,7 @@ void transform_model(ta_parameter_writer& parameter,
       c.g += color.g * 0.1;
       c.b += color.b * 0.1;
       if (n_dot_l > 0) {
-	float intensity = n_dot_l / (length(n) * length(l));
+	float intensity = n_dot_l / (magnitude(n) * magnitude(l));
         c.r += color.r * intensity * lighting_weight;
         c.g += color.g * intensity * lighting_weight;
         c.b += color.b * intensity * lighting_weight;
diff --git a/example/macaw_cube.cpp b/example/macaw_cube.cpp
index 05abd55..e757505 100644
--- a/example/macaw_cube.cpp
+++ b/example/macaw_cube.cpp
@@ -17,7 +17,7 @@
 #include "holly/background.hpp"
 #include "memorymap.hpp"
 
-#include "macaw.hpp"
+#include "texture/macaw/macaw.data.h"
 
 struct vertex {
   float x;
@@ -145,8 +145,8 @@ void init_texture_memory(const struct opb_size& opb_size)
 
 void copy_macaw_texture()
 {
-  auto src = reinterpret_cast<const uint8_t *>(&_binary_macaw_data_start);
-  auto size  = reinterpret_cast<const uint32_t>(&_binary_macaw_data_size);
+  auto src = reinterpret_cast<const uint8_t *>(&_binary_texture_macaw_macaw_data_start);
+  auto size  = reinterpret_cast<const uint32_t>(&_binary_texture_macaw_macaw_data_size);
   auto texture = reinterpret_cast<volatile uint16_t *>(&texture_memory64[texture_memory_alloc::texture.start / 4]);
   for (uint32_t px = 0; px < size / 3; px++) {
     uint8_t r = src[px * 3 + 0];
diff --git a/example/macaw_cube_render_to_texture.cpp b/example/macaw_cube_render_to_texture.cpp
index 5210f04..8e01c3c 100644
--- a/example/macaw_cube_render_to_texture.cpp
+++ b/example/macaw_cube_render_to_texture.cpp
@@ -17,7 +17,7 @@
 #include "holly/background.hpp"
 #include "memorymap.hpp"
 
-#include "macaw.hpp"
+#include "texture/macaw/macaw.data.h"
 
 struct vertex {
   float x;
@@ -178,8 +178,8 @@ void transform(ta_parameter_writer& parameter,
 
 void copy_macaw_texture()
 {
-  auto src = reinterpret_cast<const uint8_t *>(&_binary_macaw_data_start);
-  auto size  = reinterpret_cast<const uint32_t>(&_binary_macaw_data_size);
+  auto src = reinterpret_cast<const uint8_t *>(&_binary_texture_macaw_macaw_data_start);
+  auto size  = reinterpret_cast<const uint32_t>(&_binary_texture_macaw_macaw_data_size);
   auto texture = reinterpret_cast<volatile uint16_t *>(&texture_memory64[texture_memory_alloc::texture.start / 4]);
   for (uint32_t px = 0; px < size / 3; px++) {
     uint8_t r = src[px * 3 + 0];
diff --git a/example/macaw_multipass.cpp b/example/macaw_multipass.cpp
index 9b26f75..8f87f08 100644
--- a/example/macaw_multipass.cpp
+++ b/example/macaw_multipass.cpp
@@ -17,7 +17,7 @@
 #include "holly/video_output.hpp"
 #include "memorymap.hpp"
 
-#include "macaw.hpp"
+#include "texture/macaw/macaw.data.h"
 
 struct vertex {
   float x;
@@ -134,11 +134,12 @@ uint32_t transform(uint32_t * ta_parameter_buf,
 
 void init_texture_memory(const struct opb_size * opb_size, uint32_t render_passes)
 {
-  region_array_multipass(640 / 32, // width
-			 480 / 32, // height
+  region_array_multipass(640 / 32, // tile_width
+			 480 / 32, // tile_height
 			 opb_size,
-			 render_passes
-			 );
+			 render_passes,
+			 texture_memory_alloc::region_array.start,
+			 texture_memory_alloc::object_list.start);
 
