diff --git a/shader_examples/mesa/gl_teximage2d.fs.txt b/shader_examples/mesa/gl_teximage2d.fs.txt
new file mode 100644
index 0000000..241efcf
--- /dev/null
+++ b/shader_examples/mesa/gl_teximage2d.fs.txt
@@ -0,0 +1,12 @@
+0x00007807,
+0x02400000,
+0xe400f400,
+0x00000000,
+0x00000000,
+0x00000000,
+0x00078005,
+0x08020000,
+0x08020000,
+0x1c440220,
+0x1c60c003,
+0x00000005,
diff --git a/shader_examples/mesa/gl_teximage2d.vs.txt b/shader_examples/mesa/gl_teximage2d.vs.txt
new file mode 100644
index 0000000..d45b68e
--- /dev/null
+++ b/shader_examples/mesa/gl_teximage2d.vs.txt
@@ -0,0 +1,8 @@
+0x00f00203,
+0x00d10001,
+0x01248001,
+0x01248001,
+0x00f02203,
+0x00d10021,
+0x01248021,
+0x01248021,
diff --git a/src/Makefile b/src/Makefile
index d4b3875..b16eb33 100644
--- a/src/Makefile
+++ b/src/Makefile
@@ -17,6 +17,7 @@ R500_COMMON = \
 	r500/display_controller.o \
 	r500/indirect_buffer.o \
 	r500/shader.o \
+	r500/texture.o \
 	drm/buffer.o \
 	drm/drm.o \
 	file.o
diff --git a/src/drm/buffer.c b/src/drm/buffer.c
index 7b0401d..e2270fd 100644
--- a/src/drm/buffer.c
+++ b/src/drm/buffer.c
@@ -72,32 +72,3 @@ int create_flush_buffer(int fd)
   assert(args.handle != 0);
   return args.handle;
 }
-
-int * load_textures(int fd,
-                    const char ** textures,
-                    int textures_length)
-{
-  int * texturebuffer_handle = (int *)malloc((sizeof (int)) * textures_length);
-
-  for (int i = 0; i < textures_length; i++) {
-    int size = 0;
-    void * buf = file_read(textures[i], &size);
-    assert(buf != NULL);
-
-    printf("load texture[%d]: %d\n", i, size);
-
-    void * ptr = NULL;
-    int handle = create_buffer(fd, size, &ptr);
-
-    for (int i = 0; i < size / 4; i++) {
-      ((uint32_t*)ptr)[i] = ((uint32_t*)buf)[i];
-    }
-    asm volatile ("" ::: "memory");
-    free(buf);
-    munmap(ptr, size);
-
-    texturebuffer_handle[i] = handle;
-  }
-
-  return texturebuffer_handle;
-}
diff --git a/src/drm/buffer.h b/src/drm/buffer.h
index 1535132..4405a87 100644
--- a/src/drm/buffer.h
+++ b/src/drm/buffer.h
@@ -6,9 +6,6 @@ extern "C" {
 
 int create_buffer(int fd, int buffer_size, void ** out_ptr);
 int create_flush_buffer(int fd);
-int * load_textures(int fd,
-                    const char ** textures,
-                    int textures_length);
 
 #ifdef __cplusplus
 }
diff --git a/src/drm/drm.c b/src/drm/drm.c
index 14233d6..f8ab256 100644
--- a/src/drm/drm.c
+++ b/src/drm/drm.c
@@ -37,7 +37,7 @@ void drm_radeon_cs(int fd,
   };
 
   for (int i = 0; i < texturebuffer_handles_length; i++) {
-    relocs[3 + i] = (struct drm_radeon_cs_reloc){
+    relocs[TEXTUREBUFFER_RELOC_INDEX + i] = (struct drm_radeon_cs_reloc){
       .handle = texturebuffer_handles[i],
       .read_domains = 4, // RADEON_GEM_DOMAIN_VRAM
       .write_domain = 4, // RADEON_GEM_DOMAIN_VRAM
diff --git a/src/drm/drm.h b/src/drm/drm.h
index 7314dcd..ff312fa 100644
--- a/src/drm/drm.h
+++ b/src/drm/drm.h
@@ -3,7 +3,7 @@
 #define COLORBUFFER_RELOC_INDEX 0
 #define ZBUFFER_RELOC_INDEX 1
 #define FLUSH_RELOC_INDEX 2
-#define TEXTURE_RELOC_INDEX 3
+#define TEXTUREBUFFER_RELOC_INDEX 3
 
