Upate to libGL 7.11.2

Tested by jsg@, matthieu@ and ajacoutot@, ok mattieu@

18 files changed, 3809 insertions, 514 deletions

diff --git a/lib/libGL/Makefile b/lib/libGL/Makefile
index 9387986c3..b62f95e19 100644
--- a/lib/libGL/Makefile
+++ b/lib/libGL/Makefile
@@ -1,4 +1,4 @@
-# $OpenBSD: Makefile,v 1.30 2012/06/04 10:18:45 mpi Exp $
+# $OpenBSD: Makefile,v 1.31 2012/08/17 13:57:59 mpi Exp $
 .include <bsd.xconf.mk>
 
 .if ${XENOCARA_BUILD_DRI:L} == "yes"
@@ -15,8 +15,8 @@ MESA_INCLUDE=	${.CURDIR}/../../dist/Mesa/include
 MESA_DRI=	${.CURDIR}/../../dist/Mesa/src/mesa/drivers/dri
 
 MESA_MAJOR=	7
-MESA_MINOR=	10
-MESA_TINY=	3
+MESA_MINOR=	11
+MESA_TINY=	2
 
 DRI_DRIVER_INSTALL_DIR=	${X11BASE}/lib/modules/dri
 
@@ -48,6 +48,7 @@ CPPFLAGS+= \
 	-DPTHREADS \
 	-DHAVE_ALIAS \
 	-D_REENTRANT \
+	-DMAPI_MODE_UTIL \
 	-DMAPI_GLAPI_CURRENT
 
 .if ${XENOCARA_BUILD_DRI:L} == "yes"
@@ -76,9 +77,11 @@ INCS= \
 GLAPI_SOURCES= \
 	glapi_dispatch.c \
 	glapi_entrypoint.c \
+	glapi_gentable.c \
 	glapi_getproc.c \
 	glapi_nop.c \
 	glthread.c \
+	glapi.c \
 	$(ASM_API)
 
 .if ${MACHINE_ARCH} == i386
diff --git a/lib/libGL/dri/Makefile.inc b/lib/libGL/dri/Makefile.inc
index e705ff4b5..8caecd59d 100644
--- a/lib/libGL/dri/Makefile.inc
+++ b/lib/libGL/dri/Makefile.inc
@@ -1,4 +1,4 @@
-# $OpenBSD: Makefile.inc,v 1.13 2012/07/13 12:55:32 mpi Exp $
+# $OpenBSD: Makefile.inc,v 1.14 2012/08/17 13:58:00 mpi Exp $
 .include <bsd.xconf.mk>
 
 MESA= ${.CURDIR}/../../../../dist/Mesa/src/mesa
@@ -15,8 +15,7 @@ COMMON_GALLIUM_SOURCES = \
 COMMON_SOURCES = $(COMMON_GALLIUM_SOURCES) \
 	driverfuncs.c \
 	texmem.c \
-	drirenderbuffer.c \
-	dri_metaops.c
+	drirenderbuffer.c
 
 CFLAGS+= -g -DIN_DRI_DRIVER -DGLX_DIRECT_RENDERING -DPTHREADS \
 	 $(INCLUDES) $(DRIVER_DEFINES)
@@ -29,6 +28,7 @@ INCLUDES = \
 	-I$(TOP)/src/mesa/drivers/dri/common \
 	-I$(TOP)/include \
 	-I$(TOP)/include/GL/internal \
+	-I$(TOP)/src/glsl \
 	-I$(TOP)/src/mesa \
 	-I$(TOP)/src/mesa/main \
 	-I$(TOP)/src/mapi \
diff --git a/lib/libGL/dri/i915/Makefile b/lib/libGL/dri/i915/Makefile
index f308e2afb..3444de4ce 100644
--- a/lib/libGL/dri/i915/Makefile
+++ b/lib/libGL/dri/i915/Makefile
@@ -1,4 +1,4 @@
-#	$OpenBSD: Makefile,v 1.7 2011/10/23 13:57:36 matthieu Exp $
+#	$OpenBSD: Makefile,v 1.8 2012/08/17 13:58:00 mpi Exp $
 
 LIB=	i915_dri.so
 
@@ -14,7 +14,7 @@ DRIVER_SOURCES = \
 	intel_batchbuffer.c \
 	intel_clear.c \
 	intel_extensions.c \
-	intel_extensions_es2.c \
+	intel_extensions_es.c \
 	intel_mipmap_tree.c \
 	intel_tex_layout.c \
 	intel_tex_image.c \
diff --git a/lib/libGL/dri/i965/Makefile b/lib/libGL/dri/i965/Makefile
index 0c90446d5..061565e73 100644
--- a/lib/libGL/dri/i965/Makefile
+++ b/lib/libGL/dri/i965/Makefile
@@ -1,4 +1,4 @@
-#	$OpenBSD: Makefile,v 1.6 2011/10/23 13:57:36 matthieu Exp $
+#	$OpenBSD: Makefile,v 1.7 2012/08/17 13:58:00 mpi Exp $
 
 LIB=	i965_dri.so
 
@@ -11,7 +11,7 @@ DRIVER_SOURCES = \
 	intel_context.c \
 	intel_decode.c \
 	intel_extensions.c \
-	intel_extensions_es2.c \
+	intel_extensions_es.c \
 	intel_fbo.c \
 	intel_mipmap_tree.c \
 	intel_regions.c \
@@ -59,7 +59,6 @@ DRIVER_SOURCES = \
 	brw_sf.c \
 	brw_sf_emit.c \
 	brw_sf_state.c \
-	brw_state.c \
 	brw_state_batch.c \
 	brw_state_cache.c \
 	brw_state_dump.c \
@@ -95,14 +94,31 @@ DRIVER_SOURCES = \
 	gen6_urb.c \
 	gen6_viewport_state.c \
 	gen6_vs_state.c \
-	gen6_wm_state.c
+	gen6_wm_state.c \
+	gen7_cc_state.c \
+	gen7_clip_state.c \
+	gen7_disable.c \
+	gen7_misc_state.c \
+	gen7_sampler_state.c \
+	gen7_sf_state.c \
+	gen7_urb.c \
+	gen7_viewport_state.c \
+	gen7_vs_state.c \
+	gen7_wm_state.c \
+	gen7_wm_surface_state.c
 
 CXX_SOURCES = \
 	brw_cubemap_normalize.cpp \
 	brw_fs.cpp \
+	brw_fs_emit.cpp \
+	brw_fs_visitor.cpp \
 	brw_fs_channel_expressions.cpp \
 	brw_fs_reg_allocate.cpp \
-	brw_fs_vector_splitting.cpp
+	brw_fs_schedule_instructions.cpp \
+	brw_fs_vector_splitting.cpp \
+	brw_shader.cpp
+ 
+
 
 SRCS = \
 	$(COMMON_SOURCES) \
diff --git a/lib/libGL/dri/r300/Makefile b/lib/libGL/dri/r300/Makefile
index e47d4f522..ff884d9f2 100644
--- a/lib/libGL/dri/r300/Makefile
+++ b/lib/libGL/dri/r300/Makefile
@@ -1,4 +1,4 @@
-#	$OpenBSD: Makefile,v 1.5 2011/10/23 13:57:36 matthieu Exp $
+#	$OpenBSD: Makefile,v 1.6 2012/08/17 13:58:00 mpi Exp $
 
 LIB=	r300_dri.so
 
@@ -9,7 +9,7 @@ CS_SOURCES = \
 
 COMMON_SOURCES = \
 	driverfuncs.c \
-	mm.c \
+	drirenderbuffer.c \
 	utils.c \
 	texmem.c \
 	vblank.c \
@@ -70,12 +70,15 @@ COMPILER_SOURCES = \
 		radeon_pair_translate.c \
 		radeon_pair_schedule.c \
 		radeon_pair_regalloc.c \
+		radeon_pair_dead_sources.c \
 		radeon_dataflow.c \
 		radeon_dataflow_deadcode.c \
 		radeon_dataflow_swizzles.c \
+		radeon_list.c \
 		radeon_optimize.c \
 		radeon_remove_constants.c \
 		radeon_rename_regs.c \
+		radeon_variable.c \
 		r3xx_fragprog.c \
 		r300_fragprog.c \
 		r300_fragprog_swizzle.c \
diff --git a/lib/libGL/dri/r600/Makefile b/lib/libGL/dri/r600/Makefile
index eebc8a32a..28c835826 100644
--- a/lib/libGL/dri/r600/Makefile
+++ b/lib/libGL/dri/r600/Makefile
@@ -1,10 +1,10 @@
-#	$OpenBSD: Makefile,v 1.3 2011/10/23 13:57:36 matthieu Exp $
+#	$OpenBSD: Makefile,v 1.4 2012/08/17 13:58:00 mpi Exp $
 
 LIB=	r600_dri.so
 
 COMMON_SOURCES = \
 	driverfuncs.c \
-	mm.c \
+	drirenderbuffer.c \
 	utils.c \
 	texmem.c \
 	vblank.c \
diff --git a/lib/libGL/dri/swrastg/Makefile b/lib/libGL/dri/swrastg/Makefile
index 8bac1d30b..cb3a9f241 100644
--- a/lib/libGL/dri/swrastg/Makefile
+++ b/lib/libGL/dri/swrastg/Makefile
@@ -1,4 +1,4 @@
-#	$OpenBSD: Makefile,v 1.1 2012/07/13 12:54:20 mpi Exp $
+#	$OpenBSD: Makefile,v 1.2 2012/08/17 13:58:00 mpi Exp $
 
 LIB=	swrast_dri.so
 
@@ -45,7 +45,7 @@ LDADD=	../../gallium/libmesagallium/libmesagallium_pic.a \
 # XXX This is wrong but the gallium drivers depend on libpthread
 # and some bindings still fail even with the LD_PRELOAD hack when
 # starting X.
-#LDADD+= -lpthread
+LDADD+= -lpthread
 DPADD=	../../gallium/libmesagallium/libmesagallium_pic.a
 
 .include "../../gallium/Makefile.inc"
diff --git a/lib/libGL/gallium/libgallium/Makefile b/lib/libGL/gallium/libgallium/Makefile
index ca1023a3b..b5846da9f 100644
--- a/lib/libGL/gallium/libgallium/Makefile
+++ b/lib/libGL/gallium/libgallium/Makefile
@@ -1,4 +1,4 @@
-# $OpenBSD: Makefile,v 1.1 2012/07/13 12:43:30 mpi Exp $
+# $OpenBSD: Makefile,v 1.2 2012/08/17 13:58:00 mpi Exp $
 
 LIB=	gallium
 
@@ -50,7 +50,7 @@ C_SOURCES = \
 	draw_vs_exec.c \
 	draw_vs_ppc.c \
 	draw_vs_sse.c \
-	draw_vs_varient.c \
+	draw_vs_variant.c \
 	u_indices_gen.c \
 	u_unfilled_gen.c \
 	os_misc.c \
@@ -113,10 +113,13 @@ C_SOURCES = \
 	u_caps.c \
 	u_cpu_detect.c \
 	u_dl.c \
+	u_draw.c \
 	u_draw_quad.c \
 	u_format.c \
 	u_format_other.c \
+	u_format_latc.c \
 	u_format_s3tc.c \
+	u_format_rgtc.c \
 	u_format_srgb.c \
 	u_format_table.c \
 	u_format_tests.c \
@@ -135,6 +138,7 @@ C_SOURCES = \
 	u_network.c \
 	u_math.c \
 	u_mm.c \
+	u_pstipple.c \
 	u_rect.c \
 	u_ringbuffer.c \
 	u_sampler.c \
@@ -148,7 +152,8 @@ C_SOURCES = \
 	u_tile.c \
 	u_transfer.c \
 	u_resource.c \
-	u_upload_mgr.c
+	u_upload_mgr.c \
+	u_vbuf_mgr.c
 
 GALLIVM_SOURCES = \
         lp_bld_arit.c \
@@ -156,7 +161,6 @@ GALLIVM_SOURCES = \
         lp_bld_bitarit.c \
         lp_bld_const.c \
         lp_bld_conv.c \
-        lp_bld_debug.c \
         lp_bld_flow.c \
         lp_bld_format_aos.c \
         lp_bld_format_soa.c \
@@ -184,6 +188,7 @@ GALLIVM_SOURCES = \
         draw_pt_fetch_shade_pipeline_llvm.c
 
 GALLIVM_CPP_SOURCES = \
+        lp_bld_debug.cpp \
 	lp_bld_misc.cpp
 
 SRCS=	${C_SOURCES}
diff --git a/lib/libGL/gallium/libgallium/generated/u_format_table.c b/lib/libGL/gallium/libgallium/generated/u_format_table.c
index 40dd09ebf..5d02870c8 100644
--- a/lib/libGL/gallium/libgallium/generated/u_format_table.c
+++ b/lib/libGL/gallium/libgallium/generated/u_format_table.c
@@ -29,6 +29,8 @@
 
 #include "u_format.h"
 #include "u_format_s3tc.h"
+#include "u_format_rgtc.h"
+#include "u_format_latc.h"
 
 
 #include "pipe/p_compiler.h"
@@ -2116,6 +2118,127 @@ util_format_b10g10r10a2_unorm_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_st
    }
 }
 
+union util_format_b2g3r3_unorm {
+   uint8_t value;
+   struct {
+      unsigned b:2;
+      unsigned g:3;
+      unsigned r:3;
+   } chan;
+};
+
+static INLINE void
+util_format_b2g3r3_unorm_unpack_rgba_float(float *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      float *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         uint8_t value = *(const uint8_t *)src;
+         uint8_t b;
+         uint8_t g;
+         uint8_t r;
+         b = (value) & 0x3;
+         g = (value >> 2) & 0x7;
+         r = value >> 5;
+         dst[0] = (float)(r * (1.0f/0x7)); /* r */
+         dst[1] = (float)(g * (1.0f/0x7)); /* g */
+         dst[2] = (float)(b * (1.0f/0x3)); /* b */
+         dst[3] = 1; /* a */
+         src += 1;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_b2g3r3_unorm_pack_rgba_float(uint8_t *dst_row, unsigned dst_stride, const float *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const float *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         uint8_t value = 0;
+         value |= ((uint8_t)(CLAMP(src[2], 0, 1) * 0x3)) & 0x3;
+         value |= (((uint8_t)(CLAMP(src[1], 0, 1) * 0x7)) & 0x7) << 2;
+         value |= ((uint8_t)(CLAMP(src[0], 0, 1) * 0x7)) << 5;
+         *(uint8_t *)dst = value;
+         src += 4;
+         dst += 1;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+static INLINE void
+util_format_b2g3r3_unorm_fetch_rgba_float(float *dst, const uint8_t *src, unsigned i, unsigned j)
+{
+         uint8_t value = *(const uint8_t *)src;
+         uint8_t b;
+         uint8_t g;
+         uint8_t r;
+         b = (value) & 0x3;
+         g = (value >> 2) & 0x7;
+         r = value >> 5;
+         dst[0] = (float)(r * (1.0f/0x7)); /* r */
+         dst[1] = (float)(g * (1.0f/0x7)); /* g */
+         dst[2] = (float)(b * (1.0f/0x3)); /* b */
+         dst[3] = 1; /* a */
+}
+
+static INLINE void
+util_format_b2g3r3_unorm_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      uint8_t *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         uint8_t value = *(const uint8_t *)src;
+         uint8_t b;
+         uint8_t g;
+         uint8_t r;
+         b = (value) & 0x3;
+         g = (value >> 2) & 0x7;
+         r = value >> 5;
+         dst[0] = (uint8_t)(((uint32_t)r) * 0xff / 0x7); /* r */
+         dst[1] = (uint8_t)(((uint32_t)g) * 0xff / 0x7); /* g */
+         dst[2] = (uint8_t)(((uint32_t)b) * 0xff / 0x3); /* b */
+         dst[3] = 255; /* a */
+         src += 1;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_b2g3r3_unorm_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const uint8_t *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         uint8_t value = 0;
+         value |= ((uint8_t)(src[2] >> 6)) & 0x3;
+         value |= (((uint8_t)(src[1] >> 5)) & 0x7) << 2;
+         value |= ((uint8_t)(src[0] >> 5)) << 5;
+         *(uint8_t *)dst = value;
+         src += 4;
+         dst += 1;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
 union util_format_l8_unorm {
    uint8_t value;
    struct {
@@ -2782,6 +2905,2092 @@ util_format_l16_unorm_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, co
    }
 }
 