   background_parameter(0xff00ff00);
 }
@@ -147,8 +148,8 @@ uint32_t _ta_parameter_buf[((32 * (strip_length + 2)) + 32) / 4];
 
 void copy_macaw_texture()
 {
-  auto src = reinterpret_cast<const uint8_t *>(&_binary_macaw_data_start);
-  auto size  = reinterpret_cast<const uint32_t>(&_binary_macaw_data_size);
+  auto src = reinterpret_cast<const uint8_t *>(&_binary_texture_macaw_macaw_data_start);
+  auto size  = reinterpret_cast<const uint32_t>(&_binary_texture_macaw_macaw_data_size);
   auto texture = reinterpret_cast<volatile uint16_t *>(&texture_memory64[texture_memory_alloc::texture.start / 4]);
   for (uint32_t px = 0; px < size / 3; px++) {
     uint8_t r = src[px * 3 + 0];
diff --git a/example/macaw_twiddle.cpp b/example/macaw_twiddle.cpp
index 76ef105..ee9a4e6 100644
--- a/example/macaw_twiddle.cpp
+++ b/example/macaw_twiddle.cpp
@@ -18,7 +18,7 @@
 #include "memorymap.hpp"
 #include "twiddle.hpp"
 
-#include "macaw.hpp"
+#include "texture/macaw/macaw.data.h"
 
 struct vertex {
   float x;
@@ -116,8 +116,8 @@ void init_texture_memory(const struct opb_size& opb_size)
 
 void init_macaw_texture()
 {
-  auto src = reinterpret_cast<const uint8_t *>(&_binary_macaw_data_start);
-  auto size  = reinterpret_cast<const uint32_t>(&_binary_macaw_data_size);
+  auto src = reinterpret_cast<const uint8_t *>(&_binary_texture_macaw_macaw_data_start);
+  auto size  = reinterpret_cast<const uint32_t>(&_binary_texture_macaw_macaw_data_size);
   auto texture = reinterpret_cast<volatile uint16_t *>(&texture_memory64[texture_memory_alloc::texture.start / 4]);
 
   uint16_t temp[size / 3];
diff --git a/example/maple_analog.cpp b/example/maple_analog.cpp
index 4257af4..d977017 100644
--- a/example/maple_analog.cpp
+++ b/example/maple_analog.cpp
@@ -30,16 +30,16 @@
 
 static ft0::data_transfer::data_format data[4];
 
+uint32_t send_buf[1024] __attribute__((aligned(32)));
+uint32_t recv_buf[1024] __attribute__((aligned(32)));
+
 void do_get_condition()
 {
-  uint32_t send_buf[1024] __attribute__((aligned(32)));
-  uint32_t recv_buf[1024] __attribute__((aligned(32)));
+  auto writer = maple::host_command_writer(send_buf, recv_buf);
 
   using command_type = maple::get_condition;
   using response_type = maple::data_transfer<ft0::data_transfer::data_format>;
 
-  auto writer = maple::host_command_writer(send_buf, recv_buf);
-
   auto [host_command, host_response]
     = writer.append_command_all_ports<command_type, response_type>();
 
@@ -47,6 +47,7 @@ void do_get_condition()
 
   maple::dma_start(send_buf, writer.send_offset,
                    recv_buf, writer.recv_offset);
+  maple::dma_wait_complete();
 
   for (uint8_t port = 0; port < 4; port++) {
     auto& bus_data = host_response[port].bus_data;
@@ -58,8 +59,8 @@ void do_get_condition()
       return;
     }
 
-    data[port].analog_axis_3 = data_fields.data.analog_axis_3;
-    data[port].analog_axis_4 = data_fields.data.analog_axis_4;
+    data[port].analog_coordinate_axis[2] = data_fields.data.analog_coordinate_axis[2];
+    data[port].analog_coordinate_axis[3] = data_fields.data.analog_coordinate_axis[3];
   }
 }
 
@@ -188,8 +189,8 @@ void main()
 			      640 / 32,
 			      480 / 32);
 