 #ifdef __cplusplus
 extern "C" {
diff --git a/src/matrix_cubesphere.cpp b/src/matrix_cubesphere.cpp
index cd60745..f9913e5 100644
--- a/src/matrix_cubesphere.cpp
+++ b/src/matrix_cubesphere.cpp
@@ -17,6 +17,7 @@
 #include "r500/indirect_buffer.h"
 #include "r500/shader.h"
 #include "r500/display_controller.h"
+#include "r500/texture.h"
 
 #include "drm/buffer.h"
 #include "drm/drm.h"
@@ -41,18 +42,11 @@ const int vertex_shader_paths_length = (sizeof (vertex_shader_paths)) / (sizeof
 const char * fragment_shader_paths[] = {
   "clear.fs.bin",
   "matrix_cubesphere.fs.bin",
-  "light.fs.bin"
+  "light.fs.bin",
 };
 const int fragment_shader_paths_length = (sizeof (fragment_shader_paths)) / (sizeof (fragment_shader_paths[0]));
 
-struct shaders {
-  struct shader_offset * vertex;
-  struct shader_offset * fragment;
-  int vertex_length;
-  int fragment_length;
-};
-
-void _3d_clear(struct shaders& shaders)
+void _3d_clear(const shaders& shaders)
 {
   ib_rs_instructions(0);
 
@@ -67,7 +61,7 @@ void _3d_clear(struct shaders& shaders)
 
   //
 
-  ib_zbuffer(ZBUFFER_RELOC_INDEX, 7); // always
+  ib_zbuffer(ZBUFFER_RELOC_INDEX, 1600, 7); // always
 
   ib_texture__0();
 
@@ -298,7 +292,7 @@ void _3d_light_inner(mat4x4 trans)
   }
 }
 
-vec3 _3d_light(struct shaders& shaders,
+vec3 _3d_light(const shaders& shaders,
                const mat4x4& view_to_clip,
                float theta)
 {
@@ -315,7 +309,7 @@ vec3 _3d_light(struct shaders& shaders,
 
   //
 
-  ib_zbuffer(ZBUFFER_RELOC_INDEX, 1); // less
+  ib_zbuffer(ZBUFFER_RELOC_INDEX, 1600, 1); // less
 
   ib_texture__0();
 
@@ -382,7 +376,7 @@ vec3 _3d_light(struct shaders& shaders,
   return light_pos;
 }
 
-void _3d_cube(struct shaders& shaders,
+void _3d_cube(const shaders& shaders,
               const mat4x4& view_to_clip,
               float theta,
               const vec3& light_pos)
@@ -403,9 +397,17 @@ void _3d_cube(struct shaders& shaders,
 
   //
 
-  ib_zbuffer(ZBUFFER_RELOC_INDEX, 1); // less
+  ib_zbuffer(ZBUFFER_RELOC_INDEX, 1600, 1); // less
 
-  ib_texture__1(TEXTURE_RELOC_INDEX);
+  int width = 1024;
+  int height = 1024;
+  int macrotile = 0;
+  int microtile = 0;
+  int clamp = 0; // wrap/repeat
+  ib_texture__1(TEXTUREBUFFER_RELOC_INDEX,
+                width, height,
+                macrotile, microtile,
+                clamp);
 
   ib_vap_stream_cntl__323();
 
@@ -470,19 +472,28 @@ void _3d_cube(struct shaders& shaders,
   _3d_cube_inner(trans, world_trans, light_pos);
 }
 
-int indirect_buffer(shaders& shaders,
+int indirect_buffer(const shaders& shaders,
                     float theta)
 {
+  int width = 1600;
+  int height = 1200;
+  int pitch = width;
+
   ib_ix = 0;
 
   ib_generic_initialization();
 