+union util_format_a16_unorm {
+   uint16_t value;
+   struct {
+      uint16_t a;
+   } chan;
+};
+
+static INLINE void
+util_format_a16_unorm_unpack_rgba_float(float *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      float *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         uint16_t value = *(const uint16_t *)src;
+         uint16_t a;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap16(value);
+#endif
+         a = value;
+         dst[0] = 0; /* r */
+         dst[1] = 0; /* g */
+         dst[2] = 0; /* b */
+         dst[3] = (float)(a * (1.0f/0xffff)); /* a */
+         src += 2;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_a16_unorm_pack_rgba_float(uint8_t *dst_row, unsigned dst_stride, const float *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const float *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         uint16_t value = 0;
+         value |= (uint16_t)(CLAMP(src[3], 0, 1) * 0xffff);
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap16(value);
+#endif
+         *(uint16_t *)dst = value;
+         src += 4;
+         dst += 2;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+static INLINE void
+util_format_a16_unorm_fetch_rgba_float(float *dst, const uint8_t *src, unsigned i, unsigned j)
+{
+         uint16_t value = *(const uint16_t *)src;
+         uint16_t a;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap16(value);
+#endif
+         a = value;
+         dst[0] = 0; /* r */
+         dst[1] = 0; /* g */
+         dst[2] = 0; /* b */
+         dst[3] = (float)(a * (1.0f/0xffff)); /* a */
+}
+
+static INLINE void
+util_format_a16_unorm_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      uint8_t *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         uint16_t value = *(const uint16_t *)src;
+         uint16_t a;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap16(value);
+#endif
+         a = value;
+         dst[0] = 0; /* r */
+         dst[1] = 0; /* g */
+         dst[2] = 0; /* b */
+         dst[3] = (uint8_t)(a >> 8); /* a */
+         src += 2;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_a16_unorm_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const uint8_t *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         uint16_t value = 0;
+         value |= (uint16_t)(((uint32_t)src[3]) * 0xffff / 0xff);
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap16(value);
+#endif
+         *(uint16_t *)dst = value;
+         src += 4;
+         dst += 2;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+union util_format_i16_unorm {
+   uint16_t value;
+   struct {
+      uint16_t rgba;
+   } chan;
+};
+
+static INLINE void
+util_format_i16_unorm_unpack_rgba_float(float *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      float *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         uint16_t value = *(const uint16_t *)src;
+         uint16_t rgba;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap16(value);
+#endif
+         rgba = value;
+         dst[0] = (float)(rgba * (1.0f/0xffff)); /* r */
+         dst[1] = (float)(rgba * (1.0f/0xffff)); /* g */
+         dst[2] = (float)(rgba * (1.0f/0xffff)); /* b */
+         dst[3] = (float)(rgba * (1.0f/0xffff)); /* a */
+         src += 2;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_i16_unorm_pack_rgba_float(uint8_t *dst_row, unsigned dst_stride, const float *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const float *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         uint16_t value = 0;
+         value |= (uint16_t)(CLAMP(src[3], 0, 1) * 0xffff);
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap16(value);
+#endif
+         *(uint16_t *)dst = value;
+         src += 4;
+         dst += 2;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+static INLINE void
+util_format_i16_unorm_fetch_rgba_float(float *dst, const uint8_t *src, unsigned i, unsigned j)
+{
+         uint16_t value = *(const uint16_t *)src;
+         uint16_t rgba;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap16(value);
+#endif
+         rgba = value;
+         dst[0] = (float)(rgba * (1.0f/0xffff)); /* r */
+         dst[1] = (float)(rgba * (1.0f/0xffff)); /* g */
+         dst[2] = (float)(rgba * (1.0f/0xffff)); /* b */
+         dst[3] = (float)(rgba * (1.0f/0xffff)); /* a */
+}
+
+static INLINE void
+util_format_i16_unorm_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      uint8_t *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         uint16_t value = *(const uint16_t *)src;
+         uint16_t rgba;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap16(value);
+#endif
+         rgba = value;
+         dst[0] = (uint8_t)(rgba >> 8); /* r */
+         dst[1] = (uint8_t)(rgba >> 8); /* g */
+         dst[2] = (uint8_t)(rgba >> 8); /* b */
+         dst[3] = (uint8_t)(rgba >> 8); /* a */
+         src += 2;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_i16_unorm_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const uint8_t *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         uint16_t value = 0;
+         value |= (uint16_t)(((uint32_t)src[3]) * 0xffff / 0xff);
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap16(value);
+#endif
+         *(uint16_t *)dst = value;
+         src += 4;
+         dst += 2;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+union util_format_l16a16_unorm {
+   uint32_t value;
+   struct {
+      uint16_t rgb;
+      uint16_t a;
+   } chan;
+};
+
+static INLINE void
+util_format_l16a16_unorm_unpack_rgba_float(float *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      float *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         uint32_t value = *(const uint32_t *)src;
+         uint32_t rgb;
+         uint32_t a;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap32(value);
+#endif
+         rgb = (value) & 0xffff;
+         a = value >> 16;
+         dst[0] = (float)(rgb * (1.0f/0xffff)); /* r */
+         dst[1] = (float)(rgb * (1.0f/0xffff)); /* g */
+         dst[2] = (float)(rgb * (1.0f/0xffff)); /* b */
+         dst[3] = (float)(a * (1.0f/0xffff)); /* a */
+         src += 4;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_l16a16_unorm_pack_rgba_float(uint8_t *dst_row, unsigned dst_stride, const float *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const float *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         uint32_t value = 0;
+         value |= ((uint16_t)(CLAMP(src[2], 0, 1) * 0xffff)) & 0xffff;
+         value |= ((uint16_t)(CLAMP(src[3], 0, 1) * 0xffff)) << 16;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap32(value);
+#endif
+         *(uint32_t *)dst = value;
+         src += 4;
+         dst += 4;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+static INLINE void
+util_format_l16a16_unorm_fetch_rgba_float(float *dst, const uint8_t *src, unsigned i, unsigned j)
+{
+         uint32_t value = *(const uint32_t *)src;
+         uint32_t rgb;
+         uint32_t a;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap32(value);
+#endif
+         rgb = (value) & 0xffff;
+         a = value >> 16;
+         dst[0] = (float)(rgb * (1.0f/0xffff)); /* r */
+         dst[1] = (float)(rgb * (1.0f/0xffff)); /* g */
+         dst[2] = (float)(rgb * (1.0f/0xffff)); /* b */
+         dst[3] = (float)(a * (1.0f/0xffff)); /* a */
+}
+
+static INLINE void
+util_format_l16a16_unorm_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      uint8_t *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         uint32_t value = *(const uint32_t *)src;
+         uint32_t rgb;
+         uint32_t a;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap32(value);
+#endif
+         rgb = (value) & 0xffff;
+         a = value >> 16;
+         dst[0] = (uint8_t)(rgb >> 8); /* r */
+         dst[1] = (uint8_t)(rgb >> 8); /* g */
+         dst[2] = (uint8_t)(rgb >> 8); /* b */
+         dst[3] = (uint8_t)(a >> 8); /* a */
+         src += 4;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_l16a16_unorm_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const uint8_t *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         uint32_t value = 0;
+         value |= ((uint16_t)(((uint32_t)src[2]) * 0xffff / 0xff)) & 0xffff;
+         value |= ((uint16_t)(((uint32_t)src[3]) * 0xffff / 0xff)) << 16;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap32(value);
+#endif
+         *(uint32_t *)dst = value;
+         src += 4;
+         dst += 4;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+union util_format_a8_snorm {
+   uint8_t value;
+   struct {
+      int8_t a;
+   } chan;
+};
+
+static INLINE void
+util_format_a8_snorm_unpack_rgba_float(float *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      float *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         uint8_t value = *(const uint8_t *)src;
+         int8_t a;
+         a = (int8_t)(value) ;
+         dst[0] = 0; /* r */
+         dst[1] = 0; /* g */
+         dst[2] = 0; /* b */
+         dst[3] = (float)(a * (1.0f/0x7f)); /* a */
+         src += 1;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_a8_snorm_pack_rgba_float(uint8_t *dst_row, unsigned dst_stride, const float *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const float *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         uint8_t value = 0;
+         value |= (uint8_t)((int8_t)(CLAMP(src[3], -1, 1) * 0x7f)) ;
+         *(uint8_t *)dst = value;
+         src += 4;
+         dst += 1;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+static INLINE void
+util_format_a8_snorm_fetch_rgba_float(float *dst, const uint8_t *src, unsigned i, unsigned j)
+{
+         uint8_t value = *(const uint8_t *)src;
+         int8_t a;
+         a = (int8_t)(value) ;
+         dst[0] = 0; /* r */
+         dst[1] = 0; /* g */
+         dst[2] = 0; /* b */
+         dst[3] = (float)(a * (1.0f/0x7f)); /* a */
+}
+
+static INLINE void
+util_format_a8_snorm_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      uint8_t *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         uint8_t value = *(const uint8_t *)src;
+         int8_t a;
+         a = (int8_t)(value) ;
+         dst[0] = 0; /* r */
+         dst[1] = 0; /* g */
+         dst[2] = 0; /* b */
+         dst[3] = (uint8_t)(((uint32_t)MAX2(a, 0)) * 0xff / 0x7f); /* a */
+         src += 1;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_a8_snorm_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const uint8_t *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         uint8_t value = 0;
+         value |= (uint8_t)((int8_t)(src[3] >> 1)) ;
+         *(uint8_t *)dst = value;
+         src += 4;
+         dst += 1;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+union util_format_l8_snorm {
+   uint8_t value;
+   struct {
+      int8_t rgb;
+   } chan;
+};
+
+static INLINE void
+util_format_l8_snorm_unpack_rgba_float(float *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      float *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         uint8_t value = *(const uint8_t *)src;
+         int8_t rgb;
+         rgb = (int8_t)(value) ;
+         dst[0] = (float)(rgb * (1.0f/0x7f)); /* r */
+         dst[1] = (float)(rgb * (1.0f/0x7f)); /* g */
+         dst[2] = (float)(rgb * (1.0f/0x7f)); /* b */
+         dst[3] = 1; /* a */
+         src += 1;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_l8_snorm_pack_rgba_float(uint8_t *dst_row, unsigned dst_stride, const float *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const float *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         uint8_t value = 0;
+         value |= (uint8_t)((int8_t)(CLAMP(src[2], -1, 1) * 0x7f)) ;
+         *(uint8_t *)dst = value;
+         src += 4;
+         dst += 1;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+static INLINE void
+util_format_l8_snorm_fetch_rgba_float(float *dst, const uint8_t *src, unsigned i, unsigned j)
+{
+         uint8_t value = *(const uint8_t *)src;
+         int8_t rgb;
+         rgb = (int8_t)(value) ;
+         dst[0] = (float)(rgb * (1.0f/0x7f)); /* r */
+         dst[1] = (float)(rgb * (1.0f/0x7f)); /* g */
+         dst[2] = (float)(rgb * (1.0f/0x7f)); /* b */
+         dst[3] = 1; /* a */
+}
+
+static INLINE void
+util_format_l8_snorm_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      uint8_t *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         uint8_t value = *(const uint8_t *)src;
+         int8_t rgb;
+         rgb = (int8_t)(value) ;
+         dst[0] = (uint8_t)(((uint32_t)MAX2(rgb, 0)) * 0xff / 0x7f); /* r */
+         dst[1] = (uint8_t)(((uint32_t)MAX2(rgb, 0)) * 0xff / 0x7f); /* g */
+         dst[2] = (uint8_t)(((uint32_t)MAX2(rgb, 0)) * 0xff / 0x7f); /* b */
+         dst[3] = 255; /* a */
+         src += 1;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_l8_snorm_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const uint8_t *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         uint8_t value = 0;
+         value |= (uint8_t)((int8_t)(src[2] >> 1)) ;
+         *(uint8_t *)dst = value;
+         src += 4;
+         dst += 1;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+union util_format_l8a8_snorm {
+   uint16_t value;
+   struct {
+      int8_t rgb;
+      int8_t a;
+   } chan;
+};
+
+static INLINE void
+util_format_l8a8_snorm_unpack_rgba_float(float *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      float *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         uint16_t value = *(const uint16_t *)src;
+         int16_t rgb;
+         int16_t a;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap16(value);
+#endif
+         rgb = ((int16_t)(value << 8) ) >> 8;
+         a = ((int16_t)(value) ) >> 8;
+         dst[0] = (float)(rgb * (1.0f/0x7f)); /* r */
+         dst[1] = (float)(rgb * (1.0f/0x7f)); /* g */
+         dst[2] = (float)(rgb * (1.0f/0x7f)); /* b */
+         dst[3] = (float)(a * (1.0f/0x7f)); /* a */
+         src += 2;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_l8a8_snorm_pack_rgba_float(uint8_t *dst_row, unsigned dst_stride, const float *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const float *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         uint16_t value = 0;
+         value |= (uint16_t)(((int8_t)(CLAMP(src[2], -1, 1) * 0x7f)) & 0xff) ;
+         value |= (uint16_t)(((int8_t)(CLAMP(src[3], -1, 1) * 0x7f)) << 8) ;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap16(value);
+#endif
+         *(uint16_t *)dst = value;
+         src += 4;
+         dst += 2;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+static INLINE void
+util_format_l8a8_snorm_fetch_rgba_float(float *dst, const uint8_t *src, unsigned i, unsigned j)
+{
+         uint16_t value = *(const uint16_t *)src;
+         int16_t rgb;
+         int16_t a;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap16(value);
+#endif
+         rgb = ((int16_t)(value << 8) ) >> 8;
+         a = ((int16_t)(value) ) >> 8;
+         dst[0] = (float)(rgb * (1.0f/0x7f)); /* r */
+         dst[1] = (float)(rgb * (1.0f/0x7f)); /* g */
+         dst[2] = (float)(rgb * (1.0f/0x7f)); /* b */
+         dst[3] = (float)(a * (1.0f/0x7f)); /* a */
+}
+
+static INLINE void
+util_format_l8a8_snorm_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      uint8_t *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         uint16_t value = *(const uint16_t *)src;
+         int16_t rgb;
+         int16_t a;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap16(value);
+#endif
+         rgb = ((int16_t)(value << 8) ) >> 8;
+         a = ((int16_t)(value) ) >> 8;
+         dst[0] = (uint8_t)(((uint32_t)MAX2(rgb, 0)) * 0xff / 0x7f); /* r */
+         dst[1] = (uint8_t)(((uint32_t)MAX2(rgb, 0)) * 0xff / 0x7f); /* g */
+         dst[2] = (uint8_t)(((uint32_t)MAX2(rgb, 0)) * 0xff / 0x7f); /* b */
+         dst[3] = (uint8_t)(((uint32_t)MAX2(a, 0)) * 0xff / 0x7f); /* a */
+         src += 2;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_l8a8_snorm_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const uint8_t *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         uint16_t value = 0;
+         value |= (uint16_t)(((int8_t)(src[2] >> 1)) & 0xff) ;
+         value |= (uint16_t)(((int8_t)(src[3] >> 1)) << 8) ;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap16(value);
+#endif
+         *(uint16_t *)dst = value;
+         src += 4;
+         dst += 2;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+union util_format_i8_snorm {
+   uint8_t value;
+   struct {
+      int8_t rgba;
+   } chan;
+};
+
+static INLINE void
+util_format_i8_snorm_unpack_rgba_float(float *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      float *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         uint8_t value = *(const uint8_t *)src;
+         int8_t rgba;
+         rgba = (int8_t)(value) ;
+         dst[0] = (float)(rgba * (1.0f/0x7f)); /* r */
+         dst[1] = (float)(rgba * (1.0f/0x7f)); /* g */
+         dst[2] = (float)(rgba * (1.0f/0x7f)); /* b */
+         dst[3] = (float)(rgba * (1.0f/0x7f)); /* a */
+         src += 1;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_i8_snorm_pack_rgba_float(uint8_t *dst_row, unsigned dst_stride, const float *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const float *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         uint8_t value = 0;
+         value |= (uint8_t)((int8_t)(CLAMP(src[3], -1, 1) * 0x7f)) ;
+         *(uint8_t *)dst = value;
+         src += 4;
+         dst += 1;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+static INLINE void
+util_format_i8_snorm_fetch_rgba_float(float *dst, const uint8_t *src, unsigned i, unsigned j)
+{
+         uint8_t value = *(const uint8_t *)src;
+         int8_t rgba;
+         rgba = (int8_t)(value) ;
+         dst[0] = (float)(rgba * (1.0f/0x7f)); /* r */
+         dst[1] = (float)(rgba * (1.0f/0x7f)); /* g */
+         dst[2] = (float)(rgba * (1.0f/0x7f)); /* b */
+         dst[3] = (float)(rgba * (1.0f/0x7f)); /* a */
+}
+
+static INLINE void
+util_format_i8_snorm_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      uint8_t *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         uint8_t value = *(const uint8_t *)src;
+         int8_t rgba;
+         rgba = (int8_t)(value) ;
+         dst[0] = (uint8_t)(((uint32_t)MAX2(rgba, 0)) * 0xff / 0x7f); /* r */
+         dst[1] = (uint8_t)(((uint32_t)MAX2(rgba, 0)) * 0xff / 0x7f); /* g */
+         dst[2] = (uint8_t)(((uint32_t)MAX2(rgba, 0)) * 0xff / 0x7f); /* b */
+         dst[3] = (uint8_t)(((uint32_t)MAX2(rgba, 0)) * 0xff / 0x7f); /* a */
+         src += 1;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_i8_snorm_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const uint8_t *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         uint8_t value = 0;
+         value |= (uint8_t)((int8_t)(src[3] >> 1)) ;
+         *(uint8_t *)dst = value;
+         src += 4;
+         dst += 1;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+union util_format_a16_snorm {
+   uint16_t value;
+   struct {
+      int16_t a;
+   } chan;
+};
+
+static INLINE void
+util_format_a16_snorm_unpack_rgba_float(float *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      float *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         uint16_t value = *(const uint16_t *)src;
+         int16_t a;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap16(value);
+#endif
+         a = (int16_t)(value) ;
+         dst[0] = 0; /* r */
+         dst[1] = 0; /* g */
+         dst[2] = 0; /* b */
+         dst[3] = (float)(a * (1.0f/0x7fff)); /* a */
+         src += 2;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_a16_snorm_pack_rgba_float(uint8_t *dst_row, unsigned dst_stride, const float *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const float *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         uint16_t value = 0;
+         value |= (uint16_t)((int16_t)(CLAMP(src[3], -1, 1) * 0x7fff)) ;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap16(value);
+#endif
+         *(uint16_t *)dst = value;
+         src += 4;
+         dst += 2;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+static INLINE void
+util_format_a16_snorm_fetch_rgba_float(float *dst, const uint8_t *src, unsigned i, unsigned j)
+{
+         uint16_t value = *(const uint16_t *)src;
+         int16_t a;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap16(value);
+#endif
+         a = (int16_t)(value) ;
+         dst[0] = 0; /* r */
+         dst[1] = 0; /* g */
+         dst[2] = 0; /* b */
+         dst[3] = (float)(a * (1.0f/0x7fff)); /* a */
+}
+
+static INLINE void
+util_format_a16_snorm_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      uint8_t *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         uint16_t value = *(const uint16_t *)src;
+         int16_t a;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap16(value);
+#endif
+         a = (int16_t)(value) ;
+         dst[0] = 0; /* r */
+         dst[1] = 0; /* g */
+         dst[2] = 0; /* b */
+         dst[3] = (uint8_t)(MAX2(a, 0) >> 7); /* a */
+         src += 2;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_a16_snorm_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const uint8_t *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         uint16_t value = 0;
+         value |= (uint16_t)((int16_t)(((uint32_t)src[3]) * 0x7fff / 0xff)) ;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap16(value);
+#endif
+         *(uint16_t *)dst = value;
+         src += 4;
+         dst += 2;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+union util_format_l16_snorm {
+   uint16_t value;
+   struct {
+      int16_t rgb;
+   } chan;
+};
+
+static INLINE void
+util_format_l16_snorm_unpack_rgba_float(float *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      float *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         uint16_t value = *(const uint16_t *)src;
+         int16_t rgb;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap16(value);
+#endif
+         rgb = (int16_t)(value) ;
+         dst[0] = (float)(rgb * (1.0f/0x7fff)); /* r */
+         dst[1] = (float)(rgb * (1.0f/0x7fff)); /* g */
+         dst[2] = (float)(rgb * (1.0f/0x7fff)); /* b */
+         dst[3] = 1; /* a */
+         src += 2;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_l16_snorm_pack_rgba_float(uint8_t *dst_row, unsigned dst_stride, const float *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const float *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         uint16_t value = 0;
+         value |= (uint16_t)((int16_t)(CLAMP(src[2], -1, 1) * 0x7fff)) ;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap16(value);
+#endif
+         *(uint16_t *)dst = value;
+         src += 4;
+         dst += 2;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+static INLINE void
+util_format_l16_snorm_fetch_rgba_float(float *dst, const uint8_t *src, unsigned i, unsigned j)
+{
+         uint16_t value = *(const uint16_t *)src;
+         int16_t rgb;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap16(value);
+#endif
+         rgb = (int16_t)(value) ;
+         dst[0] = (float)(rgb * (1.0f/0x7fff)); /* r */
+         dst[1] = (float)(rgb * (1.0f/0x7fff)); /* g */
+         dst[2] = (float)(rgb * (1.0f/0x7fff)); /* b */
+         dst[3] = 1; /* a */
+}
+
+static INLINE void
+util_format_l16_snorm_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      uint8_t *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         uint16_t value = *(const uint16_t *)src;
+         int16_t rgb;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap16(value);
+#endif
+         rgb = (int16_t)(value) ;
+         dst[0] = (uint8_t)(MAX2(rgb, 0) >> 7); /* r */
+         dst[1] = (uint8_t)(MAX2(rgb, 0) >> 7); /* g */
+         dst[2] = (uint8_t)(MAX2(rgb, 0) >> 7); /* b */
+         dst[3] = 255; /* a */
+         src += 2;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_l16_snorm_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const uint8_t *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         uint16_t value = 0;
+         value |= (uint16_t)((int16_t)(((uint32_t)src[2]) * 0x7fff / 0xff)) ;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap16(value);
+#endif
+         *(uint16_t *)dst = value;
+         src += 4;
+         dst += 2;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+union util_format_l16a16_snorm {
+   uint32_t value;
+   struct {
+      int16_t rgb;
+      int16_t a;
+   } chan;
+};
+
+static INLINE void
+util_format_l16a16_snorm_unpack_rgba_float(float *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      float *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         uint32_t value = *(const uint32_t *)src;
+         int32_t rgb;
+         int32_t a;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap32(value);
+#endif
+         rgb = ((int32_t)(value << 16) ) >> 16;
+         a = ((int32_t)(value) ) >> 16;
+         dst[0] = (float)(rgb * (1.0f/0x7fff)); /* r */
+         dst[1] = (float)(rgb * (1.0f/0x7fff)); /* g */
+         dst[2] = (float)(rgb * (1.0f/0x7fff)); /* b */
+         dst[3] = (float)(a * (1.0f/0x7fff)); /* a */
+         src += 4;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_l16a16_snorm_pack_rgba_float(uint8_t *dst_row, unsigned dst_stride, const float *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const float *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         uint32_t value = 0;
+         value |= (uint32_t)(((int16_t)(CLAMP(src[2], -1, 1) * 0x7fff)) & 0xffff) ;
+         value |= (uint32_t)(((int16_t)(CLAMP(src[3], -1, 1) * 0x7fff)) << 16) ;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap32(value);
+#endif
+         *(uint32_t *)dst = value;
+         src += 4;
+         dst += 4;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+static INLINE void
+util_format_l16a16_snorm_fetch_rgba_float(float *dst, const uint8_t *src, unsigned i, unsigned j)
+{
+         uint32_t value = *(const uint32_t *)src;
+         int32_t rgb;
+         int32_t a;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap32(value);
+#endif
+         rgb = ((int32_t)(value << 16) ) >> 16;
+         a = ((int32_t)(value) ) >> 16;
+         dst[0] = (float)(rgb * (1.0f/0x7fff)); /* r */
+         dst[1] = (float)(rgb * (1.0f/0x7fff)); /* g */
+         dst[2] = (float)(rgb * (1.0f/0x7fff)); /* b */
+         dst[3] = (float)(a * (1.0f/0x7fff)); /* a */
+}
+
+static INLINE void
+util_format_l16a16_snorm_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      uint8_t *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         uint32_t value = *(const uint32_t *)src;
+         int32_t rgb;
+         int32_t a;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap32(value);
+#endif
+         rgb = ((int32_t)(value << 16) ) >> 16;
+         a = ((int32_t)(value) ) >> 16;
+         dst[0] = (uint8_t)(MAX2(rgb, 0) >> 7); /* r */
+         dst[1] = (uint8_t)(MAX2(rgb, 0) >> 7); /* g */
+         dst[2] = (uint8_t)(MAX2(rgb, 0) >> 7); /* b */
+         dst[3] = (uint8_t)(MAX2(a, 0) >> 7); /* a */
+         src += 4;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_l16a16_snorm_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const uint8_t *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         uint32_t value = 0;
+         value |= (uint32_t)(((int16_t)(((uint32_t)src[2]) * 0x7fff / 0xff)) & 0xffff) ;
+         value |= (uint32_t)(((int16_t)(((uint32_t)src[3]) * 0x7fff / 0xff)) << 16) ;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap32(value);
+#endif
+         *(uint32_t *)dst = value;
+         src += 4;
+         dst += 4;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+union util_format_i16_snorm {
+   uint16_t value;
+   struct {
+      int16_t rgba;
+   } chan;
+};
+
+static INLINE void
+util_format_i16_snorm_unpack_rgba_float(float *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      float *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         uint16_t value = *(const uint16_t *)src;
+         int16_t rgba;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap16(value);
+#endif
+         rgba = (int16_t)(value) ;
+         dst[0] = (float)(rgba * (1.0f/0x7fff)); /* r */
+         dst[1] = (float)(rgba * (1.0f/0x7fff)); /* g */
+         dst[2] = (float)(rgba * (1.0f/0x7fff)); /* b */
+         dst[3] = (float)(rgba * (1.0f/0x7fff)); /* a */
+         src += 2;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_i16_snorm_pack_rgba_float(uint8_t *dst_row, unsigned dst_stride, const float *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const float *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         uint16_t value = 0;
+         value |= (uint16_t)((int16_t)(CLAMP(src[3], -1, 1) * 0x7fff)) ;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap16(value);
+#endif
+         *(uint16_t *)dst = value;
+         src += 4;
+         dst += 2;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+static INLINE void
+util_format_i16_snorm_fetch_rgba_float(float *dst, const uint8_t *src, unsigned i, unsigned j)
+{
+         uint16_t value = *(const uint16_t *)src;
+         int16_t rgba;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap16(value);
+#endif
+         rgba = (int16_t)(value) ;
+         dst[0] = (float)(rgba * (1.0f/0x7fff)); /* r */
+         dst[1] = (float)(rgba * (1.0f/0x7fff)); /* g */
+         dst[2] = (float)(rgba * (1.0f/0x7fff)); /* b */
+         dst[3] = (float)(rgba * (1.0f/0x7fff)); /* a */
+}
+
+static INLINE void
+util_format_i16_snorm_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      uint8_t *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         uint16_t value = *(const uint16_t *)src;
+         int16_t rgba;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap16(value);
+#endif
+         rgba = (int16_t)(value) ;
+         dst[0] = (uint8_t)(MAX2(rgba, 0) >> 7); /* r */
+         dst[1] = (uint8_t)(MAX2(rgba, 0) >> 7); /* g */
+         dst[2] = (uint8_t)(MAX2(rgba, 0) >> 7); /* b */
+         dst[3] = (uint8_t)(MAX2(rgba, 0) >> 7); /* a */
+         src += 2;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_i16_snorm_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const uint8_t *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         uint16_t value = 0;
+         value |= (uint16_t)((int16_t)(((uint32_t)src[3]) * 0x7fff / 0xff)) ;
+#ifdef PIPE_ARCH_BIG_ENDIAN
+         value = util_bswap16(value);
+#endif
+         *(uint16_t *)dst = value;
+         src += 4;
+         dst += 2;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+union util_format_a16_float {
+   uint16_t value;
+   struct {
+      uint16_t a;
+   } chan;
+};
+
+static INLINE void
+util_format_a16_float_unpack_rgba_float(float *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      float *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_a16_float pixel;
+         memcpy(&pixel, src, sizeof pixel);
+         dst[0] = 0; /* r */
+         dst[1] = 0; /* g */
+         dst[2] = 0; /* b */
+         dst[3] = util_half_to_float(pixel.chan.a); /* a */
+         src += 2;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_a16_float_pack_rgba_float(uint8_t *dst_row, unsigned dst_stride, const float *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const float *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_a16_float pixel;
+         pixel.chan.a = util_float_to_half(src[3]);
+         memcpy(dst, &pixel, sizeof pixel);
+         src += 4;
+         dst += 2;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+static INLINE void
+util_format_a16_float_fetch_rgba_float(float *dst, const uint8_t *src, unsigned i, unsigned j)
+{
+         union util_format_a16_float pixel;
+         memcpy(&pixel, src, sizeof pixel);
+         dst[0] = 0; /* r */
+         dst[1] = 0; /* g */
+         dst[2] = 0; /* b */
+         dst[3] = util_half_to_float(pixel.chan.a); /* a */
+}
+
+static INLINE void
+util_format_a16_float_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      uint8_t *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_a16_float pixel;
+         memcpy(&pixel, src, sizeof pixel);
+         dst[0] = 0; /* r */
+         dst[1] = 0; /* g */
+         dst[2] = 0; /* b */
+         dst[3] = float_to_ubyte(util_half_to_float(pixel.chan.a)); /* a */
+         src += 2;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_a16_float_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const uint8_t *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_a16_float pixel;
+         pixel.chan.a = util_float_to_half((float)(src[3] * (1.0f/0xff)));
+         memcpy(dst, &pixel, sizeof pixel);
+         src += 4;
+         dst += 2;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+union util_format_l16_float {
+   uint16_t value;
+   struct {
+      uint16_t rgb;
+   } chan;
+};
+
+static INLINE void
+util_format_l16_float_unpack_rgba_float(float *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      float *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_l16_float pixel;
+         memcpy(&pixel, src, sizeof pixel);
+         dst[0] = util_half_to_float(pixel.chan.rgb); /* r */
+         dst[1] = util_half_to_float(pixel.chan.rgb); /* g */
+         dst[2] = util_half_to_float(pixel.chan.rgb); /* b */
+         dst[3] = 1; /* a */
+         src += 2;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_l16_float_pack_rgba_float(uint8_t *dst_row, unsigned dst_stride, const float *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const float *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_l16_float pixel;
+         pixel.chan.rgb = util_float_to_half(src[2]);
+         memcpy(dst, &pixel, sizeof pixel);
+         src += 4;
+         dst += 2;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+static INLINE void
+util_format_l16_float_fetch_rgba_float(float *dst, const uint8_t *src, unsigned i, unsigned j)
+{
+         union util_format_l16_float pixel;
+         memcpy(&pixel, src, sizeof pixel);
+         dst[0] = util_half_to_float(pixel.chan.rgb); /* r */
+         dst[1] = util_half_to_float(pixel.chan.rgb); /* g */
+         dst[2] = util_half_to_float(pixel.chan.rgb); /* b */
+         dst[3] = 1; /* a */
+}
+
+static INLINE void
+util_format_l16_float_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      uint8_t *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_l16_float pixel;
+         memcpy(&pixel, src, sizeof pixel);
+         dst[0] = float_to_ubyte(util_half_to_float(pixel.chan.rgb)); /* r */
+         dst[1] = float_to_ubyte(util_half_to_float(pixel.chan.rgb)); /* g */
+         dst[2] = float_to_ubyte(util_half_to_float(pixel.chan.rgb)); /* b */
+         dst[3] = 255; /* a */
+         src += 2;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_l16_float_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const uint8_t *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_l16_float pixel;
+         pixel.chan.rgb = util_float_to_half((float)(src[2] * (1.0f/0xff)));
+         memcpy(dst, &pixel, sizeof pixel);
+         src += 4;
+         dst += 2;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+union util_format_l16a16_float {
+   uint32_t value;
+   struct {
+      uint16_t rgb;
+      uint16_t a;
+   } chan;
+};
+
+static INLINE void
+util_format_l16a16_float_unpack_rgba_float(float *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      float *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_l16a16_float pixel;
+         memcpy(&pixel, src, sizeof pixel);
+         dst[0] = util_half_to_float(pixel.chan.rgb); /* r */
+         dst[1] = util_half_to_float(pixel.chan.rgb); /* g */
+         dst[2] = util_half_to_float(pixel.chan.rgb); /* b */
+         dst[3] = util_half_to_float(pixel.chan.a); /* a */
+         src += 4;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_l16a16_float_pack_rgba_float(uint8_t *dst_row, unsigned dst_stride, const float *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const float *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_l16a16_float pixel;
+         pixel.chan.rgb = util_float_to_half(src[2]);
+         pixel.chan.a = util_float_to_half(src[3]);
+         memcpy(dst, &pixel, sizeof pixel);
+         src += 4;
+         dst += 4;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+static INLINE void
+util_format_l16a16_float_fetch_rgba_float(float *dst, const uint8_t *src, unsigned i, unsigned j)
+{
+         union util_format_l16a16_float pixel;
+         memcpy(&pixel, src, sizeof pixel);
+         dst[0] = util_half_to_float(pixel.chan.rgb); /* r */
+         dst[1] = util_half_to_float(pixel.chan.rgb); /* g */
+         dst[2] = util_half_to_float(pixel.chan.rgb); /* b */
+         dst[3] = util_half_to_float(pixel.chan.a); /* a */
+}
+
+static INLINE void
+util_format_l16a16_float_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      uint8_t *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_l16a16_float pixel;
+         memcpy(&pixel, src, sizeof pixel);
+         dst[0] = float_to_ubyte(util_half_to_float(pixel.chan.rgb)); /* r */
+         dst[1] = float_to_ubyte(util_half_to_float(pixel.chan.rgb)); /* g */
+         dst[2] = float_to_ubyte(util_half_to_float(pixel.chan.rgb)); /* b */
+         dst[3] = float_to_ubyte(util_half_to_float(pixel.chan.a)); /* a */
+         src += 4;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_l16a16_float_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const uint8_t *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_l16a16_float pixel;
+         pixel.chan.rgb = util_float_to_half((float)(src[2] * (1.0f/0xff)));
+         pixel.chan.a = util_float_to_half((float)(src[3] * (1.0f/0xff)));
+         memcpy(dst, &pixel, sizeof pixel);
+         src += 4;
+         dst += 4;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+union util_format_i16_float {
+   uint16_t value;
+   struct {
+      uint16_t rgba;
+   } chan;
+};
+
+static INLINE void
+util_format_i16_float_unpack_rgba_float(float *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      float *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_i16_float pixel;
+         memcpy(&pixel, src, sizeof pixel);
+         dst[0] = util_half_to_float(pixel.chan.rgba); /* r */
+         dst[1] = util_half_to_float(pixel.chan.rgba); /* g */
+         dst[2] = util_half_to_float(pixel.chan.rgba); /* b */
+         dst[3] = util_half_to_float(pixel.chan.rgba); /* a */
+         src += 2;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_i16_float_pack_rgba_float(uint8_t *dst_row, unsigned dst_stride, const float *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const float *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_i16_float pixel;
+         pixel.chan.rgba = util_float_to_half(src[3]);
+         memcpy(dst, &pixel, sizeof pixel);
+         src += 4;
+         dst += 2;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+static INLINE void
+util_format_i16_float_fetch_rgba_float(float *dst, const uint8_t *src, unsigned i, unsigned j)
+{
+         union util_format_i16_float pixel;
+         memcpy(&pixel, src, sizeof pixel);
+         dst[0] = util_half_to_float(pixel.chan.rgba); /* r */
+         dst[1] = util_half_to_float(pixel.chan.rgba); /* g */
+         dst[2] = util_half_to_float(pixel.chan.rgba); /* b */
+         dst[3] = util_half_to_float(pixel.chan.rgba); /* a */
+}
+
+static INLINE void
+util_format_i16_float_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      uint8_t *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_i16_float pixel;
+         memcpy(&pixel, src, sizeof pixel);
+         dst[0] = float_to_ubyte(util_half_to_float(pixel.chan.rgba)); /* r */
+         dst[1] = float_to_ubyte(util_half_to_float(pixel.chan.rgba)); /* g */
+         dst[2] = float_to_ubyte(util_half_to_float(pixel.chan.rgba)); /* b */
+         dst[3] = float_to_ubyte(util_half_to_float(pixel.chan.rgba)); /* a */
+         src += 2;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_i16_float_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const uint8_t *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_i16_float pixel;
+         pixel.chan.rgba = util_float_to_half((float)(src[3] * (1.0f/0xff)));
+         memcpy(dst, &pixel, sizeof pixel);
+         src += 4;
+         dst += 2;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+union util_format_a32_float {
+   uint32_t value;
+   struct {
+      float a;
+   } chan;
+};
+
+static INLINE void
+util_format_a32_float_unpack_rgba_float(float *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      float *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_a32_float pixel;
+         memcpy(&pixel, src, sizeof pixel);
+         dst[0] = 0; /* r */
+         dst[1] = 0; /* g */
+         dst[2] = 0; /* b */
+         dst[3] = pixel.chan.a; /* a */
+         src += 4;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_a32_float_pack_rgba_float(uint8_t *dst_row, unsigned dst_stride, const float *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const float *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_a32_float pixel;
+         pixel.chan.a = src[3];
+         memcpy(dst, &pixel, sizeof pixel);
+         src += 4;
+         dst += 4;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+static INLINE void
+util_format_a32_float_fetch_rgba_float(float *dst, const uint8_t *src, unsigned i, unsigned j)
+{
+         union util_format_a32_float pixel;
+         memcpy(&pixel, src, sizeof pixel);
+         dst[0] = 0; /* r */
+         dst[1] = 0; /* g */
+         dst[2] = 0; /* b */
+         dst[3] = pixel.chan.a; /* a */
+}
+
+static INLINE void
+util_format_a32_float_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      uint8_t *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_a32_float pixel;
+         memcpy(&pixel, src, sizeof pixel);
+         dst[0] = 0; /* r */
+         dst[1] = 0; /* g */
+         dst[2] = 0; /* b */
+         dst[3] = float_to_ubyte(pixel.chan.a); /* a */
+         src += 4;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_a32_float_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const uint8_t *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_a32_float pixel;
+         pixel.chan.a = ubyte_to_float(src[3]);
+         memcpy(dst, &pixel, sizeof pixel);
+         src += 4;
+         dst += 4;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+union util_format_l32_float {
+   uint32_t value;
+   struct {
+      float rgb;
+   } chan;
+};
+
+static INLINE void
+util_format_l32_float_unpack_rgba_float(float *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      float *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_l32_float pixel;
+         memcpy(&pixel, src, sizeof pixel);
+         dst[0] = pixel.chan.rgb; /* r */
+         dst[1] = pixel.chan.rgb; /* g */
+         dst[2] = pixel.chan.rgb; /* b */
+         dst[3] = 1; /* a */
+         src += 4;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_l32_float_pack_rgba_float(uint8_t *dst_row, unsigned dst_stride, const float *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const float *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_l32_float pixel;
+         pixel.chan.rgb = src[2];
+         memcpy(dst, &pixel, sizeof pixel);
+         src += 4;
+         dst += 4;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+static INLINE void
+util_format_l32_float_fetch_rgba_float(float *dst, const uint8_t *src, unsigned i, unsigned j)
+{
+         union util_format_l32_float pixel;
+         memcpy(&pixel, src, sizeof pixel);
+         dst[0] = pixel.chan.rgb; /* r */
+         dst[1] = pixel.chan.rgb; /* g */
+         dst[2] = pixel.chan.rgb; /* b */
+         dst[3] = 1; /* a */
+}
+
+static INLINE void
+util_format_l32_float_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      uint8_t *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_l32_float pixel;
+         memcpy(&pixel, src, sizeof pixel);
+         dst[0] = float_to_ubyte(pixel.chan.rgb); /* r */
+         dst[1] = float_to_ubyte(pixel.chan.rgb); /* g */
+         dst[2] = float_to_ubyte(pixel.chan.rgb); /* b */
+         dst[3] = 255; /* a */
+         src += 4;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_l32_float_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const uint8_t *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_l32_float pixel;
+         pixel.chan.rgb = ubyte_to_float(src[2]);
+         memcpy(dst, &pixel, sizeof pixel);
+         src += 4;
+         dst += 4;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+union util_format_l32a32_float {
+   uint64_t value;
+   struct {
+      float rgb;
+      float a;
+   } chan;
+};
+
+static INLINE void
+util_format_l32a32_float_unpack_rgba_float(float *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      float *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_l32a32_float pixel;
+         memcpy(&pixel, src, sizeof pixel);
+         dst[0] = pixel.chan.rgb; /* r */
+         dst[1] = pixel.chan.rgb; /* g */
+         dst[2] = pixel.chan.rgb; /* b */
+         dst[3] = pixel.chan.a; /* a */
+         src += 8;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_l32a32_float_pack_rgba_float(uint8_t *dst_row, unsigned dst_stride, const float *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const float *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_l32a32_float pixel;
+         pixel.chan.rgb = src[2];
+         pixel.chan.a = src[3];
+         memcpy(dst, &pixel, sizeof pixel);
+         src += 4;
+         dst += 8;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+static INLINE void
+util_format_l32a32_float_fetch_rgba_float(float *dst, const uint8_t *src, unsigned i, unsigned j)
+{
+         union util_format_l32a32_float pixel;
+         memcpy(&pixel, src, sizeof pixel);
+         dst[0] = pixel.chan.rgb; /* r */
+         dst[1] = pixel.chan.rgb; /* g */
+         dst[2] = pixel.chan.rgb; /* b */
+         dst[3] = pixel.chan.a; /* a */
+}
+
+static INLINE void
+util_format_l32a32_float_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      uint8_t *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_l32a32_float pixel;
+         memcpy(&pixel, src, sizeof pixel);
+         dst[0] = float_to_ubyte(pixel.chan.rgb); /* r */
+         dst[1] = float_to_ubyte(pixel.chan.rgb); /* g */
+         dst[2] = float_to_ubyte(pixel.chan.rgb); /* b */
+         dst[3] = float_to_ubyte(pixel.chan.a); /* a */
+         src += 8;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_l32a32_float_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const uint8_t *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_l32a32_float pixel;
+         pixel.chan.rgb = ubyte_to_float(src[2]);
+         pixel.chan.a = ubyte_to_float(src[3]);
+         memcpy(dst, &pixel, sizeof pixel);
+         src += 4;
+         dst += 8;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+union util_format_i32_float {
+   uint32_t value;
+   struct {
+      float rgba;
+   } chan;
+};
+
+static INLINE void
+util_format_i32_float_unpack_rgba_float(float *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      float *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_i32_float pixel;
+         memcpy(&pixel, src, sizeof pixel);
+         dst[0] = pixel.chan.rgba; /* r */
+         dst[1] = pixel.chan.rgba; /* g */
+         dst[2] = pixel.chan.rgba; /* b */
+         dst[3] = pixel.chan.rgba; /* a */
+         src += 4;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_i32_float_pack_rgba_float(uint8_t *dst_row, unsigned dst_stride, const float *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const float *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_i32_float pixel;
+         pixel.chan.rgba = src[3];
+         memcpy(dst, &pixel, sizeof pixel);
+         src += 4;
+         dst += 4;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
+static INLINE void
+util_format_i32_float_fetch_rgba_float(float *dst, const uint8_t *src, unsigned i, unsigned j)
+{
+         union util_format_i32_float pixel;
+         memcpy(&pixel, src, sizeof pixel);
+         dst[0] = pixel.chan.rgba; /* r */
+         dst[1] = pixel.chan.rgba; /* g */
+         dst[2] = pixel.chan.rgba; /* b */
+         dst[3] = pixel.chan.rgba; /* a */
+}
+
+static INLINE void
+util_format_i32_float_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      uint8_t *dst = dst_row;
+      const uint8_t *src = src_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_i32_float pixel;
+         memcpy(&pixel, src, sizeof pixel);
+         dst[0] = float_to_ubyte(pixel.chan.rgba); /* r */
+         dst[1] = float_to_ubyte(pixel.chan.rgba); /* g */
+         dst[2] = float_to_ubyte(pixel.chan.rgba); /* b */
+         dst[3] = float_to_ubyte(pixel.chan.rgba); /* a */
+         src += 4;
+         dst += 4;
+      }
+      src_row += src_stride;
+      dst_row += dst_stride/sizeof(*dst_row);
+   }
+}
+
+static INLINE void
+util_format_i32_float_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
+{
+   unsigned x, y;
+   for(y = 0; y < height; y += 1) {
+      const uint8_t *src = src_row;
+      uint8_t *dst = dst_row;
+      for(x = 0; x < width; x += 1) {
+         union util_format_i32_float pixel;
+         pixel.chan.rgba = ubyte_to_float(src[3]);
+         memcpy(dst, &pixel, sizeof pixel);
+         src += 4;
+         dst += 4;
+      }
+      dst_row += dst_stride;
+      src_row += src_stride/sizeof(*src_row);
+   }
+}
+
 union util_format_l8_srgb {
    uint8_t value;
    struct {
@@ -4671,131 +6880,6 @@ util_format_r10g10b10a2_uscaled_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_
    }
 }
 