-    float x_pos = static_cast<float>(data[0].analog_axis_3 - 0x80) * (0.5 / 127);
-    float y_pos = static_cast<float>(data[0].analog_axis_4 - 0x80) * (0.5 / 127);
+    float x_pos = static_cast<float>(data[0].analog_coordinate_axis[2] - 0x80) * (0.5 / 127);
+    float y_pos = static_cast<float>(data[0].analog_coordinate_axis[3] - 0x80) * (0.5 / 127);
 
     auto parameter = ta_parameter_writer(ta_parameter_buf);
     for (uint32_t i = 0; i < border::num_faces; i++) {
diff --git a/example/maple_controller.cpp b/example/maple_controller.cpp
index 3dd21e0..cc44c2c 100644
--- a/example/maple_controller.cpp
+++ b/example/maple_controller.cpp
@@ -21,8 +21,8 @@ void do_get_condition()
 
   auto writer = maple::host_command_writer(send_buf, recv_buf);
 
-  using command_type = get_condition;
-  using response_type = data_transfer<ft0::data_transfer::data_format>;
+  using command_type = maple::get_condition;
+  using response_type = maple::data_transfer<ft0::data_transfer::data_format>;
 
   auto [host_command, host_response]
     = writer.append_command_all_ports<command_type, response_type>();
@@ -31,6 +31,7 @@ void do_get_condition()
 
   maple::dma_start(send_buf, writer.send_offset,
                    recv_buf, writer.recv_offset);
+  maple::dma_wait_complete();
 
   for (uint8_t port = 0; port < 4; port++) {
     auto& bus_data = host_response[port].bus_data;
@@ -61,8 +62,8 @@ void do_lm_request(uint8_t port, uint8_t lm)
   uint32_t host_port_select = host_instruction_port_select(port);
   uint32_t destination_ap = ap_port_select(port) | ap::de::expansion_device | lm;
 
-  using command_type = device_request;
-  using response_type = device_status;
+  using command_type = maple::device_request;
+  using response_type = maple::device_status;
 
   auto [host_command, host_response]
     = writer.append_command<command_type, response_type>(host_port_select,
@@ -71,10 +72,11 @@ void do_lm_request(uint8_t port, uint8_t lm)
 
   maple::dma_start(send_buf, writer.send_offset,
                    recv_buf, writer.recv_offset);
+  maple::dma_wait_complete();
 
   auto& bus_data = host_response->bus_data;
   auto& data_fields = bus_data.data_fields;
-  if (bus_data.command_code != device_status::command_code) {
+  if (bus_data.command_code != maple::device_status::command_code) {
     serial::string("lm did not reply: ");
     serial::integer<uint8_t>(port, ' ');
     serial::integer<uint8_t>(lm);
@@ -115,19 +117,20 @@ void do_device_request()
 
   auto writer = maple::host_command_writer(send_buf, recv_buf);
 
-  using command_type = device_request;
-  using response_type = device_status;
+  using command_type = maple::device_request;
+  using response_type = maple::device_status;
 
   auto [host_command, host_response]
     = writer.append_command_all_ports<command_type, response_type>();
 
   maple::dma_start(send_buf, writer.send_offset,
                    recv_buf, writer.recv_offset);
+  maple::dma_wait_complete();
 
   for (uint8_t port = 0; port < 4; port++) {
     auto& bus_data = host_response[port].bus_data;
     auto& data_fields = bus_data.data_fields;
-    if (bus_data.command_code != device_status::command_code) {
+    if (bus_data.command_code != maple::device_status::command_code) {
       serial::string("port: ");
       serial::integer<uint8_t>(port);
       serial::string("  disconnected\n");
@@ -153,6 +156,7 @@ void do_device_request()
 
 void main()
 {
+  serial::init(0);
   // flycast needs this in HLE mode, or else it won't start the vcount
   // counter.
   video_output::set_mode_vga();
diff --git a/example/maple_device_request.cpp b/example/maple_device_request.cpp
index 68ffb6e..0aed218 100644
--- a/example/maple_device_request.cpp
+++ b/example/maple_device_request.cpp
@@ -7,7 +7,7 @@
 
 void main()
 {
-  serial::init(4);
+  serial::init(0);
 