+  ib_viewport(width, height);
+  ib_colorbuffer(COLORBUFFER_RELOC_INDEX, pitch, 0, 0);
+
+  T0V(GB_ENABLE, 0);
+
   T0V(US_OUT_FMT_0
       , US_OUT_FMT__OUT_FMT(0)  // C4_8
-      | US_OUT_FMT__C0_SEL(3)   // Blue
-      | US_OUT_FMT__C1_SEL(2)   // Green
-      | US_OUT_FMT__C2_SEL(1)   // Red
-      | US_OUT_FMT__C3_SEL(0)   // Alpha
+      | US_OUT_FMT__C0_SEL__BLUE
+      | US_OUT_FMT__C1_SEL__GREEN
+      | US_OUT_FMT__C2_SEL__RED
+      | US_OUT_FMT__C3_SEL__ALPHA
       | US_OUT_FMT__OUT_SIGN(0)
       );
   T0V(US_OUT_FMT_1
@@ -491,12 +502,10 @@ int indirect_buffer(shaders& shaders,
   T0V(US_OUT_FMT_2
       , US_OUT_FMT__OUT_FMT(15) // render target is not used
       );
-  T0V(US_OUT_FMT_2
+  T0V(US_OUT_FMT_3
       , US_OUT_FMT__OUT_FMT(15) // render target is not used
       );
 
-  ib_colorbuffer(COLORBUFFER_RELOC_INDEX);
-
   load_pvs_shaders(shaders.vertex, shaders.vertex_length);
   load_us_shaders(shaders.fragment, shaders.fragment_length);
 
@@ -559,7 +568,14 @@ int main()
   colorbuffer_handle[1] = create_buffer(fd, colorbuffer_size, &colorbuffer_ptr[1]);
   zbuffer_handle = create_buffer(fd, colorbuffer_size, &zbuffer_ptr);
   flush_handle = create_flush_buffer(fd);
-  texturebuffer_handle = load_textures(fd, textures, textures_length);
+  //texturebuffer_handle = load_textures(fd, textures, textures_length);
+  texturebuffer_handle = load_textures_tiled(fd,
+                                             zbuffer_handle,
+                                             flush_handle,
+                                             textures,
+                                             textures_length,
+                                             1024,
+                                             1024);
 
   fprintf(stderr, "colorbuffer handle[0] %d\n", colorbuffer_handle[0]);
   fprintf(stderr, "colorbuffer handle[1] %d\n", colorbuffer_handle[1]);
@@ -584,6 +600,8 @@ int main()
     // next state
     theta += 0.01f;
     colorbuffer_ix = (colorbuffer_ix + 1) & 1;
+
+    break;
   }
 
   close(fd);
diff --git a/src/r500/indirect_buffer.c b/src/r500/indirect_buffer.c
index 9cdd37a..cce0916 100644
--- a/src/r500/indirect_buffer.c
+++ b/src/r500/indirect_buffer.c
@@ -136,11 +136,6 @@ void ib_generic_initialization()
       | GA_ROUND_MODE__GEOMETRY_MASK(0)
       );
 
-  T0Vf(GA_POINT_S0, 0.0f);
-  T0Vf(GA_POINT_T0, 1.0f);
-  T0Vf(GA_POINT_S1, 1.0f);
-  T0Vf(GA_POINT_T1, 0.0f);
-
   T0V(GA_COLOR_CONTROL
       , GA_COLOR_CONTROL__RGB0_SHADING(2)
       | GA_COLOR_CONTROL__ALPHA0_SHADING(2)
@@ -160,7 +155,6 @@ void ib_generic_initialization()
       );
 
   T0V(GB_SELECT, 0x00000000);
-  T0V(GB_ENABLE, 0x00000000);
 