-static INLINE void
-util_format_r11g11b10_float_unpack_rgba_float(float *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
-{
-}
-
-static INLINE void
-util_format_r11g11b10_float_pack_rgba_float(uint8_t *dst_row, unsigned dst_stride, const float *src_row, unsigned src_stride, unsigned width, unsigned height)
-{
-}
-
-static INLINE void
-util_format_r11g11b10_float_fetch_rgba_float(float *dst, const uint8_t *src, unsigned i, unsigned j)
-{
-}
-
-static INLINE void
-util_format_r11g11b10_float_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
-{
-}
-
-static INLINE void
-util_format_r11g11b10_float_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
-{
-}
-
-static INLINE void
-util_format_rgtc1_unorm_unpack_rgba_float(float *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
-{
-}
-
-static INLINE void
-util_format_rgtc1_unorm_pack_rgba_float(uint8_t *dst_row, unsigned dst_stride, const float *src_row, unsigned src_stride, unsigned width, unsigned height)
-{
-}
-
-static INLINE void
-util_format_rgtc1_unorm_fetch_rgba_float(float *dst, const uint8_t *src, unsigned i, unsigned j)
-{
-}
-
-static INLINE void
-util_format_rgtc1_unorm_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
-{
-}
-
-static INLINE void
-util_format_rgtc1_unorm_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
-{
-}
-
-static INLINE void
-util_format_rgtc1_snorm_unpack_rgba_float(float *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
-{
-}
-
-static INLINE void
-util_format_rgtc1_snorm_pack_rgba_float(uint8_t *dst_row, unsigned dst_stride, const float *src_row, unsigned src_stride, unsigned width, unsigned height)
-{
-}
-
-static INLINE void
-util_format_rgtc1_snorm_fetch_rgba_float(float *dst, const uint8_t *src, unsigned i, unsigned j)
-{
-}
-
-static INLINE void
-util_format_rgtc1_snorm_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
-{
-}
-
-static INLINE void
-util_format_rgtc1_snorm_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
-{
-}
-
-static INLINE void
-util_format_rgtc2_unorm_unpack_rgba_float(float *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
-{
-}
-
-static INLINE void
-util_format_rgtc2_unorm_pack_rgba_float(uint8_t *dst_row, unsigned dst_stride, const float *src_row, unsigned src_stride, unsigned width, unsigned height)
-{
-}
-
-static INLINE void
-util_format_rgtc2_unorm_fetch_rgba_float(float *dst, const uint8_t *src, unsigned i, unsigned j)
-{
-}
-
-static INLINE void
-util_format_rgtc2_unorm_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
-{
-}
-
-static INLINE void
-util_format_rgtc2_unorm_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
-{
-}
-
-static INLINE void
-util_format_rgtc2_snorm_unpack_rgba_float(float *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
-{
-}
-
-static INLINE void
-util_format_rgtc2_snorm_pack_rgba_float(uint8_t *dst_row, unsigned dst_stride, const float *src_row, unsigned src_stride, unsigned width, unsigned height)
-{
-}
-
-static INLINE void
-util_format_rgtc2_snorm_fetch_rgba_float(float *dst, const uint8_t *src, unsigned i, unsigned j)
-{
-}
-
-static INLINE void
-util_format_rgtc2_snorm_unpack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
-{
-}
-
-static INLINE void
-util_format_rgtc2_snorm_pack_rgba_8unorm(uint8_t *dst_row, unsigned dst_stride, const uint8_t *src_row, unsigned src_stride, unsigned width, unsigned height)
-{
-}
-
 union util_format_r64_float {
    uint64_t value;
    struct {
@@ -12797,6 +14881,44 @@ util_format_b10g10r10a2_unorm_description = {
 };
 