   uint32_t send_buf[1024] __attribute__((aligned(32)));
   uint32_t recv_buf[1024] __attribute__((aligned(32)));
@@ -22,6 +22,7 @@ void main()
 
   maple::dma_start(send_buf, writer.send_offset,
                    recv_buf, writer.recv_offset);
+  maple::dma_wait_complete();
 
   for (uint8_t port = 0; port < 4; port++) {
     auto& bus_data = host_response[port].bus_data;
diff --git a/example/maple_font.cpp b/example/maple_font.cpp
index df9f5c1..5b9231e 100644
--- a/example/maple_font.cpp
+++ b/example/maple_font.cpp
@@ -72,6 +72,7 @@ void send_vmu_framebuffer(uint8_t port, uint8_t lm)
 
   maple::dma_start(send_buf, writer.send_offset,
 		   recv_buf, writer.recv_offset);
+  maple::dma_wait_complete();
 
   serial::integer<uint8_t>(host_response->bus_data.command_code);
   serial::integer<uint8_t>(host_response->bus_data.destination_ap);
@@ -99,6 +100,7 @@ void do_lm_request(uint8_t port, uint8_t lm)
 
   maple::dma_start(send_buf, writer.send_offset,
                    recv_buf, writer.recv_offset);
+  maple::dma_wait_complete();
 
   auto& bus_data = host_response->bus_data;
   auto& data_fields = bus_data.data_fields;
@@ -156,6 +158,7 @@ void do_device_request()
 
   maple::dma_start(send_buf, writer.send_offset,
                    recv_buf, writer.recv_offset);
+  maple::dma_wait_complete();
 
   for (uint8_t port = 0; port < 4; port++) {
     auto& bus_data = host_response[port].bus_data;
diff --git a/example/maple_mouse.cpp b/example/maple_mouse.cpp
index 900a34c..0559c8f 100644
--- a/example/maple_mouse.cpp
+++ b/example/maple_mouse.cpp
@@ -49,6 +49,7 @@ void do_get_condition()
 
   maple::dma_start(send_buf, writer.send_offset,
                    recv_buf, writer.recv_offset);
+  maple::dma_wait_complete();
 
   for (uint8_t port = 0; port < 4; port++) {
     auto& bus_data = host_response[port].bus_data;
diff --git a/example/maple_storage.cpp b/example/maple_storage.cpp
index a138ce4..2fcca49 100644
--- a/example/maple_storage.cpp
+++ b/example/maple_storage.cpp
@@ -146,6 +146,7 @@ bool do_block_read(storage_state& state, uint16_t block_number, uint8_t * dest)
 
   maple::dma_start(state.send_buf, writer.send_offset,
 		   state.recv_buf, writer.recv_offset);
+  maple::dma_wait_complete();
 
   for (uint32_t phase = 0; phase < state.device_status.read_accesses; phase++) {
     auto& bus_data = host_responses[phase]->bus_data;
@@ -227,6 +228,7 @@ bool do_block_write(storage_state& state, uint16_t block_number, uint8_t * src)
 
   maple::dma_start(state.send_buf, writer.send_offset,
 		   state.recv_buf, writer.recv_offset);
+  maple::dma_wait_complete();
 
   serial::string("block write status: ");
   serial::integer(host_response->bus_data.command_code);
@@ -316,6 +318,7 @@ void do_lm_request(uint8_t port, uint8_t lm)
 
     maple::dma_start(send_buf, writer.send_offset,
 		     recv_buf, writer.recv_offset);
+    maple::dma_wait_complete();
 
     auto& bus_data = host_response->bus_data;
     auto& data_fields = bus_data.data_fields;
@@ -361,6 +364,7 @@ void do_lm_request(uint8_t port, uint8_t lm)
 
     maple::dma_start(send_buf, writer.send_offset,
 		     recv_buf, writer.recv_offset);
+    maple::dma_wait_complete();
 
     auto& bus_data = host_response->bus_data;
     auto& data_fields = bus_data.data_fields;
@@ -556,6 +560,7 @@ void do_device_request()
 