   T0V(GB_MSPOS0
       , GB_MSPOS0__MS_X0(6)
@@ -278,33 +272,9 @@ void ib_generic_initialization()
   T0V(US_W_FMT
       , US_W_FMT__W_FMT(0) // W is always zero
       );
-
-  //////////////////////////////////////////////////////////////////////////////
-  // SC
-  //////////////////////////////////////////////////////////////////////////////
-
-  T0V(SC_SCISSOR0
-      , SC_SCISSOR0__XS0(0)
-      | SC_SCISSOR0__YS0(0)
-      );
-  T0V(SC_SCISSOR1
-      , SC_SCISSOR1__XS1(1600 - 1)
-      | SC_SCISSOR1__YS1(1200 - 1)
-      );
-
-  //////////////////////////////////////////////////////////////////////////////
-  // VAP
-  //////////////////////////////////////////////////////////////////////////////
-
-  T0Vf(VAP_VPORT_XSCALE,   800.0f);
-  T0Vf(VAP_VPORT_XOFFSET,  800.0f);
-  T0Vf(VAP_VPORT_YSCALE,  -600.0f);
-  T0Vf(VAP_VPORT_YOFFSET,  600.0f);
-  T0Vf(VAP_VPORT_ZSCALE,     0.5f);
-  T0Vf(VAP_VPORT_ZOFFSET,    0.5f);
 }
 
-void ib_colorbuffer(int reloc_index)
+void ib_colorbuffer(int reloc_index, int pitch, int macrotile, int microtile)
 {
 
   //////////////////////////////////////////////////////////////////////////////
@@ -318,8 +288,10 @@ void ib_colorbuffer(int reloc_index)
   TU(reloc_index * 4); // index into relocs array
 
   T0V(RB3D_COLORPITCH0
-      , RB3D_COLORPITCH__COLORPITCH(1600 >> 1)
-      | RB3D_COLORPITCH__COLORFORMAT(6) // ARGB8888
+      , RB3D_COLORPITCH__COLORPITCH(pitch >> 1)
+      | RB3D_COLORPITCH__COLORTILE(macrotile)
+      | RB3D_COLORPITCH__COLORMICROTILE(microtile)
+      | RB3D_COLORPITCH__COLORFORMAT__ARGB8888
       );
   // The COLORPITCH NOP is ignored/not applied due to
   // RADEON_CS_KEEP_TILING_FLAGS, but is still required.
@@ -327,7 +299,37 @@ void ib_colorbuffer(int reloc_index)
   TU(reloc_index * 4); // index into relocs array
 }
 
-void ib_zbuffer(int reloc_index, int zfunc)
+void ib_viewport(int width, int height)
+{
+  //////////////////////////////////////////////////////////////////////////////
+  // SC
+  //////////////////////////////////////////////////////////////////////////////
+
+  T0V(SC_SCISSOR0
+      , SC_SCISSOR0__XS0(0)
+      | SC_SCISSOR0__YS0(0)
+      );
+  T0V(SC_SCISSOR1
+      , SC_SCISSOR1__XS1(width - 1)
+      | SC_SCISSOR1__YS1(height - 1)
+      );
+
+  //////////////////////////////////////////////////////////////////////////////
+  // VAP
+  //////////////////////////////////////////////////////////////////////////////
+
+  float x = ((float)width) * 0.5f;
+  float y = ((float)height) * 0.5f;
+
+  T0Vf(VAP_VPORT_XSCALE,   x);
+  T0Vf(VAP_VPORT_XOFFSET,  x);
+  T0Vf(VAP_VPORT_YSCALE,  -y);
+  T0Vf(VAP_VPORT_YOFFSET,  y);
+  T0Vf(VAP_VPORT_ZSCALE,   0.5f);
+  T0Vf(VAP_VPORT_ZOFFSET,  0.5f);
+}
+
+void ib_zbuffer(int reloc_index, int pitch, int zfunc)
 {
   //////////////////////////////////////////////////////////////////////////////
   // ZB
@@ -349,7 +351,7 @@ void ib_zbuffer(int reloc_index, int zfunc)
   TU(reloc_index * 4); // index into relocs array
 
   T0V(ZB_DEPTHPITCH
-      , ZB_DEPTHPITCH__DEPTHPITCH(1600 >> 2)
+      , ZB_DEPTHPITCH__DEPTHPITCH(pitch >> 2)
       | ZB_DEPTHPITCH__DEPTHMACROTILE(1)
       | ZB_DEPTHPITCH__DEPTHMICROTILE(1)
       );
@@ -518,7 +520,10 @@ void ib_texture__0()
   T0V(TX_ENABLE, 0x00000000);
 }
 