 const struct util_format_description
+util_format_b2g3r3_unorm_description = {
+   PIPE_FORMAT_B2G3R3_UNORM,
+   "PIPE_FORMAT_B2G3R3_UNORM",
+   "b2g3r3_unorm",
+   {1, 1, 8},	/* block */
+   UTIL_FORMAT_LAYOUT_PLAIN,
+   3,	/* nr_channels */
+   FALSE,	/* is_array */
+   TRUE,	/* is_bitmask */
+   FALSE,	/* is_mixed */
+   {
+      {UTIL_FORMAT_TYPE_UNSIGNED, TRUE, 2},	/* x = b */
+      {UTIL_FORMAT_TYPE_UNSIGNED, TRUE, 3},	/* y = g */
+      {UTIL_FORMAT_TYPE_UNSIGNED, TRUE, 3},	/* z = r */
+      {0, 0, 0}
+   },
+   {
+      UTIL_FORMAT_SWIZZLE_Z,	/* r */
+      UTIL_FORMAT_SWIZZLE_Y,	/* g */
+      UTIL_FORMAT_SWIZZLE_X,	/* b */
+      UTIL_FORMAT_SWIZZLE_1	/* a */
+   },
+   UTIL_FORMAT_COLORSPACE_RGB,
+   &util_format_b2g3r3_unorm_unpack_rgba_8unorm,
+   &util_format_b2g3r3_unorm_pack_rgba_8unorm,
+   NULL, /* fetch_rgba_8unorm */
+   &util_format_b2g3r3_unorm_unpack_rgba_float,
+   &util_format_b2g3r3_unorm_pack_rgba_float,
+   &util_format_b2g3r3_unorm_fetch_rgba_float,
+   NULL, /* unpack_z_32unorm */
+   NULL, /* pack_z_32unorm */
+   NULL, /* unpack_z_float */
+   NULL, /* pack_z_float */
+   NULL, /* unpack_s_8uscaled */
+   NULL /* pack_s_8uscaled */
+};
+
+const struct util_format_description
 util_format_l8_unorm_description = {
    PIPE_FORMAT_L8_UNORM,
    "PIPE_FORMAT_L8_UNORM",
@@ -13025,6 +15147,728 @@ util_format_l16_unorm_description = {
 };
 