   maple::dma_start(send_buf, writer.send_offset,
                    recv_buf, writer.recv_offset);
+  maple::dma_wait_complete();
 
   for (uint8_t port = 0; port < 4; port++) {
     auto& bus_data = host_response[port].bus_data;
diff --git a/example/maple_vibrator.cpp b/example/maple_vibrator.cpp
index 9ae5fbb..97c9bfc 100644
--- a/example/maple_vibrator.cpp
+++ b/example/maple_vibrator.cpp
@@ -63,6 +63,7 @@ void do_lm_request(uint8_t port, uint8_t lm)
   serial::string("dma start\n");
   maple::dma_start(command_buf, maple::sizeof_command(host_command),
                    receive_buf, maple::sizeof_command(host_response));
+  maple::dma_wait_complete();
 
   auto& bus_data = host_response->bus_data;
   if (bus_data.command_code != response_type::command_code) {
@@ -123,6 +124,7 @@ void do_lm_request(uint8_t port, uint8_t lm)
     auto host_response = reinterpret_cast<host_response_type *>(receive_buf);
     maple::dma_start(command_buf, maple::sizeof_command(host_command),
                      receive_buf, maple::sizeof_command(host_response));
+    maple::dma_wait_complete();
 
     auto& bus_data = host_response->bus_data;
 
@@ -160,6 +162,7 @@ void do_device_request()
   const uint32_t command_size = maple::init_host_command_all_ports<command_type, response_type>(command_buf, receive_buf);
   maple::dma_start(command_buf, command_size,
                    receive_buf, maple::sizeof_command(host_response) * 4);
+  maple::dma_wait_complete();
 
   for (uint8_t port = 0; port < 4; port++) {
     auto& bus_data = host_response[port].bus_data;
diff --git a/example/maple_wink.cpp b/example/maple_wink.cpp
index 30817c4..87228f7 100644
--- a/example/maple_wink.cpp
+++ b/example/maple_wink.cpp
@@ -76,6 +76,7 @@ void send_wink(uint8_t port, uint8_t lm)
 
   maple::dma_start(send_buf, writer.send_offset,
 		   recv_buf, writer.recv_offset);
+  maple::dma_wait_complete();
 
   serial::integer<uint8_t>(host_response->bus_data.command_code);
   serial::integer<uint8_t>(host_response->bus_data.destination_ap);
@@ -103,6 +104,7 @@ void do_lm_request(uint8_t port, uint8_t lm)
 
   maple::dma_start(send_buf, writer.send_offset,
                    recv_buf, writer.recv_offset);
+  maple::dma_wait_complete();
 
   auto& bus_data = host_response->bus_data;
   auto& data_fields = bus_data.data_fields;
@@ -160,6 +162,7 @@ void do_device_request()
 
   maple::dma_start(send_buf, writer.send_offset,
                    recv_buf, writer.recv_offset);
+  maple::dma_wait_complete();
 
   for (uint8_t port = 0; port < 4; port++) {
     auto& bus_data = host_response[port].bus_data;
diff --git a/example/modifier_volume_with_two_volumes.cpp b/example/modifier_volume_with_two_volumes.cpp
index 9b5d438..e9bd09e 100644
--- a/example/modifier_volume_with_two_volumes.cpp
+++ b/example/modifier_volume_with_two_volumes.cpp
@@ -52,6 +52,7 @@ void do_get_condition(uint32_t * command_buf,
 
   maple::dma_start(command_buf, command_size,
                    receive_buf, maple::sizeof_command(host_response));
+  maple::dma_wait_complete();
 
   for (uint8_t port = 0; port < 4; port++) {
     auto& bus_data = host_response[port].bus_data;
diff --git a/example/texture_filtering_maple.cpp b/example/texture_filtering_maple.cpp
index 1684c59..5e4c9d0 100644
--- a/example/texture_filtering_maple.cpp
+++ b/example/texture_filtering_maple.cpp
@@ -249,6 +249,7 @@ void do_get_condition(uint32_t * command_buf,
 
   maple::dma_start(command_buf, command_size,
                    receive_buf, maple::sizeof_command(host_response));
+  maple::dma_wait_complete();
 