-void ib_texture__1(int reloc_index)
+void ib_texture__1(int reloc_index,
+                   int width, int height,
+                   int macrotile, int microtile,
+                   int clamp)
 {
   //////////////////////////////////////////////////////////////////////////////
   // TX
@@ -528,17 +533,21 @@ void ib_texture__1(int reloc_index)
 
   T0V(TX_ENABLE
       , TX_ENABLE__TEX_0_ENABLE__ENABLE);
+
   T0V(TX_FILTER0_0
-      , TX_FILTER0__MAG_FILTER__LINEAR
+      , TX_FILTER0__CLAMP_S(clamp)
+      | TX_FILTER0__CLAMP_T(clamp)
+      | TX_FILTER0__MAG_FILTER__LINEAR
       | TX_FILTER0__MIN_FILTER__LINEAR
       );
   T0V(TX_FILTER1_0
       , TX_FILTER1__LOD_BIAS(1)
+      | TX_FILTER1__BORDER_FIX(1)
       );
   T0V(TX_BORDER_COLOR_0, 0);
   T0V(TX_FORMAT0_0
-      , TX_FORMAT0__TXWIDTH(1024 - 1)
-      | TX_FORMAT0__TXHEIGHT(1024 - 1)
+      , TX_FORMAT0__TXWIDTH(width - 1)
+      | TX_FORMAT0__TXHEIGHT(height - 1)
       );
 
   T0V(TX_FORMAT1_0
@@ -552,9 +561,8 @@ void ib_texture__1(int reloc_index)
   T0V(TX_FORMAT2_0, 0);
 
   T0V(TX_OFFSET_0
-      //, TX_OFFSET__MACRO_TILE(1)
-      //| TX_OFFSET__MICRO_TILE(1)
-      , 0
+      , TX_OFFSET__MACRO_TILE(macrotile)
+      | TX_OFFSET__MICRO_TILE(microtile)
       );
 
   T3(_NOP, 0);
diff --git a/src/r500/indirect_buffer.h b/src/r500/indirect_buffer.h
index 87db0c7..980015b 100644
--- a/src/r500/indirect_buffer.h
+++ b/src/r500/indirect_buffer.h
@@ -55,11 +55,15 @@ extern union u32_f32 ib[16384];
 extern volatile int ib_ix;
 
 void ib_generic_initialization();
-void ib_colorbuffer(int reloc_index);
-void ib_zbuffer(int reloc_index, int zfunc);
+void ib_viewport(int width, int height);
+void ib_colorbuffer(int reloc_index, int pitch, int macrotile, int microtile);
+void ib_zbuffer(int reloc_index, int pitch, int zfunc);
 void ib_rs_instructions(int count);
 void ib_texture__0();
-void ib_texture__1(int reloc_index);
+void ib_texture__1(int reloc_index,
+                   int width, int height,
+                   int macrotile, int microtile,
+                   int clamp);
 void ib_vap_pvs(struct shader_offset * offset);
 void ib_ga_us(struct shader_offset * offset);
 void ib_vap_pvs_const_cntl(const float * consts, int size);
diff --git a/src/r500/shader.h b/src/r500/shader.h
index 8d7b960..bd64624 100644
--- a/src/r500/shader.h
+++ b/src/r500/shader.h
@@ -16,6 +16,13 @@ void load_us_shaders(struct shader_offset * offsets, int offsets_length);
 
 struct shader_offset * load_shaders(const char ** paths, int paths_length);
 