 const struct util_format_description
+util_format_a16_unorm_description = {
+   PIPE_FORMAT_A16_UNORM,
+   "PIPE_FORMAT_A16_UNORM",
+   "a16_unorm",
+   {1, 1, 16},	/* block */
+   UTIL_FORMAT_LAYOUT_PLAIN,
+   1,	/* nr_channels */
+   TRUE,	/* is_array */
+   TRUE,	/* is_bitmask */
+   FALSE,	/* is_mixed */
+   {
+      {UTIL_FORMAT_TYPE_UNSIGNED, TRUE, 16},	/* x = a */
+      {0, 0, 0},
+      {0, 0, 0},
+      {0, 0, 0}
+   },
+   {
+      UTIL_FORMAT_SWIZZLE_0,	/* r */
+      UTIL_FORMAT_SWIZZLE_0,	/* g */
+      UTIL_FORMAT_SWIZZLE_0,	/* b */
+      UTIL_FORMAT_SWIZZLE_X	/* a */
+   },
+   UTIL_FORMAT_COLORSPACE_RGB,
+   &util_format_a16_unorm_unpack_rgba_8unorm,
+   &util_format_a16_unorm_pack_rgba_8unorm,
+   NULL, /* fetch_rgba_8unorm */
+   &util_format_a16_unorm_unpack_rgba_float,
+   &util_format_a16_unorm_pack_rgba_float,
+   &util_format_a16_unorm_fetch_rgba_float,
+   NULL, /* unpack_z_32unorm */
+   NULL, /* pack_z_32unorm */
+   NULL, /* unpack_z_float */
+   NULL, /* pack_z_float */
+   NULL, /* unpack_s_8uscaled */
+   NULL /* pack_s_8uscaled */
+};
+
+const struct util_format_description
+util_format_i16_unorm_description = {
+   PIPE_FORMAT_I16_UNORM,
+   "PIPE_FORMAT_I16_UNORM",
+   "i16_unorm",
+   {1, 1, 16},	/* block */
+   UTIL_FORMAT_LAYOUT_PLAIN,
+   1,	/* nr_channels */
+   TRUE,	/* is_array */
+   TRUE,	/* is_bitmask */
+   FALSE,	/* is_mixed */
+   {
+      {UTIL_FORMAT_TYPE_UNSIGNED, TRUE, 16},	/* x = rgba */
+      {0, 0, 0},
+      {0, 0, 0},
+      {0, 0, 0}
+   },
+   {
+      UTIL_FORMAT_SWIZZLE_X,	/* r */
+      UTIL_FORMAT_SWIZZLE_X,	/* g */
+      UTIL_FORMAT_SWIZZLE_X,	/* b */
+      UTIL_FORMAT_SWIZZLE_X	/* a */
+   },
+   UTIL_FORMAT_COLORSPACE_RGB,
+   &util_format_i16_unorm_unpack_rgba_8unorm,
+   &util_format_i16_unorm_pack_rgba_8unorm,
+   NULL, /* fetch_rgba_8unorm */
+   &util_format_i16_unorm_unpack_rgba_float,
+   &util_format_i16_unorm_pack_rgba_float,
+   &util_format_i16_unorm_fetch_rgba_float,
+   NULL, /* unpack_z_32unorm */
+   NULL, /* pack_z_32unorm */
+   NULL, /* unpack_z_float */
+   NULL, /* pack_z_float */
+   NULL, /* unpack_s_8uscaled */
+   NULL /* pack_s_8uscaled */
+};
+
+const struct util_format_description
+util_format_l16a16_unorm_description = {
+   PIPE_FORMAT_L16A16_UNORM,
+   "PIPE_FORMAT_L16A16_UNORM",
+   "l16a16_unorm",
+   {1, 1, 32},	/* block */
+   UTIL_FORMAT_LAYOUT_PLAIN,
+   2,	/* nr_channels */
+   TRUE,	/* is_array */
+   TRUE,	/* is_bitmask */
+   FALSE,	/* is_mixed */
+   {
+      {UTIL_FORMAT_TYPE_UNSIGNED, TRUE, 16},	/* x = rgb */
+      {UTIL_FORMAT_TYPE_UNSIGNED, TRUE, 16},	/* y = a */
+      {0, 0, 0},
+      {0, 0, 0}
+   },
+   {
+      UTIL_FORMAT_SWIZZLE_X,	/* r */
+      UTIL_FORMAT_SWIZZLE_X,	/* g */
+      UTIL_FORMAT_SWIZZLE_X,	/* b */
+      UTIL_FORMAT_SWIZZLE_Y	/* a */
+   },
+   UTIL_FORMAT_COLORSPACE_RGB,
+   &util_format_l16a16_unorm_unpack_rgba_8unorm,
+   &util_format_l16a16_unorm_pack_rgba_8unorm,
+   NULL, /* fetch_rgba_8unorm */
+   &util_format_l16a16_unorm_unpack_rgba_float,
+   &util_format_l16a16_unorm_pack_rgba_float,
+   &util_format_l16a16_unorm_fetch_rgba_float,
+   NULL, /* unpack_z_32unorm */
+   NULL, /* pack_z_32unorm */
+   NULL, /* unpack_z_float */
+   NULL, /* pack_z_float */
+   NULL, /* unpack_s_8uscaled */
+   NULL /* pack_s_8uscaled */
+};
+
+const struct util_format_description
+util_format_a8_snorm_description = {
+   PIPE_FORMAT_A8_SNORM,
+   "PIPE_FORMAT_A8_SNORM",
+   "a8_snorm",
+   {1, 1, 8},	/* block */
+   UTIL_FORMAT_LAYOUT_PLAIN,
+   1,	/* nr_channels */
+   TRUE,	/* is_array */
+   TRUE,	/* is_bitmask */
+   FALSE,	/* is_mixed */
+   {
+      {UTIL_FORMAT_TYPE_SIGNED, TRUE, 8},	/* x = a */
+      {0, 0, 0},
+      {0, 0, 0},
+      {0, 0, 0}
+   },
+   {
+      UTIL_FORMAT_SWIZZLE_0,	/* r */
+      UTIL_FORMAT_SWIZZLE_0,	/* g */
+      UTIL_FORMAT_SWIZZLE_0,	/* b */
+      UTIL_FORMAT_SWIZZLE_X	/* a */
+   },
+   UTIL_FORMAT_COLORSPACE_RGB,
+   &util_format_a8_snorm_unpack_rgba_8unorm,
+   &util_format_a8_snorm_pack_rgba_8unorm,
+   NULL, /* fetch_rgba_8unorm */
+   &util_format_a8_snorm_unpack_rgba_float,
+   &util_format_a8_snorm_pack_rgba_float,
+   &util_format_a8_snorm_fetch_rgba_float,
+   NULL, /* unpack_z_32unorm */
+   NULL, /* pack_z_32unorm */
+   NULL, /* unpack_z_float */
+   NULL, /* pack_z_float */
+   NULL, /* unpack_s_8uscaled */
+   NULL /* pack_s_8uscaled */
+};
+
+const struct util_format_description
+util_format_l8_snorm_description = {
+   PIPE_FORMAT_L8_SNORM,
+   "PIPE_FORMAT_L8_SNORM",
+   "l8_snorm",
+   {1, 1, 8},	/* block */
+   UTIL_FORMAT_LAYOUT_PLAIN,
+   1,	/* nr_channels */
+   TRUE,	/* is_array */
+   TRUE,	/* is_bitmask */
+   FALSE,	/* is_mixed */
+   {
+      {UTIL_FORMAT_TYPE_SIGNED, TRUE, 8},	/* x = rgb */
+      {0, 0, 0},
+      {0, 0, 0},
+      {0, 0, 0}
+   },
+   {
+      UTIL_FORMAT_SWIZZLE_X,	/* r */
+      UTIL_FORMAT_SWIZZLE_X,	/* g */
+      UTIL_FORMAT_SWIZZLE_X,	/* b */
+      UTIL_FORMAT_SWIZZLE_1	/* a */
+   },
+   UTIL_FORMAT_COLORSPACE_RGB,
+   &util_format_l8_snorm_unpack_rgba_8unorm,
+   &util_format_l8_snorm_pack_rgba_8unorm,
+   NULL, /* fetch_rgba_8unorm */
+   &util_format_l8_snorm_unpack_rgba_float,
+   &util_format_l8_snorm_pack_rgba_float,
+   &util_format_l8_snorm_fetch_rgba_float,
+   NULL, /* unpack_z_32unorm */
+   NULL, /* pack_z_32unorm */
+   NULL, /* unpack_z_float */
+   NULL, /* pack_z_float */
+   NULL, /* unpack_s_8uscaled */
+   NULL /* pack_s_8uscaled */
+};
+
+const struct util_format_description
+util_format_l8a8_snorm_description = {
+   PIPE_FORMAT_L8A8_SNORM,
+   "PIPE_FORMAT_L8A8_SNORM",
+   "l8a8_snorm",
+   {1, 1, 16},	/* block */
+   UTIL_FORMAT_LAYOUT_PLAIN,
+   2,	/* nr_channels */
+   TRUE,	/* is_array */
+   TRUE,	/* is_bitmask */
+   FALSE,	/* is_mixed */
+   {
+      {UTIL_FORMAT_TYPE_SIGNED, TRUE, 8},	/* x = rgb */
+      {UTIL_FORMAT_TYPE_SIGNED, TRUE, 8},	/* y = a */
+      {0, 0, 0},
+      {0, 0, 0}
+   },
+   {
+      UTIL_FORMAT_SWIZZLE_X,	/* r */
+      UTIL_FORMAT_SWIZZLE_X,	/* g */
+      UTIL_FORMAT_SWIZZLE_X,	/* b */
+      UTIL_FORMAT_SWIZZLE_Y	/* a */
+   },
+   UTIL_FORMAT_COLORSPACE_RGB,
+   &util_format_l8a8_snorm_unpack_rgba_8unorm,
+   &util_format_l8a8_snorm_pack_rgba_8unorm,
+   NULL, /* fetch_rgba_8unorm */
+   &util_format_l8a8_snorm_unpack_rgba_float,
+   &util_format_l8a8_snorm_pack_rgba_float,
+   &util_format_l8a8_snorm_fetch_rgba_float,
+   NULL, /* unpack_z_32unorm */
+   NULL, /* pack_z_32unorm */
+   NULL, /* unpack_z_float */
+   NULL, /* pack_z_float */
+   NULL, /* unpack_s_8uscaled */
+   NULL /* pack_s_8uscaled */
+};
+
+const struct util_format_description
+util_format_i8_snorm_description = {
+   PIPE_FORMAT_I8_SNORM,
+   "PIPE_FORMAT_I8_SNORM",
+   "i8_snorm",
+   {1, 1, 8},	/* block */
+   UTIL_FORMAT_LAYOUT_PLAIN,
+   1,	/* nr_channels */
+   TRUE,	/* is_array */
+   TRUE,	/* is_bitmask */
+   FALSE,	/* is_mixed */
+   {
+      {UTIL_FORMAT_TYPE_SIGNED, TRUE, 8},	/* x = rgba */
+      {0, 0, 0},
+      {0, 0, 0},
+      {0, 0, 0}
+   },
+   {
+      UTIL_FORMAT_SWIZZLE_X,	/* r */
+      UTIL_FORMAT_SWIZZLE_X,	/* g */
+      UTIL_FORMAT_SWIZZLE_X,	/* b */
+      UTIL_FORMAT_SWIZZLE_X	/* a */
+   },
+   UTIL_FORMAT_COLORSPACE_RGB,
+   &util_format_i8_snorm_unpack_rgba_8unorm,
+   &util_format_i8_snorm_pack_rgba_8unorm,
+   NULL, /* fetch_rgba_8unorm */
+   &util_format_i8_snorm_unpack_rgba_float,
+   &util_format_i8_snorm_pack_rgba_float,
+   &util_format_i8_snorm_fetch_rgba_float,
+   NULL, /* unpack_z_32unorm */
+   NULL, /* pack_z_32unorm */
+   NULL, /* unpack_z_float */
+   NULL, /* pack_z_float */
+   NULL, /* unpack_s_8uscaled */
+   NULL /* pack_s_8uscaled */
+};
+
+const struct util_format_description
+util_format_a16_snorm_description = {
+   PIPE_FORMAT_A16_SNORM,
+   "PIPE_FORMAT_A16_SNORM",
+   "a16_snorm",
+   {1, 1, 16},	/* block */
+   UTIL_FORMAT_LAYOUT_PLAIN,
+   1,	/* nr_channels */
+   TRUE,	/* is_array */
+   TRUE,	/* is_bitmask */
+   FALSE,	/* is_mixed */
+   {
+      {UTIL_FORMAT_TYPE_SIGNED, TRUE, 16},	/* x = a */
+      {0, 0, 0},
+      {0, 0, 0},
+      {0, 0, 0}
+   },
+   {
+      UTIL_FORMAT_SWIZZLE_0,	/* r */
+      UTIL_FORMAT_SWIZZLE_0,	/* g */
+      UTIL_FORMAT_SWIZZLE_0,	/* b */
+      UTIL_FORMAT_SWIZZLE_X	/* a */
+   },
+   UTIL_FORMAT_COLORSPACE_RGB,
+   &util_format_a16_snorm_unpack_rgba_8unorm,
+   &util_format_a16_snorm_pack_rgba_8unorm,
+   NULL, /* fetch_rgba_8unorm */
+   &util_format_a16_snorm_unpack_rgba_float,
+   &util_format_a16_snorm_pack_rgba_float,
+   &util_format_a16_snorm_fetch_rgba_float,
+   NULL, /* unpack_z_32unorm */
+   NULL, /* pack_z_32unorm */
+   NULL, /* unpack_z_float */
+   NULL, /* pack_z_float */
+   NULL, /* unpack_s_8uscaled */
+   NULL /* pack_s_8uscaled */
+};
+
+const struct util_format_description
+util_format_l16_snorm_description = {
+   PIPE_FORMAT_L16_SNORM,
+   "PIPE_FORMAT_L16_SNORM",
+   "l16_snorm",
+   {1, 1, 16},	/* block */
+   UTIL_FORMAT_LAYOUT_PLAIN,
+   1,	/* nr_channels */
+   TRUE,	/* is_array */
+   TRUE,	/* is_bitmask */
+   FALSE,	/* is_mixed */
+   {
+      {UTIL_FORMAT_TYPE_SIGNED, TRUE, 16},	/* x = rgb */
+      {0, 0, 0},
+      {0, 0, 0},
+      {0, 0, 0}
+   },
+   {
+      UTIL_FORMAT_SWIZZLE_X,	/* r */
+      UTIL_FORMAT_SWIZZLE_X,	/* g */
+      UTIL_FORMAT_SWIZZLE_X,	/* b */
+      UTIL_FORMAT_SWIZZLE_1	/* a */
+   },
+   UTIL_FORMAT_COLORSPACE_RGB,
+   &util_format_l16_snorm_unpack_rgba_8unorm,
+   &util_format_l16_snorm_pack_rgba_8unorm,
+   NULL, /* fetch_rgba_8unorm */
+   &util_format_l16_snorm_unpack_rgba_float,
+   &util_format_l16_snorm_pack_rgba_float,
+   &util_format_l16_snorm_fetch_rgba_float,
+   NULL, /* unpack_z_32unorm */
+   NULL, /* pack_z_32unorm */
+   NULL, /* unpack_z_float */
+   NULL, /* pack_z_float */
+   NULL, /* unpack_s_8uscaled */
+   NULL /* pack_s_8uscaled */
+};
+
+const struct util_format_description
+util_format_l16a16_snorm_description = {
+   PIPE_FORMAT_L16A16_SNORM,
+   "PIPE_FORMAT_L16A16_SNORM",
+   "l16a16_snorm",
+   {1, 1, 32},	/* block */
+   UTIL_FORMAT_LAYOUT_PLAIN,
+   2,	/* nr_channels */
+   TRUE,	/* is_array */
+   TRUE,	/* is_bitmask */
+   FALSE,	/* is_mixed */
+   {
+      {UTIL_FORMAT_TYPE_SIGNED, TRUE, 16},	/* x = rgb */
+      {UTIL_FORMAT_TYPE_SIGNED, TRUE, 16},	/* y = a */
+      {0, 0, 0},
+      {0, 0, 0}
+   },
+   {
+      UTIL_FORMAT_SWIZZLE_X,	/* r */
+      UTIL_FORMAT_SWIZZLE_X,	/* g */
+      UTIL_FORMAT_SWIZZLE_X,	/* b */
+      UTIL_FORMAT_SWIZZLE_Y	/* a */
+   },
+   UTIL_FORMAT_COLORSPACE_RGB,
+   &util_format_l16a16_snorm_unpack_rgba_8unorm,
+   &util_format_l16a16_snorm_pack_rgba_8unorm,
+   NULL, /* fetch_rgba_8unorm */
+   &util_format_l16a16_snorm_unpack_rgba_float,
+   &util_format_l16a16_snorm_pack_rgba_float,
+   &util_format_l16a16_snorm_fetch_rgba_float,
+   NULL, /* unpack_z_32unorm */
+   NULL, /* pack_z_32unorm */
+   NULL, /* unpack_z_float */
+   NULL, /* pack_z_float */
+   NULL, /* unpack_s_8uscaled */
+   NULL /* pack_s_8uscaled */
+};
+
+const struct util_format_description
+util_format_i16_snorm_description = {
+   PIPE_FORMAT_I16_SNORM,
+   "PIPE_FORMAT_I16_SNORM",
+   "i16_snorm",
+   {1, 1, 16},	/* block */
+   UTIL_FORMAT_LAYOUT_PLAIN,
+   1,	/* nr_channels */
+   TRUE,	/* is_array */
+   TRUE,	/* is_bitmask */
+   FALSE,	/* is_mixed */
+   {
+      {UTIL_FORMAT_TYPE_SIGNED, TRUE, 16},	/* x = rgba */
+      {0, 0, 0},
+      {0, 0, 0},
+      {0, 0, 0}
+   },
+   {
+      UTIL_FORMAT_SWIZZLE_X,	/* r */
+      UTIL_FORMAT_SWIZZLE_X,	/* g */
+      UTIL_FORMAT_SWIZZLE_X,	/* b */
+      UTIL_FORMAT_SWIZZLE_X	/* a */
+   },
+   UTIL_FORMAT_COLORSPACE_RGB,
+   &util_format_i16_snorm_unpack_rgba_8unorm,
+   &util_format_i16_snorm_pack_rgba_8unorm,
+   NULL, /* fetch_rgba_8unorm */
+   &util_format_i16_snorm_unpack_rgba_float,
+   &util_format_i16_snorm_pack_rgba_float,
+   &util_format_i16_snorm_fetch_rgba_float,
+   NULL, /* unpack_z_32unorm */
+   NULL, /* pack_z_32unorm */
+   NULL, /* unpack_z_float */
+   NULL, /* pack_z_float */
+   NULL, /* unpack_s_8uscaled */
+   NULL /* pack_s_8uscaled */
+};
+
+const struct util_format_description
+util_format_a16_float_description = {
+   PIPE_FORMAT_A16_FLOAT,
+   "PIPE_FORMAT_A16_FLOAT",
+   "a16_float",
+   {1, 1, 16},	/* block */
+   UTIL_FORMAT_LAYOUT_PLAIN,
+   1,	/* nr_channels */
+   TRUE,	/* is_array */
+   FALSE,	/* is_bitmask */
+   FALSE,	/* is_mixed */
+   {
+      {UTIL_FORMAT_TYPE_FLOAT, FALSE, 16},	/* x = a */
+      {0, 0, 0},
+      {0, 0, 0},
+      {0, 0, 0}
+   },
+   {
+      UTIL_FORMAT_SWIZZLE_0,	/* r */
+      UTIL_FORMAT_SWIZZLE_0,	/* g */
+      UTIL_FORMAT_SWIZZLE_0,	/* b */
+      UTIL_FORMAT_SWIZZLE_X	/* a */
+   },
+   UTIL_FORMAT_COLORSPACE_RGB,
+   &util_format_a16_float_unpack_rgba_8unorm,
+   &util_format_a16_float_pack_rgba_8unorm,
+   NULL, /* fetch_rgba_8unorm */
+   &util_format_a16_float_unpack_rgba_float,
+   &util_format_a16_float_pack_rgba_float,
+   &util_format_a16_float_fetch_rgba_float,
+   NULL, /* unpack_z_32unorm */
+   NULL, /* pack_z_32unorm */
+   NULL, /* unpack_z_float */
+   NULL, /* pack_z_float */
+   NULL, /* unpack_s_8uscaled */
+   NULL /* pack_s_8uscaled */
+};
+
+const struct util_format_description
+util_format_l16_float_description = {
+   PIPE_FORMAT_L16_FLOAT,
+   "PIPE_FORMAT_L16_FLOAT",
+   "l16_float",
+   {1, 1, 16},	/* block */
+   UTIL_FORMAT_LAYOUT_PLAIN,
+   1,	/* nr_channels */
+   TRUE,	/* is_array */
+   FALSE,	/* is_bitmask */
+   FALSE,	/* is_mixed */
+   {
+      {UTIL_FORMAT_TYPE_FLOAT, FALSE, 16},	/* x = rgb */
+      {0, 0, 0},
+      {0, 0, 0},
+      {0, 0, 0}
+   },
+   {
+      UTIL_FORMAT_SWIZZLE_X,	/* r */
+      UTIL_FORMAT_SWIZZLE_X,	/* g */
+      UTIL_FORMAT_SWIZZLE_X,	/* b */
+      UTIL_FORMAT_SWIZZLE_1	/* a */
+   },
+   UTIL_FORMAT_COLORSPACE_RGB,
+   &util_format_l16_float_unpack_rgba_8unorm,
+   &util_format_l16_float_pack_rgba_8unorm,
+   NULL, /* fetch_rgba_8unorm */
+   &util_format_l16_float_unpack_rgba_float,
+   &util_format_l16_float_pack_rgba_float,
+   &util_format_l16_float_fetch_rgba_float,
+   NULL, /* unpack_z_32unorm */
+   NULL, /* pack_z_32unorm */
+   NULL, /* unpack_z_float */
+   NULL, /* pack_z_float */
+   NULL, /* unpack_s_8uscaled */
+   NULL /* pack_s_8uscaled */
+};
+
+const struct util_format_description
+util_format_l16a16_float_description = {
+   PIPE_FORMAT_L16A16_FLOAT,
+   "PIPE_FORMAT_L16A16_FLOAT",
+   "l16a16_float",
+   {1, 1, 32},	/* block */
+   UTIL_FORMAT_LAYOUT_PLAIN,
+   2,	/* nr_channels */
+   TRUE,	/* is_array */
+   FALSE,	/* is_bitmask */
+   FALSE,	/* is_mixed */
+   {
+      {UTIL_FORMAT_TYPE_FLOAT, FALSE, 16},	/* x = rgb */
+      {UTIL_FORMAT_TYPE_FLOAT, FALSE, 16},	/* y = a */
+      {0, 0, 0},
+      {0, 0, 0}
+   },
+   {
+      UTIL_FORMAT_SWIZZLE_X,	/* r */
+      UTIL_FORMAT_SWIZZLE_X,	/* g */
+      UTIL_FORMAT_SWIZZLE_X,	/* b */
+      UTIL_FORMAT_SWIZZLE_Y	/* a */
+   },
+   UTIL_FORMAT_COLORSPACE_RGB,
+   &util_format_l16a16_float_unpack_rgba_8unorm,
+   &util_format_l16a16_float_pack_rgba_8unorm,
+   NULL, /* fetch_rgba_8unorm */
+   &util_format_l16a16_float_unpack_rgba_float,
+   &util_format_l16a16_float_pack_rgba_float,
+   &util_format_l16a16_float_fetch_rgba_float,
+   NULL, /* unpack_z_32unorm */
+   NULL, /* pack_z_32unorm */
+   NULL, /* unpack_z_float */
+   NULL, /* pack_z_float */
+   NULL, /* unpack_s_8uscaled */
+   NULL /* pack_s_8uscaled */
+};
+
+const struct util_format_description
+util_format_i16_float_description = {
+   PIPE_FORMAT_I16_FLOAT,
+   "PIPE_FORMAT_I16_FLOAT",
+   "i16_float",
+   {1, 1, 16},	/* block */
+   UTIL_FORMAT_LAYOUT_PLAIN,
+   1,	/* nr_channels */
+   TRUE,	/* is_array */
+   FALSE,	/* is_bitmask */
+   FALSE,	/* is_mixed */
+   {
+      {UTIL_FORMAT_TYPE_FLOAT, FALSE, 16},	/* x = rgba */
+      {0, 0, 0},
+      {0, 0, 0},
+      {0, 0, 0}
+   },
+   {
+      UTIL_FORMAT_SWIZZLE_X,	/* r */
+      UTIL_FORMAT_SWIZZLE_X,	/* g */
+      UTIL_FORMAT_SWIZZLE_X,	/* b */
+      UTIL_FORMAT_SWIZZLE_X	/* a */
+   },
+   UTIL_FORMAT_COLORSPACE_RGB,
+   &util_format_i16_float_unpack_rgba_8unorm,
+   &util_format_i16_float_pack_rgba_8unorm,
+   NULL, /* fetch_rgba_8unorm */
+   &util_format_i16_float_unpack_rgba_float,
+   &util_format_i16_float_pack_rgba_float,
+   &util_format_i16_float_fetch_rgba_float,
+   NULL, /* unpack_z_32unorm */
+   NULL, /* pack_z_32unorm */
+   NULL, /* unpack_z_float */
+   NULL, /* pack_z_float */
+   NULL, /* unpack_s_8uscaled */
+   NULL /* pack_s_8uscaled */
+};
+
+const struct util_format_description
+util_format_a32_float_description = {
+   PIPE_FORMAT_A32_FLOAT,
+   "PIPE_FORMAT_A32_FLOAT",
+   "a32_float",
+   {1, 1, 32},	/* block */
+   UTIL_FORMAT_LAYOUT_PLAIN,
+   1,	/* nr_channels */
+   TRUE,	/* is_array */
+   FALSE,	/* is_bitmask */
+   FALSE,	/* is_mixed */
+   {
+      {UTIL_FORMAT_TYPE_FLOAT, FALSE, 32},	/* x = a */
+      {0, 0, 0},
+      {0, 0, 0},
+      {0, 0, 0}
+   },
+   {
+      UTIL_FORMAT_SWIZZLE_0,	/* r */
+      UTIL_FORMAT_SWIZZLE_0,	/* g */
+      UTIL_FORMAT_SWIZZLE_0,	/* b */
+      UTIL_FORMAT_SWIZZLE_X	/* a */
+   },
+   UTIL_FORMAT_COLORSPACE_RGB,
+   &util_format_a32_float_unpack_rgba_8unorm,
+   &util_format_a32_float_pack_rgba_8unorm,
+   NULL, /* fetch_rgba_8unorm */
+   &util_format_a32_float_unpack_rgba_float,
+   &util_format_a32_float_pack_rgba_float,
+   &util_format_a32_float_fetch_rgba_float,
+   NULL, /* unpack_z_32unorm */
+   NULL, /* pack_z_32unorm */
+   NULL, /* unpack_z_float */
+   NULL, /* pack_z_float */
+   NULL, /* unpack_s_8uscaled */
+   NULL /* pack_s_8uscaled */
+};
+
+const struct util_format_description
+util_format_l32_float_description = {
+   PIPE_FORMAT_L32_FLOAT,
+   "PIPE_FORMAT_L32_FLOAT",
+   "l32_float",
+   {1, 1, 32},	/* block */
+   UTIL_FORMAT_LAYOUT_PLAIN,
+   1,	/* nr_channels */
+   TRUE,	/* is_array */
+   FALSE,	/* is_bitmask */
+   FALSE,	/* is_mixed */
+   {
+      {UTIL_FORMAT_TYPE_FLOAT, FALSE, 32},	/* x = rgb */
+      {0, 0, 0},
+      {0, 0, 0},
+      {0, 0, 0}
+   },
+   {
+      UTIL_FORMAT_SWIZZLE_X,	/* r */
+      UTIL_FORMAT_SWIZZLE_X,	/* g */
+      UTIL_FORMAT_SWIZZLE_X,	/* b */
+      UTIL_FORMAT_SWIZZLE_1	/* a */
+   },
+   UTIL_FORMAT_COLORSPACE_RGB,
+   &util_format_l32_float_unpack_rgba_8unorm,
+   &util_format_l32_float_pack_rgba_8unorm,
+   NULL, /* fetch_rgba_8unorm */
+   &util_format_l32_float_unpack_rgba_float,
+   &util_format_l32_float_pack_rgba_float,
+   &util_format_l32_float_fetch_rgba_float,
+   NULL, /* unpack_z_32unorm */
+   NULL, /* pack_z_32unorm */
+   NULL, /* unpack_z_float */
+   NULL, /* pack_z_float */
+   NULL, /* unpack_s_8uscaled */
+   NULL /* pack_s_8uscaled */
+};
+
+const struct util_format_description
+util_format_l32a32_float_description = {
+   PIPE_FORMAT_L32A32_FLOAT,
+   "PIPE_FORMAT_L32A32_FLOAT",
+   "l32a32_float",
+   {1, 1, 64},	/* block */
+   UTIL_FORMAT_LAYOUT_PLAIN,
+   2,	/* nr_channels */
+   TRUE,	/* is_array */
+   FALSE,	/* is_bitmask */
+   FALSE,	/* is_mixed */
+   {
+      {UTIL_FORMAT_TYPE_FLOAT, FALSE, 32},	/* x = rgb */
+      {UTIL_FORMAT_TYPE_FLOAT, FALSE, 32},	/* y = a */
+      {0, 0, 0},
+      {0, 0, 0}
+   },
+   {
+      UTIL_FORMAT_SWIZZLE_X,	/* r */
+      UTIL_FORMAT_SWIZZLE_X,	/* g */
+      UTIL_FORMAT_SWIZZLE_X,	/* b */
+      UTIL_FORMAT_SWIZZLE_Y	/* a */
+   },
+   UTIL_FORMAT_COLORSPACE_RGB,
+   &util_format_l32a32_float_unpack_rgba_8unorm,
+   &util_format_l32a32_float_pack_rgba_8unorm,
+   NULL, /* fetch_rgba_8unorm */
+   &util_format_l32a32_float_unpack_rgba_float,
+   &util_format_l32a32_float_pack_rgba_float,
+   &util_format_l32a32_float_fetch_rgba_float,
+   NULL, /* unpack_z_32unorm */
+   NULL, /* pack_z_32unorm */
+   NULL, /* unpack_z_float */
+   NULL, /* pack_z_float */
+   NULL, /* unpack_s_8uscaled */
+   NULL /* pack_s_8uscaled */
+};
+
+const struct util_format_description
+util_format_i32_float_description = {
+   PIPE_FORMAT_I32_FLOAT,
+   "PIPE_FORMAT_I32_FLOAT",
+   "i32_float",
+   {1, 1, 32},	/* block */
+   UTIL_FORMAT_LAYOUT_PLAIN,
+   1,	/* nr_channels */
+   TRUE,	/* is_array */
+   FALSE,	/* is_bitmask */
+   FALSE,	/* is_mixed */
+   {
+      {UTIL_FORMAT_TYPE_FLOAT, FALSE, 32},	/* x = rgba */
+      {0, 0, 0},
+      {0, 0, 0},
+      {0, 0, 0}
+   },
+   {
+      UTIL_FORMAT_SWIZZLE_X,	/* r */
+      UTIL_FORMAT_SWIZZLE_X,	/* g */
+      UTIL_FORMAT_SWIZZLE_X,	/* b */
+      UTIL_FORMAT_SWIZZLE_X	/* a */
+   },
+   UTIL_FORMAT_COLORSPACE_RGB,
+   &util_format_i32_float_unpack_rgba_8unorm,
+   &util_format_i32_float_pack_rgba_8unorm,
+   NULL, /* fetch_rgba_8unorm */
+   &util_format_i32_float_unpack_rgba_float,
+   &util_format_i32_float_pack_rgba_float,
+   &util_format_i32_float_fetch_rgba_float,
+   NULL, /* unpack_z_32unorm */
+   NULL, /* pack_z_32unorm */
+   NULL, /* unpack_z_float */
+   NULL, /* pack_z_float */
+   NULL, /* unpack_s_8uscaled */
+   NULL /* pack_s_8uscaled */
+};
+
+const struct util_format_description
 util_format_l8_srgb_description = {
    PIPE_FORMAT_L8_SRGB,
    "PIPE_FORMAT_L8_SRGB",
@@ -14170,15 +17014,15 @@ util_format_r11g11b10_float_description = {
    "PIPE_FORMAT_R11G11B10_FLOAT",
    "r11g11b10_float",
    {1, 1, 32},	/* block */
-   UTIL_FORMAT_LAYOUT_PLAIN,
-   3,	/* nr_channels */
+   UTIL_FORMAT_LAYOUT_OTHER,
+   1,	/* nr_channels */
    FALSE,	/* is_array */
    FALSE,	/* is_bitmask */
    FALSE,	/* is_mixed */
    {
-      {UTIL_FORMAT_TYPE_FLOAT, FALSE, 11},	/* x = r */
-      {UTIL_FORMAT_TYPE_FLOAT, FALSE, 11},	/* y = g */
-      {UTIL_FORMAT_TYPE_FLOAT, FALSE, 10},	/* z = b */
+      {UTIL_FORMAT_TYPE_VOID, FALSE, 32},	/* x = x */
+      {0, 0, 0},
+      {0, 0, 0},
       {0, 0, 0}
    },
    {
@@ -14646,7 +17490,7 @@ util_format_rgtc1_unorm_description = {
    UTIL_FORMAT_COLORSPACE_RGB,
    &util_format_rgtc1_unorm_unpack_rgba_8unorm,
    &util_format_rgtc1_unorm_pack_rgba_8unorm,
-   NULL, /* fetch_rgba_8unorm */
+   &util_format_rgtc1_unorm_fetch_rgba_8unorm,
    &util_format_rgtc1_unorm_unpack_rgba_float,
    &util_format_rgtc1_unorm_pack_rgba_float,
    &util_format_rgtc1_unorm_fetch_rgba_float,
@@ -14684,7 +17528,7 @@ util_format_rgtc1_snorm_description = {
    UTIL_FORMAT_COLORSPACE_RGB,
    &util_format_rgtc1_snorm_unpack_rgba_8unorm,
    &util_format_rgtc1_snorm_pack_rgba_8unorm,
-   NULL, /* fetch_rgba_8unorm */
+   &util_format_rgtc1_snorm_fetch_rgba_8unorm,
    &util_format_rgtc1_snorm_unpack_rgba_float,
    &util_format_rgtc1_snorm_pack_rgba_float,
    &util_format_rgtc1_snorm_fetch_rgba_float,
@@ -14722,7 +17566,7 @@ util_format_rgtc2_unorm_description = {
    UTIL_FORMAT_COLORSPACE_RGB,
    &util_format_rgtc2_unorm_unpack_rgba_8unorm,
    &util_format_rgtc2_unorm_pack_rgba_8unorm,
-   NULL, /* fetch_rgba_8unorm */
+   &util_format_rgtc2_unorm_fetch_rgba_8unorm,
    &util_format_rgtc2_unorm_unpack_rgba_float,
    &util_format_rgtc2_unorm_pack_rgba_float,
    &util_format_rgtc2_unorm_fetch_rgba_float,
@@ -14760,7 +17604,7 @@ util_format_rgtc2_snorm_description = {
    UTIL_FORMAT_COLORSPACE_RGB,
    &util_format_rgtc2_snorm_unpack_rgba_8unorm,
    &util_format_rgtc2_snorm_pack_rgba_8unorm,
-   NULL, /* fetch_rgba_8unorm */
+   &util_format_rgtc2_snorm_fetch_rgba_8unorm,
    &util_format_rgtc2_snorm_unpack_rgba_float,
    &util_format_rgtc2_snorm_pack_rgba_float,
    &util_format_rgtc2_snorm_fetch_rgba_float,
@@ -14773,6 +17617,158 @@ util_format_rgtc2_snorm_description = {
 };
 