   buttons.reset();
 
diff --git a/font/dejavusansmono/dejavusansmono.data b/font/dejavusansmono/dejavusansmono.data
index a072ce7..7cafa02 100644
Binary files a/font/dejavusansmono/dejavusansmono.data and b/font/dejavusansmono/dejavusansmono.data differ
diff --git a/font/dejavusansmono/dejavusansmono_mono.data.h b/font/dejavusansmono/dejavusansmono_mono.data.h
new file mode 100644
index 0000000..3025b64
--- /dev/null
+++ b/font/dejavusansmono/dejavusansmono_mono.data.h
@@ -0,0 +1,5 @@
+#pragma once
+#include <cstdint>
+extern uint32_t _binary_font_dejavusansmono_dejavusansmono_mono_data_start __asm("_binary_font_dejavusansmono_dejavusansmono_mono_data_start");
+extern uint32_t _binary_font_dejavusansmono_dejavusansmono_mono_data_end __asm("_binary_font_dejavusansmono_dejavusansmono_mono_data_end");
+extern uint32_t _binary_font_dejavusansmono_dejavusansmono_mono_data_size __asm("_binary_font_dejavusansmono_dejavusansmono_mono_data_size");
diff --git a/font/dejavusansmono/dejavusansmono_mono.hpp b/font/dejavusansmono/dejavusansmono_mono.hpp
deleted file mode 100644
index ca93a73..0000000
--- a/font/dejavusansmono/dejavusansmono_mono.hpp
+++ /dev/null
@@ -1,5 +0,0 @@
-#include <cstdint>
-
-extern uint32_t _binary_dejavusansmono_mono_data_start __asm("_binary_dejavusansmono_mono_data_start");
-extern uint32_t _binary_dejavusansmono_mono_data_end __asm("_binary_dejavusansmono_mono_data_end");
-extern uint32_t _binary_dejavusansmono_mono_data_size __asm("_binary_dejavusansmono_mono_data_size");
diff --git a/font/europc_mono_9x14/europc_mono_9x14.hpp b/font/europc_mono_9x14/europc_mono_9x14.hpp
deleted file mode 100644
index a2be09f..0000000
--- a/font/europc_mono_9x14/europc_mono_9x14.hpp
+++ /dev/null
@@ -1,7 +0,0 @@
-#pragma once
-
-#include <cstdint>
-
-extern uint32_t _binary_europc_mono_9x14_data_start __asm("_binary_europc_mono_9x14_data_start");
-extern uint32_t _binary_europc_mono_9x14_data_end __asm("_binary_europc_mono_9x14_data_end");
-extern uint32_t _binary_europc_mono_9x14_data_size __asm("_binary_europc_mono_9x14_data_size");
diff --git a/font/ter_u20n/ter_u20n.hpp b/font/ter_u20n/ter_u20n.hpp
deleted file mode 100644
index 151c6b9..0000000
--- a/font/ter_u20n/ter_u20n.hpp
+++ /dev/null
@@ -1,5 +0,0 @@
-#include <cstdint>
-
-extern uint32_t _binary_ter_u20n_data_start __asm("_binary_ter_u20n_data_start");
-extern uint32_t _binary_ter_u20n_data_end __asm("_binary_ter_u20n_data_end");
-extern uint32_t _binary_ter_u20n_data_size __asm("_binary_ter_u20n_data_size");
diff --git a/font/verite_8x16/verite_8x16.hpp b/font/verite_8x16/verite_8x16.hpp
deleted file mode 100644
index 5e5bc6c..0000000
--- a/font/verite_8x16/verite_8x16.hpp
+++ /dev/null
@@ -1,7 +0,0 @@
-#pragma once
-
-#include <cstdint>
-
-extern uint32_t _binary_verite_8x16_data_start __asm("_binary_verite_8x16_data_start");
-extern uint32_t _binary_verite_8x16_data_end __asm("_binary_verite_8x16_data_end");
-extern uint32_t _binary_verite_8x16_data_size __asm("_binary_verite_8x16_data_size");