+struct shaders {
+  struct shader_offset * vertex;
+  struct shader_offset * fragment;
+  int vertex_length;
+  int fragment_length;
+};
+
 #ifdef __cplusplus
 }
 #endif
diff --git a/src/r500/texture.c b/src/r500/texture.c
new file mode 100644
index 0000000..c6bdfe3
--- /dev/null
+++ b/src/r500/texture.c
@@ -0,0 +1,310 @@
+#include <assert.h>
+#include <stdlib.h>
+#include <stdio.h>
+#include <sys/mman.h>
+#include <unistd.h>
+
+#include "3d_registers.h"
+#include "3d_registers_bits.h"
+
+#include "../drm/buffer.h"
+#include "../drm/drm.h"
+#include "../file.h"
+
+#include "texture.h"
+#include "shader.h"
+#include "indirect_buffer.h"
+
+#define TEXTURE_TILE_SHADER 0
+
+static int _tile_texture(const struct shaders * shaders,
+                         int input_reloc_index,
+                         int output_reloc_index)
+{
+  int width = 1024;
+  int height = 1024;
+  int pitch = width;
+  float x = (float)width * 0.5f;
+  float y = (float)height * 0.5f;
+
+  ib_ix = 0;
+
+  ib_generic_initialization();
+
+  ib_viewport(width, height);
+  ib_colorbuffer(output_reloc_index, pitch, 1, 1); // macrotile, microtile
+
+  T0V(US_OUT_FMT_0
+      , US_OUT_FMT__OUT_FMT(0)  // C4_8
+      | US_OUT_FMT__C0_SEL__RED
+      | US_OUT_FMT__C1_SEL__GREEN
+      | US_OUT_FMT__C2_SEL__BLUE
+      | US_OUT_FMT__C3_SEL__ALPHA
+      | US_OUT_FMT__OUT_SIGN(0)
+      );
+  T0V(US_OUT_FMT_1
+      , US_OUT_FMT__OUT_FMT(15) // render target is not used
+      );
+  T0V(US_OUT_FMT_2
+      , US_OUT_FMT__OUT_FMT(15) // render target is not used
+      );
+  T0V(US_OUT_FMT_3
+      , US_OUT_FMT__OUT_FMT(15) // render target is not used
+      );
+
+  // GA
+
+  T0V(GB_ENABLE
+    , GB_ENABLE__POINT_STUFF_ENABLE(1)
+    | GB_ENABLE__TEX0_SOURCE(2) // stuff with source texture coordinates s,t
+    );
+
+  T0Vf(GA_POINT_S0, 0.0f);
+  T0Vf(GA_POINT_T0, 1.0f);
+  T0Vf(GA_POINT_S1, 1.0f);
+  T0Vf(GA_POINT_T1, 0.0f);
+
+  //////////////////////////////////////////////////////////////////////////////
+  // RS
+  //////////////////////////////////////////////////////////////////////////////
+
+  int rs_instructions = 1;
+
+  //ib_rs_instructions(0);
+
+  T0V(RS_IP_0
+    , RS_IP__TEX_PTR_S(0)
+    | RS_IP__TEX_PTR_T(1)
+    | RS_IP__TEX_PTR_R(62) // constant 0.0
+    | RS_IP__TEX_PTR_Q(63) // constant 1.0
+    );
+
+  T0V(RS_COUNT
+      , RS_COUNT__IT_COUNT(2)
+      | RS_COUNT__IC_COUNT(0)
+      | RS_COUNT__W_ADDR(0)
+      | RS_COUNT__HIRES_EN(1)
+      );
+
+  T0V(RS_INST_COUNT
+      , RS_INST_COUNT__INST_COUNT(rs_instructions - 1));
+
+  T0V(RS_INST_0
+      , RS_INST__TEX_ID(0)
+      | RS_INST__TEX_CN(1)
+      | RS_INST__TEX_ADDR(0)
+      );
+
+  //////////////////////////////////////////////////////////////////////////////
+  // VAP OUT
+  //////////////////////////////////////////////////////////////////////////////
+
+  T0V(VAP_OUT_VTX_FMT_0
+      , VAP_OUT_VTX_FMT_0__VTX_POS_PRESENT(1)
+      );
+  T0V(VAP_OUT_VTX_FMT_1
+      , 0
+      );
+
+  //
+
+  T0V(ZB_CNTL, 0);
+  T0V(ZB_ZSTENCILCNTL, 0);
+
+  //
+
+  int macrotile = 0;
+  int microtile = 0;
+  int clamp = 2; // clamp to [0.0, 1.0]
+  ib_texture__1(input_reloc_index,
+                width, height,
+                macrotile, microtile,
+                clamp);
+
+  ib_vap_stream_cntl__2();
+
+  T0V(US_PIXSIZE
+      , US_PIXSIZE__PIX_SIZE(1)
+      );
+
+  ib_ga_us(&shaders->fragment[TEXTURE_TILE_SHADER]);
+  ib_vap_pvs(&shaders->vertex[TEXTURE_TILE_SHADER]);
+