 const struct util_format_description
+util_format_latc1_unorm_description = {
+   PIPE_FORMAT_LATC1_UNORM,
+   "PIPE_FORMAT_LATC1_UNORM",
+   "latc1_unorm",
+   {4, 4, 64},	/* block */
+   UTIL_FORMAT_LAYOUT_RGTC,
+   1,	/* nr_channels */
+   FALSE,	/* is_array */
+   FALSE,	/* is_bitmask */
+   FALSE,	/* is_mixed */
+   {
+      {UTIL_FORMAT_TYPE_VOID, FALSE, 64},	/* x = x */
+      {0, 0, 0},
+      {0, 0, 0},
+      {0, 0, 0}
+   },
+   {
+      UTIL_FORMAT_SWIZZLE_X,	/* r */
+      UTIL_FORMAT_SWIZZLE_X,	/* g */
+      UTIL_FORMAT_SWIZZLE_X,	/* b */
+      UTIL_FORMAT_SWIZZLE_1	/* a */
+   },
+   UTIL_FORMAT_COLORSPACE_RGB,
+   &util_format_latc1_unorm_unpack_rgba_8unorm,
+   &util_format_latc1_unorm_pack_rgba_8unorm,
+   &util_format_latc1_unorm_fetch_rgba_8unorm,
+   &util_format_latc1_unorm_unpack_rgba_float,
+   &util_format_latc1_unorm_pack_rgba_float,
+   &util_format_latc1_unorm_fetch_rgba_float,
+   NULL, /* unpack_z_32unorm */
+   NULL, /* pack_z_32unorm */
+   NULL, /* unpack_z_float */
+   NULL, /* pack_z_float */
+   NULL, /* unpack_s_8uscaled */
+   NULL /* pack_s_8uscaled */
+};
+
+const struct util_format_description
+util_format_latc1_snorm_description = {
+   PIPE_FORMAT_LATC1_SNORM,
+   "PIPE_FORMAT_LATC1_SNORM",
+   "latc1_snorm",
+   {4, 4, 64},	/* block */
+   UTIL_FORMAT_LAYOUT_RGTC,
+   1,	/* nr_channels */
+   FALSE,	/* is_array */
+   FALSE,	/* is_bitmask */
+   FALSE,	/* is_mixed */
+   {
+      {UTIL_FORMAT_TYPE_VOID, FALSE, 64},	/* x = x */
+      {0, 0, 0},
+      {0, 0, 0},
+      {0, 0, 0}
+   },
+   {
+      UTIL_FORMAT_SWIZZLE_X,	/* r */
+      UTIL_FORMAT_SWIZZLE_X,	/* g */
+      UTIL_FORMAT_SWIZZLE_X,	/* b */
+      UTIL_FORMAT_SWIZZLE_1	/* a */
+   },
+   UTIL_FORMAT_COLORSPACE_RGB,
+   &util_format_latc1_snorm_unpack_rgba_8unorm,
+   &util_format_latc1_snorm_pack_rgba_8unorm,
+   &util_format_latc1_snorm_fetch_rgba_8unorm,
+   &util_format_latc1_snorm_unpack_rgba_float,
+   &util_format_latc1_snorm_pack_rgba_float,
+   &util_format_latc1_snorm_fetch_rgba_float,
+   NULL, /* unpack_z_32unorm */
+   NULL, /* pack_z_32unorm */
+   NULL, /* unpack_z_float */
+   NULL, /* pack_z_float */
+   NULL, /* unpack_s_8uscaled */
+   NULL /* pack_s_8uscaled */
+};
+
+const struct util_format_description
+util_format_latc2_unorm_description = {
+   PIPE_FORMAT_LATC2_UNORM,
+   "PIPE_FORMAT_LATC2_UNORM",
+   "latc2_unorm",
+   {4, 4, 128},	/* block */
+   UTIL_FORMAT_LAYOUT_RGTC,
+   1,	/* nr_channels */
+   FALSE,	/* is_array */
+   FALSE,	/* is_bitmask */
+   FALSE,	/* is_mixed */
+   {
+      {UTIL_FORMAT_TYPE_VOID, FALSE, 128},	/* x = x */
+      {0, 0, 0},
+      {0, 0, 0},
+      {0, 0, 0}
+   },
+   {
+      UTIL_FORMAT_SWIZZLE_X,	/* r */
+      UTIL_FORMAT_SWIZZLE_X,	/* g */
+      UTIL_FORMAT_SWIZZLE_X,	/* b */
+      UTIL_FORMAT_SWIZZLE_Y	/* a */
+   },
+   UTIL_FORMAT_COLORSPACE_RGB,
+   &util_format_latc2_unorm_unpack_rgba_8unorm,
+   &util_format_latc2_unorm_pack_rgba_8unorm,
+   &util_format_latc2_unorm_fetch_rgba_8unorm,
+   &util_format_latc2_unorm_unpack_rgba_float,
+   &util_format_latc2_unorm_pack_rgba_float,
+   &util_format_latc2_unorm_fetch_rgba_float,
+   NULL, /* unpack_z_32unorm */
+   NULL, /* pack_z_32unorm */
+   NULL, /* unpack_z_float */
+   NULL, /* pack_z_float */
+   NULL, /* unpack_s_8uscaled */
+   NULL /* pack_s_8uscaled */
+};
+
+const struct util_format_description
+util_format_latc2_snorm_description = {
+   PIPE_FORMAT_LATC2_SNORM,
+   "PIPE_FORMAT_LATC2_SNORM",
+   "latc2_snorm",
+   {4, 4, 128},	/* block */
+   UTIL_FORMAT_LAYOUT_RGTC,
+   1,	/* nr_channels */
+   FALSE,	/* is_array */
+   FALSE,	/* is_bitmask */
+   FALSE,	/* is_mixed */
+   {
+      {UTIL_FORMAT_TYPE_VOID, FALSE, 128},	/* x = x */
+      {0, 0, 0},
+      {0, 0, 0},
+      {0, 0, 0}
+   },
+   {
+      UTIL_FORMAT_SWIZZLE_X,	/* r */
+      UTIL_FORMAT_SWIZZLE_X,	/* g */
+      UTIL_FORMAT_SWIZZLE_X,	/* b */
+      UTIL_FORMAT_SWIZZLE_Y	/* a */
+   },
+   UTIL_FORMAT_COLORSPACE_RGB,
+   &util_format_latc2_snorm_unpack_rgba_8unorm,
+   &util_format_latc2_snorm_pack_rgba_8unorm,
+   &util_format_latc2_snorm_fetch_rgba_8unorm,
+   &util_format_latc2_snorm_unpack_rgba_float,
+   &util_format_latc2_snorm_pack_rgba_float,
+   &util_format_latc2_snorm_fetch_rgba_float,
+   NULL, /* unpack_z_32unorm */
+   NULL, /* pack_z_32unorm */
+   NULL, /* unpack_z_float */
+   NULL, /* pack_z_float */
+   NULL, /* unpack_s_8uscaled */
+   NULL /* pack_s_8uscaled */
+};
+
+const struct util_format_description
 util_format_r64_float_description = {
    PIPE_FORMAT_R64_FLOAT,
    "PIPE_FORMAT_R64_FLOAT",
@@ -17318,6 +20314,8 @@ util_format_description(enum pipe_format format)
       return &util_format_r10g10b10a2_unorm_description;
    case PIPE_FORMAT_B10G10R10A2_UNORM:
       return &util_format_b10g10r10a2_unorm_description;
+   case PIPE_FORMAT_B2G3R3_UNORM:
+      return &util_format_b2g3r3_unorm_description;
    case PIPE_FORMAT_L8_UNORM:
       return &util_format_l8_unorm_description;
    case PIPE_FORMAT_A8_UNORM:
@@ -17330,6 +20328,44 @@ util_format_description(enum pipe_format format)
       return &util_format_l8a8_unorm_description;
    case PIPE_FORMAT_L16_UNORM:
       return &util_format_l16_unorm_description;
+   case PIPE_FORMAT_A16_UNORM:
+      return &util_format_a16_unorm_description;
+   case PIPE_FORMAT_I16_UNORM:
+      return &util_format_i16_unorm_description;
+   case PIPE_FORMAT_L16A16_UNORM:
+      return &util_format_l16a16_unorm_description;
+   case PIPE_FORMAT_A8_SNORM:
+      return &util_format_a8_snorm_description;
+   case PIPE_FORMAT_L8_SNORM:
+      return &util_format_l8_snorm_description;
+   case PIPE_FORMAT_L8A8_SNORM:
+      return &util_format_l8a8_snorm_description;
+   case PIPE_FORMAT_I8_SNORM:
+      return &util_format_i8_snorm_description;
+   case PIPE_FORMAT_A16_SNORM:
+      return &util_format_a16_snorm_description;
+   case PIPE_FORMAT_L16_SNORM:
+      return &util_format_l16_snorm_description;
+   case PIPE_FORMAT_L16A16_SNORM:
+      return &util_format_l16a16_snorm_description;
+   case PIPE_FORMAT_I16_SNORM:
+      return &util_format_i16_snorm_description;
+   case PIPE_FORMAT_A16_FLOAT:
+      return &util_format_a16_float_description;
+   case PIPE_FORMAT_L16_FLOAT:
+      return &util_format_l16_float_description;
+   case PIPE_FORMAT_L16A16_FLOAT:
+      return &util_format_l16a16_float_description;
+   case PIPE_FORMAT_I16_FLOAT:
+      return &util_format_i16_float_description;
+   case PIPE_FORMAT_A32_FLOAT:
+      return &util_format_a32_float_description;
+   case PIPE_FORMAT_L32_FLOAT:
+      return &util_format_l32_float_description;
+   case PIPE_FORMAT_L32A32_FLOAT:
+      return &util_format_l32a32_float_description;
+   case PIPE_FORMAT_I32_FLOAT:
+      return &util_format_i32_float_description;
    case PIPE_FORMAT_L8_SRGB:
       return &util_format_l8_srgb_description;
    case PIPE_FORMAT_L8A8_SRGB:
@@ -17422,6 +20458,14 @@ util_format_description(enum pipe_format format)
       return &util_format_rgtc2_unorm_description;
    case PIPE_FORMAT_RGTC2_SNORM:
       return &util_format_rgtc2_snorm_description;
+   case PIPE_FORMAT_LATC1_UNORM:
+      return &util_format_latc1_unorm_description;
+   case PIPE_FORMAT_LATC1_SNORM:
+      return &util_format_latc1_snorm_description;
+   case PIPE_FORMAT_LATC2_UNORM:
+      return &util_format_latc2_unorm_description;
+   case PIPE_FORMAT_LATC2_SNORM:
+      return &util_format_latc2_snorm_description;
    case PIPE_FORMAT_R64_FLOAT:
       return &util_format_r64_float_description;
    case PIPE_FORMAT_R64G64_FLOAT:
diff --git a/lib/libGL/gallium/libgallium/generated/u_unfilled_gen.c b/lib/libGL/gallium/libgallium/generated/u_unfilled_gen.c
index 93897c98d..e71cd2feb 100644
--- a/lib/libGL/gallium/libgallium/generated/u_unfilled_gen.c
+++ b/lib/libGL/gallium/libgallium/generated/u_unfilled_gen.c
@@ -154,13 +154,9 @@ static void generate_polygon_ushort(
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
-  for (j = i = 0; j < nr; j+=6, i++) { 
-      (out+j)[0] = (ushort)(0);
-      (out+j)[1] = (ushort)(i+1);
-      (out+j+2)[0] = (ushort)(i+1);
-      (out+j+2)[1] = (ushort)(i+2);
-      (out+j+4)[0] = (ushort)(i+2);
-      (out+j+4)[1] = (ushort)(0);
+  for (j = i = 0; j < nr; j+=2, i++) { 
+      (out+j)[0] = (ushort)(i);
+      (out+j)[1] = (ushort)((i+1)%(nr/2));
    }
 }
 static void generate_tris_uint(
@@ -254,13 +250,9 @@ static void generate_polygon_uint(
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
-  for (j = i = 0; j < nr; j+=6, i++) { 
-      (out+j)[0] = (uint)(0);
-      (out+j)[1] = (uint)(i+1);
-      (out+j+2)[0] = (uint)(i+1);
-      (out+j+2)[1] = (uint)(i+2);
-      (out+j+4)[0] = (uint)(i+2);
-      (out+j+4)[1] = (uint)(0);
+  for (j = i = 0; j < nr; j+=2, i++) { 
+      (out+j)[0] = (uint)(i);
+      (out+j)[1] = (uint)((i+1)%(nr/2));
    }
 }
 static void translate_tris_ubyte2ushort(
@@ -366,13 +358,9 @@ static void translate_polygon_ubyte2ushort(
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
-  for (j = i = 0; j < nr; j+=6, i++) { 
-      (out+j)[0] = (ushort)in[0];
-      (out+j)[1] = (ushort)in[i+1];
-      (out+j+2)[0] = (ushort)in[i+1];
-      (out+j+2)[1] = (ushort)in[i+2];
-      (out+j+4)[0] = (ushort)in[i+2];
-      (out+j+4)[1] = (ushort)in[0];
+  for (j = i = 0; j < nr; j+=2, i++) { 
+      (out+j)[0] = (ushort)in[i];
+      (out+j)[1] = (ushort)in[(i+1)%(nr/2)];
    }
 }
 static void translate_tris_ubyte2uint(
@@ -478,13 +466,9 @@ static void translate_polygon_ubyte2uint(
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
-  for (j = i = 0; j < nr; j+=6, i++) { 
-      (out+j)[0] = (uint)in[0];
-      (out+j)[1] = (uint)in[i+1];
-      (out+j+2)[0] = (uint)in[i+1];
-      (out+j+2)[1] = (uint)in[i+2];
-      (out+j+4)[0] = (uint)in[i+2];
-      (out+j+4)[1] = (uint)in[0];
+  for (j = i = 0; j < nr; j+=2, i++) { 
+      (out+j)[0] = (uint)in[i];
+      (out+j)[1] = (uint)in[(i+1)%(nr/2)];
    }
 }
 static void translate_tris_ushort2ushort(
@@ -590,13 +574,9 @@ static void translate_polygon_ushort2ushort(
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
-  for (j = i = 0; j < nr; j+=6, i++) { 
-      (out+j)[0] = (ushort)in[0];
-      (out+j)[1] = (ushort)in[i+1];
-      (out+j+2)[0] = (ushort)in[i+1];
-      (out+j+2)[1] = (ushort)in[i+2];
-      (out+j+4)[0] = (ushort)in[i+2];
-      (out+j+4)[1] = (ushort)in[0];
+  for (j = i = 0; j < nr; j+=2, i++) { 
+      (out+j)[0] = (ushort)in[i];
+      (out+j)[1] = (ushort)in[(i+1)%(nr/2)];
    }
 }
 static void translate_tris_ushort2uint(
@@ -702,13 +682,9 @@ static void translate_polygon_ushort2uint(
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
-  for (j = i = 0; j < nr; j+=6, i++) { 
-      (out+j)[0] = (uint)in[0];
-      (out+j)[1] = (uint)in[i+1];
-      (out+j+2)[0] = (uint)in[i+1];
-      (out+j+2)[1] = (uint)in[i+2];
-      (out+j+4)[0] = (uint)in[i+2];
-      (out+j+4)[1] = (uint)in[0];
+  for (j = i = 0; j < nr; j+=2, i++) { 
+      (out+j)[0] = (uint)in[i];
+      (out+j)[1] = (uint)in[(i+1)%(nr/2)];
    }
 }
 static void translate_tris_uint2ushort(
@@ -814,13 +790,9 @@ static void translate_polygon_uint2ushort(
   ushort *out = (ushort*)_out;
   unsigned i, j;
   (void)j;
-  for (j = i = 0; j < nr; j+=6, i++) { 
-      (out+j)[0] = (ushort)in[0];
-      (out+j)[1] = (ushort)in[i+1];
-      (out+j+2)[0] = (ushort)in[i+1];
-      (out+j+2)[1] = (ushort)in[i+2];
-      (out+j+4)[0] = (ushort)in[i+2];
-      (out+j+4)[1] = (ushort)in[0];
+  for (j = i = 0; j < nr; j+=2, i++) { 
+      (out+j)[0] = (ushort)in[i];
+      (out+j)[1] = (ushort)in[(i+1)%(nr/2)];
    }
 }
 static void translate_tris_uint2uint(
@@ -926,13 +898,9 @@ static void translate_polygon_uint2uint(
   uint *out = (uint*)_out;
   unsigned i, j;
   (void)j;
-  for (j = i = 0; j < nr; j+=6, i++) { 
-      (out+j)[0] = (uint)in[0];
-      (out+j)[1] = (uint)in[i+1];
-      (out+j+2)[0] = (uint)in[i+1];
-      (out+j+2)[1] = (uint)in[i+2];
-      (out+j+4)[0] = (uint)in[i+2];
-      (out+j+4)[1] = (uint)in[0];
+  for (j = i = 0; j < nr; j+=2, i++) { 
+      (out+j)[0] = (uint)in[i];
+      (out+j)[1] = (uint)in[(i+1)%(nr/2)];
    }
 }
 void u_unfilled_init( void )
diff --git a/lib/libGL/gallium/libmesagallium/Makefile b/lib/libGL/gallium/libmesagallium/Makefile
index ce9db71d6..54dedbee4 100644
--- a/lib/libGL/gallium/libmesagallium/Makefile
+++ b/lib/libGL/gallium/libmesagallium/Makefile
@@ -1,4 +1,4 @@
-# $OpenBSD: Makefile,v 1.1 2012/07/13 12:43:34 mpi Exp $
+# $OpenBSD: Makefile,v 1.2 2012/08/17 13:58:00 mpi Exp $
 
 LIB=	mesagallium
 
@@ -103,6 +103,7 @@ MAIN_SOURCES = \
 	multisample.c \
 	nvprogram.c \
 	pack.c \
+	pbo.c \
 	pixel.c \
 	pixelstore.c \
 	pixeltransfer.c \
@@ -114,6 +115,7 @@ MAIN_SOURCES = \
 	readpix.c \
 	remap.c \
 	renderbuffer.c \
+	samplerobj.c \
 	scissor.c \
 	shaderapi.c \
 	shaderobj.c \
@@ -122,10 +124,10 @@ MAIN_SOURCES = \
 	stencil.c \
 	syncobj.c \
 	texcompress.c \
+	texcompress_rgtc.c \
 	texcompress_s3tc.c \
 	texcompress_fxt1.c \
 	texenv.c \
-	texenvprogram.c \
 	texfetch.c \
 	texformat.c \
 	texgen.c \
@@ -134,9 +136,9 @@ MAIN_SOURCES = \
 	texobj.c \
 	texpal.c \
 	texparam.c \
-	texrender.c \
 	texstate.c \
 	texstore.c \
+	texturebarrier.c \
 	transformfeedback.c \
 	uniforms.c \
 	varray.c \
@@ -145,6 +147,9 @@ MAIN_SOURCES = \
 	vtxfmt.c \
 	${MAIN_ES_SOURCES}
 
+MAIN_CXX_SOURCES = \
+	ff_fragment_shader.cpp
+
 # math
 MATH_SOURCES= \
 	m_debug_clip.c \
@@ -203,8 +208,10 @@ STATETRACKER_SOURCES = \
 	st_cb_queryobj.c \
 	st_cb_rasterpos.c \
 	st_cb_readpixels.c \
+	st_cb_syncobj.c \
 	st_cb_strings.c \
 	st_cb_texture.c \
+	st_cb_texturebarrier.c \
 	st_cb_viewport.c \
 	st_cb_xformfb.c \
 	st_context.c \
@@ -272,6 +279,7 @@ MESAGALLIUM_SOURCES = \
 	${DRISW_SOURCES}
 
 MESAGALLIUM_CXX_SOURCES = \
+	${MAIN_CXX_SOURCES} \
 	${SHADER_CXX_SOURCES}
 
 SRCS=	${MESAGALLIUM_SOURCES}	\
diff --git a/lib/libGL/gallium/libmesagallium/generated/api_exec_es1.c b/lib/libGL/gallium/libmesagallium/generated/api_exec_es1.c
index 54b3f0d62..e4d9566e0 100644
--- a/lib/libGL/gallium/libmesagallium/generated/api_exec_es1.c
+++ b/lib/libGL/gallium/libmesagallium/generated/api_exec_es1.c
@@ -1,17 +1,23 @@
 /* DO NOT EDIT *************************************************
  * THIS FILE AUTOMATICALLY GENERATED BY THE es_generator.py SCRIPT
- * API specification file:   /share/OpenBSD/xenocara/lib/libGL/libmesa/generated/../../../../dist/Mesa/src/mesa/main/APIspec.xml
+ * API specification file:   /usr/xenocara/lib/libGL/gallium/libmesagallium/generated/../../../../../dist/Mesa/src/mesa/main/APIspec.xml
  * GLES version:             GLES1.1
- * date:                     2011-05-15 15:41:33
+ * date:                     2012-07-13 17:41:04
  */
 