+  //////////////////////////////////////////////////////////////////////////////
+  // VAP
+  //////////////////////////////////////////////////////////////////////////////
+
+  T0V(VAP_CLIP_CNTL
+      , VAP_CLIP_CNTL__CLIP_DISABLE(1)
+      );
+
+  T0V(VAP_VTE_CNTL
+      , VAP_VTE_CNTL__VTX_XY_FMT(1) // disable W division
+      | VAP_VTE_CNTL__VTX_Z_FMT(1)  // disable W division
+      );
+
+  T0V(VAP_CNTL_STATUS
+      , VAP_CNTL_STATUS__PVS_BYPASS(0)
+      );
+
+  //////////////////////////////////////////////////////////////////////////////
+  // GA POINT SIZE
+  //////////////////////////////////////////////////////////////////////////////
+
+  T0V(GA_POINT_SIZE
+      , GA_POINT_SIZE__HEIGHT((int)(x * 12.0f))
+      | GA_POINT_SIZE__WIDTH((int)(y * 12.0f))
+      );
+
+  //////////////////////////////////////////////////////////////////////////////
+  // 3D_DRAW
+  //////////////////////////////////////////////////////////////////////////////
+
+  const int dwords_per_vtx = 2;
+
+  T0V(VAP_VTX_SIZE
+      , VAP_VTX_SIZE__DWORDS_PER_VTX(dwords_per_vtx)
+      );
+
+  const float center[] = {
+    x, y,
+  };
+  const int vertex_count = 1;
+  T3(_3D_DRAW_IMMD_2, (1 + vertex_count * dwords_per_vtx) - 1);
+  TU( VAP_VF_CNTL__PRIM_TYPE(1) // point list
+    | VAP_VF_CNTL__PRIM_WALK(3)
+    | VAP_VF_CNTL__INDEX_SIZE(0)
+    | VAP_VF_CNTL__VTX_REUSE_DIS(0)
+    | VAP_VF_CNTL__DUAL_INDEX_MODE(0)
+    | VAP_VF_CNTL__USE_ALT_NUM_VERTS(0)
+    | VAP_VF_CNTL__NUM_VERTICES(vertex_count)
+    );
+  for (int i = 0; i < 2; i++) {
+    TF(center[i]);
+  }
+
+  //////////////////////////////////////////////////////////////////////////////
+  // padding
+  //////////////////////////////////////////////////////////////////////////////
+
+  while ((ib_ix % 8) != 0) {
+    TU(0x80000000);
+  }
+
+  return ib_ix;
+}
+
+int * load_textures(int fd,
+                    const char ** textures,
+                    int textures_length)
+{
+  int * texturebuffer_handle = (int *)malloc((sizeof (int)) * textures_length);
+
+  for (int i = 0; i < textures_length; i++) {
+    int size = 0;
+    void * buf = file_read(textures[i], &size);
+    assert(buf != NULL);
+
+    printf("load texture[%d]: %d\n", i, size);
+
+    void * ptr = NULL;
+    int handle = create_buffer(fd, size, &ptr);
+
+    for (int i = 0; i < size / 4; i++) {
+      ((uint32_t*)ptr)[i] = ((uint32_t*)buf)[i];
+    }
+    asm volatile ("" ::: "memory");
+    free(buf);
+    munmap(ptr, size);
+
+    texturebuffer_handle[i] = handle;
+  }
+
+  return texturebuffer_handle;
+}
+
+static void tile_texture(int fd,
+                         const struct shaders * shaders,
+                         int colorbuffer_handle,
+                         int zbuffer_handle,
+                         int flush_handle,
+                         int texturebuffer_handle)
+{
+  int ib_dwords = _tile_texture(shaders,
+                                TEXTUREBUFFER_RELOC_INDEX, // input
+                                COLORBUFFER_RELOC_INDEX);  // output
+
+  drm_radeon_cs(fd,
+                colorbuffer_handle,
+                zbuffer_handle,
+                flush_handle,
+                &texturebuffer_handle,
+                1,
+                ib_dwords);
+}
+
+static const char * vertex_shader_paths[] = {
+  "texture_tile.vs.bin",
+};
+static const int vertex_shader_paths_length = (sizeof (vertex_shader_paths)) / (sizeof (vertex_shader_paths[0]));
+static const char * fragment_shader_paths[] = {
+  "texture_tile.fs.bin",
+};
+static const int fragment_shader_paths_length = (sizeof (fragment_shader_paths)) / (sizeof (fragment_shader_paths[0]));
+
+int * load_textures_tiled(int fd,
+                          int zbuffer_handle,
+                          int flush_handle,
+                          const char ** textures,
+                          int textures_length,
+                          int max_width,
+                          int max_height)
+{
+  struct shaders shaders = {
+    .vertex = load_shaders(vertex_shader_paths, vertex_shader_paths_length),
+    .fragment = load_shaders(fragment_shader_paths, fragment_shader_paths_length),
+    .vertex_length = vertex_shader_paths_length,
+    .fragment_length = fragment_shader_paths_length,
+  };
+
+  // shaders
+  load_pvs_shaders(shaders.vertex, shaders.vertex_length);
+  load_us_shaders(shaders.fragment, shaders.fragment_length);
+
+  int * texturebuffer_handle = (int *)malloc((sizeof (int)) * textures_length);
+
+  int max_size = max_width * max_height * 4;
+  void * temp_ptr;
+  int temp_handle = create_buffer(fd, max_size, &temp_ptr);
+
+  for (int i = 0; i < textures_length; i++) {
+    int size = 0;
+    void * buf = file_read(textures[i], &size);
+    assert(buf != NULL);
+
+    printf("load tiled texture[%d]: %d\n", i, size);
+
+    for (int i = 0; i < size / 4; i++) {
+      ((uint32_t*)temp_ptr)[i] = ((uint32_t*)buf)[i];
+    }
+    asm volatile ("" ::: "memory");
+    free(buf);
+
+    assert(size <= max_size);
+    int handle = create_buffer(fd, size, NULL);
+
+    tile_texture(fd,
+                 &shaders,
+                 handle,
+                 zbuffer_handle,
+                 flush_handle,
+                 temp_handle);
+
+    texturebuffer_handle[i] = handle;
+  }
+  sleep(1);
+
+  munmap(temp_ptr, max_size);
+
+  return texturebuffer_handle;
+}
diff --git a/src/r500/texture.h b/src/r500/texture.h
new file mode 100644
index 0000000..6013beb
--- /dev/null
+++ b/src/r500/texture.h
@@ -0,0 +1,21 @@
+#pragma once
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+int * load_textures(int fd,
+                    const char ** textures,
+                    int textures_length);
+
+int * load_textures_tiled(int fd,
+                          int zbuffer_handle,
+                          int flush_handle,
+                          const char ** textures,
+                          int textures_length,
+                          int max_width,
+                          int max_height);
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/src/texture_tile.fs.asm b/src/texture_tile.fs.asm
new file mode 100644
index 0000000..3e534e8
--- /dev/null
+++ b/src/texture_tile.fs.asm
@@ -0,0 +1,8 @@
+TEX TEX_SEM_WAIT TEX_SEM_ACQUIRE
+  temp[0].rgba = LD tex[0].rgba temp[0].rgaa ;
+
+OUT TEX_SEM_WAIT
+src0.a = temp[0],
+src0.rgb = temp[0] :
+  out[0].a    = MAX src0.a src0.a ,
+  out[0].rgb  = MAX src0.rgb src0.rgb ;
diff --git a/src/texture_tile.fs.bin b/src/texture_tile.fs.bin
new file mode 100644
index 0000000..bbfaa90
Binary files /dev/null and b/src/texture_tile.fs.bin differ
diff --git a/src/texture_tile.vs.asm b/src/texture_tile.vs.asm
new file mode 100644
index 0000000..e779e8c
--- /dev/null
+++ b/src/texture_tile.vs.asm
@@ -0,0 +1 @@
+out[0].xyzw = VE_ADD  input[0].xyzw  input[0].0000 ;
diff --git a/src/texture_tile.vs.bin b/src/texture_tile.vs.bin
new file mode 100644
index 0000000..69e1685
Binary files /dev/null and b/src/texture_tile.vs.bin differ