 
 #include "GLES/gl.h"
 #include "GLES/glext.h"
 #include "main/mfeatures.h"
+#include "main/compiler.h"
+#include "main/api_exec.h"
 
 #if FEATURE_ES1
 
+#ifndef GLAPIENTRYP
+#define GLAPIENTRYP GL_APIENTRYP
+#endif
+
 
 /* These types are needed for the Mesa veneer, but are not defined in
  * the standard GLES headers.
@@ -23,49 +29,6 @@ typedef double GLclampd;
 extern void *_mesa_get_current_context(void);
 extern void _mesa_error(void *ctx, GLenum error, const char *fmtString, ... );
 
-#include "main/compiler.h"
-#include "main/api_exec.h"
-#include "main/remap.h"
-
-/* cannot include main/dispatch.h here */
-#ifdef IN_DRI_DRIVER
-#define _GLAPI_USE_REMAP_TABLE
-#endif
-/* glapi uses GLAPIENTRY while GLES headers define GL_APIENTRY */
-#ifndef GLAPIENTRY
-#define GLAPIENTRY GL_APIENTRY
-#endif
-#include "es1api/glapi/glapitable.h"
-#include "es1api/glapi/glapioffsets.h"
-#include "es1api/glapi/glapidispatch.h"
-
-#if FEATURE_remap_table
-
-#if !FEATURE_GL
-int driDispatchRemapTable[driDispatchRemapTable_size];
-#endif
-
-#define need_MESA_remap_table
-
-#include "es1api/main/remap_helper.h"
-
-void
-_mesa_init_remap_table_es1(void)
-{
-   _mesa_do_init_remap_table(_mesa_function_pool,
-                             driDispatchRemapTable_size,
-                             MESA_remap_table_functions);
-}
-
-void
-_mesa_map_static_functions_es1(void)
-{
-}
-
-#endif
-
-typedef void (*_glapi_proc)(void); /* generic function pointer */
-
 /*************************************************************
  * Generated functions begin here
  */
@@ -586,10 +549,11 @@ static void GL_APIENTRY _es_ClipPlanex(GLenum plane, const GLfixed *equation)
     _mesa_ClipPlane(plane, converted_equation);
 }
 
-extern void GL_APIENTRY _vbo_Color4f(GLfloat red, GLfloat green, GLfloat blue, GLfloat alpha);
-#define _es_Color4f _vbo_Color4f
+/* this function is special and is defined elsewhere */
+extern void GL_APIENTRY _es_Color4f(GLfloat red, GLfloat green, GLfloat blue, GLfloat alpha);
+#define _check_Color4f _es_Color4f
 
-extern void GL_APIENTRY _vbo_Color4f(GLfloat  red, GLfloat  green, GLfloat  blue, GLfloat  alpha);
+extern void GL_APIENTRY _es_Color4f(GLfloat  red, GLfloat  green, GLfloat  blue, GLfloat  alpha);
 static void GL_APIENTRY _es_Color4ub(GLubyte red, GLubyte green, GLubyte blue, GLubyte alpha)
 {
     GLfloat converted_red;
@@ -602,10 +566,10 @@ static void GL_APIENTRY _es_Color4ub(GLubyte red, GLubyte green, GLubyte blue, G
     converted_blue = (GLfloat) (blue / 255.0f);
     converted_alpha = (GLfloat) (alpha / 255.0f);
 
-    _vbo_Color4f(converted_red, converted_green, converted_blue, converted_alpha);
+    _es_Color4f(converted_red, converted_green, converted_blue, converted_alpha);
 }
 
-extern void GL_APIENTRY _vbo_Color4f(GLfloat  red, GLfloat  green, GLfloat  blue, GLfloat  alpha);
+extern void GL_APIENTRY _es_Color4f(GLfloat  red, GLfloat  green, GLfloat  blue, GLfloat  alpha);
 static void GL_APIENTRY _es_Color4x(GLfixed red, GLfixed green, GLfixed blue, GLfixed alpha)
 {
     GLfloat converted_red;
@@ -618,7 +582,7 @@ static void GL_APIENTRY _es_Color4x(GLfixed red, GLfixed green, GLfixed blue, GL
     converted_blue = (GLfloat) (blue / 65536.0f);
     converted_alpha = (GLfloat) (alpha / 65536.0f);
 
-    _vbo_Color4f(converted_red, converted_green, converted_blue, converted_alpha);
+    _es_Color4f(converted_red, converted_green, converted_blue, converted_alpha);
 }
 
 extern void GL_APIENTRY _mesa_ColorMask(GLboolean red, GLboolean green, GLboolean blue, GLboolean alpha);
@@ -2548,8 +2512,9 @@ static void * GL_APIENTRY _es_MapBufferOES(GLenum target, GLenum access)
     return _mesa_MapBufferARB(target, access);
 }
 
-extern void GL_APIENTRY _vbo_Materialf(GLenum face, GLenum pname, GLfloat param);
-static void GL_APIENTRY _es_Materialf(GLenum face, GLenum pname, GLfloat param)
+/* this function is special and is defined elsewhere */
+extern void GL_APIENTRY _es_Materialf(GLenum face, GLenum pname, GLfloat param);
+static void GL_APIENTRY _check_Materialf(GLenum face, GLenum pname, GLfloat param)
 {
     switch(face) {
         case GL_FRONT_AND_BACK:
@@ -2566,11 +2531,12 @@ static void GL_APIENTRY _es_Materialf(GLenum face, GLenum pname, GLfloat param)
             return;
     }
 
-    _vbo_Materialf(face, pname, param);
+    _es_Materialf(face, pname, param);
 }
 
-extern void GL_APIENTRY _vbo_Materialfv(GLenum face, GLenum pname, const GLfloat * params);
-static void GL_APIENTRY _es_Materialfv(GLenum face, GLenum pname, const GLfloat *params)
+/* this function is special and is defined elsewhere */
+extern void GL_APIENTRY _es_Materialfv(GLenum face, GLenum pname, const GLfloat * params);
+static void GL_APIENTRY _check_Materialfv(GLenum face, GLenum pname, const GLfloat *params)
 {
     switch(face) {
         case GL_FRONT_AND_BACK:
@@ -2592,10 +2558,10 @@ static void GL_APIENTRY _es_Materialfv(GLenum face, GLenum pname, const GLfloat
             return;
     }
 
-    _vbo_Materialfv(face, pname, params);
+    _es_Materialfv(face, pname, params);
 }
 
-extern void GL_APIENTRY _vbo_Materialf(GLenum face, GLenum pname, GLfloat  param);
+extern void GL_APIENTRY _es_Materialf(GLenum face, GLenum pname, GLfloat  param);
 static void GL_APIENTRY _es_Materialx(GLenum face, GLenum pname, GLfixed param)
 {
     GLfloat converted_param;
@@ -2617,10 +2583,10 @@ static void GL_APIENTRY _es_Materialx(GLenum face, GLenum pname, GLfixed param)
 
     converted_param = (GLfloat) (param / 65536.0f);
 
-    _vbo_Materialf(face, pname, converted_param);
+    _es_Materialf(face, pname, converted_param);
 }
 
-extern void GL_APIENTRY _vbo_Materialfv(GLenum face, GLenum pname, GLfloat const * params);
+extern void GL_APIENTRY _es_Materialfv(GLenum face, GLenum pname, GLfloat const * params);
 static void GL_APIENTRY _es_Materialxv(GLenum face, GLenum pname, const GLfixed *params)
 {
     register unsigned int i;
@@ -2654,7 +2620,7 @@ static void GL_APIENTRY _es_Materialxv(GLenum face, GLenum pname, const GLfixed
         converted_params[i] = (GLfloat) (params[i] / 65536.0f);
     }
 
-    _vbo_Materialfv(face, pname, converted_params);
+    _es_Materialfv(face, pname, converted_params);
 }
 
 extern void GL_APIENTRY _mesa_MatrixMode(GLenum mode);
@@ -2740,8 +2706,9 @@ static void GL_APIENTRY _es_MultiDrawElementsEXT(GLenum mode, const GLsizei *cou
     _mesa_MultiDrawElementsEXT(mode, count, type, indices, primcount);
 }
 
-extern void GL_APIENTRY _vbo_MultiTexCoord4f(GLenum texture, GLfloat s, GLfloat t, GLfloat r, GLfloat q);
-static void GL_APIENTRY _es_MultiTexCoord4f(GLenum texture, GLfloat s, GLfloat t, GLfloat r, GLfloat q)
+/* this function is special and is defined elsewhere */
+extern void GL_APIENTRY _es_MultiTexCoord4f(GLenum texture, GLfloat s, GLfloat t, GLfloat r, GLfloat q);
+static void GL_APIENTRY _check_MultiTexCoord4f(GLenum texture, GLfloat s, GLfloat t, GLfloat r, GLfloat q)
 {
     switch(texture) {
         case GL_TEXTURE0:
@@ -2782,10 +2749,10 @@ static void GL_APIENTRY _es_MultiTexCoord4f(GLenum texture, GLfloat s, GLfloat t
             return;
     }
 
-    _vbo_MultiTexCoord4f(texture, s, t, r, q);
+    _es_MultiTexCoord4f(texture, s, t, r, q);
 }
 
-extern void GL_APIENTRY _vbo_MultiTexCoord4f(GLenum texture, GLfloat  s, GLfloat  t, GLfloat  r, GLfloat  q);
+extern void GL_APIENTRY _es_MultiTexCoord4f(GLenum texture, GLfloat  s, GLfloat  t, GLfloat  r, GLfloat  q);
 static void GL_APIENTRY _es_MultiTexCoord4x(GLenum texture, GLfixed s, GLfixed t, GLfixed r, GLfixed q)
 {
     GLfloat converted_s;
@@ -2837,13 +2804,14 @@ static void GL_APIENTRY _es_MultiTexCoord4x(GLenum texture, GLfixed s, GLfixed t
     converted_r = (GLfloat) (r / 65536.0f);
     converted_q = (GLfloat) (q / 65536.0f);
 
-    _vbo_MultiTexCoord4f(texture, converted_s, converted_t, converted_r, converted_q);
+    _es_MultiTexCoord4f(texture, converted_s, converted_t, converted_r, converted_q);
 }
 
-extern void GL_APIENTRY _vbo_Normal3f(GLfloat nx, GLfloat ny, GLfloat nz);
-#define _es_Normal3f _vbo_Normal3f
+/* this function is special and is defined elsewhere */
+extern void GL_APIENTRY _es_Normal3f(GLfloat nx, GLfloat ny, GLfloat nz);
+#define _check_Normal3f _es_Normal3f
 
-extern void GL_APIENTRY _vbo_Normal3f(GLfloat  nx, GLfloat  ny, GLfloat  nz);
+extern void GL_APIENTRY _es_Normal3f(GLfloat  nx, GLfloat  ny, GLfloat  nz);
 static void GL_APIENTRY _es_Normal3x(GLfixed nx, GLfixed ny, GLfixed nz)
 {
     GLfloat converted_nx;
@@ -2854,7 +2822,7 @@ static void GL_APIENTRY _es_Normal3x(GLfixed nx, GLfixed ny, GLfixed nz)
     converted_ny = (GLfloat) (ny / 65536.0f);
     converted_nz = (GLfloat) (nz / 65536.0f);
 
-    _vbo_Normal3f(converted_nx, converted_ny, converted_nz);
+    _es_Normal3f(converted_nx, converted_ny, converted_nz);
 }
 
 extern void GL_APIENTRY _mesa_NormalPointer(GLenum type, GLsizei stride, const GLvoid * pointer);
@@ -4131,10 +4099,34 @@ static void GL_APIENTRY _es_TexImage2D(GLenum target, GLint level, GLint interna
     }
     switch(internalFormat) {
         case GL_ALPHA:
+            if (format != GL_ALPHA) {
+                _mesa_error(_mesa_get_current_context(), GL_INVALID_VALUE, "glTexImage2D(internalFormat=0x%x)", internalFormat);
+                return;
+            }
+            break;
         case GL_RGB:
+            if (format != GL_RGB) {
+                _mesa_error(_mesa_get_current_context(), GL_INVALID_VALUE, "glTexImage2D(internalFormat=0x%x)", internalFormat);
+                return;
+            }
+            break;
         case GL_RGBA:
+            if (format != GL_RGBA) {
+                _mesa_error(_mesa_get_current_context(), GL_INVALID_VALUE, "glTexImage2D(internalFormat=0x%x)", internalFormat);
+                return;
+            }
+            break;
         case GL_LUMINANCE:
+            if (format != GL_LUMINANCE) {
+                _mesa_error(_mesa_get_current_context(), GL_INVALID_VALUE, "glTexImage2D(internalFormat=0x%x)", internalFormat);
+                return;
+            }
+            break;
         case GL_LUMINANCE_ALPHA:
+            if (format != GL_LUMINANCE_ALPHA) {
+                _mesa_error(_mesa_get_current_context(), GL_INVALID_VALUE, "glTexImage2D(internalFormat=0x%x)", internalFormat);
+                return;
+            }
             break;
         default:
             _mesa_error(_mesa_get_current_context(), GL_INVALID_VALUE, "glTexImage2D(internalFormat=%d)", internalFormat);
@@ -4617,14 +4609,66 @@ extern void GL_APIENTRY _mesa_Viewport(GLint x, GLint y, GLsizei width, GLsizei
 #define _es_Viewport _mesa_Viewport
 
 
+#include "glapi/glapi.h"
+
+#if FEATURE_remap_table
+
+/* define esLocalRemapTable */
+#include "es1api/main/dispatch.h"
+
+#define need_MESA_remap_table
+#include "es1api/main/remap_helper.h"
+
+static void
+init_remap_table(void)
+{
+   _glthread_DECLARE_STATIC_MUTEX(mutex);
+   static GLboolean initialized = GL_FALSE;
+   const struct gl_function_pool_remap *remap = MESA_remap_table_functions;
+   int i;
+
+   _glthread_LOCK_MUTEX(mutex);
+   if (initialized) {
+      _glthread_UNLOCK_MUTEX(mutex);
+      return;
+   }
+
+   for (i = 0; i < esLocalRemapTable_size; i++) {
+      GLint offset;
+      const char *spec;
+
+      /* sanity check */
+      ASSERT(i == remap[i].remap_index);
+      spec = _mesa_function_pool + remap[i].pool_index;
+
+      offset = _mesa_map_function_spec(spec);
+      esLocalRemapTable[i] = offset;
+   }
+   initialized = GL_TRUE;
+   _glthread_UNLOCK_MUTEX(mutex);
+}
+
+#else /* FEATURE_remap_table */
+
+#include "es1api/main/dispatch.h"
+
+static INLINE void
+init_remap_table(void)
+{
+}
+
+#endif /* FEATURE_remap_table */
+
 struct _glapi_table *
 _mesa_create_exec_table_es1(void)
 {
    struct _glapi_table *exec;
-   exec = _mesa_alloc_dispatch_table(sizeof *exec);
+
+   exec = _mesa_alloc_dispatch_table(_gloffset_COUNT);
    if (exec == NULL)
       return NULL;
 
+   init_remap_table();
 
     SET_ActiveTexture(exec, _es_ActiveTexture);
     SET_AlphaFunc(exec, _es_AlphaFunc);
@@ -4649,7 +4693,7 @@ _mesa_create_exec_table_es1(void)
     SET_ClientActiveTexture(exec, _es_ClientActiveTexture);
     SET_ClipPlanef(exec, _es_ClipPlanef);
     SET_ClipPlanex(exec, _es_ClipPlanex);
-    SET_Color4f(exec, _es_Color4f);
+    SET_Color4f(exec, _check_Color4f);
     SET_Color4ub(exec, _es_Color4ub);
     SET_Color4x(exec, _es_Color4x);
     SET_ColorMask(exec, _es_ColorMask);
@@ -4746,8 +4790,8 @@ _mesa_create_exec_table_es1(void)
     SET_LoadMatrixx(exec, _es_LoadMatrixx);
     SET_LogicOp(exec, _es_LogicOp);
     SET_MapBufferOES(exec, _es_MapBufferOES);
-    SET_Materialf(exec, _es_Materialf);
-    SET_Materialfv(exec, _es_Materialfv);
+    SET_Materialf(exec, _check_Materialf);
+    SET_Materialfv(exec, _check_Materialfv);
     SET_Materialx(exec, _es_Materialx);
     SET_Materialxv(exec, _es_Materialxv);
     SET_MatrixMode(exec, _es_MatrixMode);
@@ -4755,9 +4799,9 @@ _mesa_create_exec_table_es1(void)
     SET_MultMatrixx(exec, _es_MultMatrixx);
     SET_MultiDrawArraysEXT(exec, _es_MultiDrawArraysEXT);
     SET_MultiDrawElementsEXT(exec, _es_MultiDrawElementsEXT);
-    SET_MultiTexCoord4f(exec, _es_MultiTexCoord4f);
+    SET_MultiTexCoord4f(exec, _check_MultiTexCoord4f);
     SET_MultiTexCoord4x(exec, _es_MultiTexCoord4x);
-    SET_Normal3f(exec, _es_Normal3f);
+    SET_Normal3f(exec, _check_Normal3f);
     SET_Normal3x(exec, _es_Normal3x);
     SET_NormalPointer(exec, _es_NormalPointer);
     SET_Orthof(exec, _es_Orthof);
diff --git a/lib/libGL/gallium/libmesagallium/generated/api_exec_es2.c b/lib/libGL/gallium/libmesagallium/generated/api_exec_es2.c
index d28291b0e..f00b3d4db 100644
--- a/lib/libGL/gallium/libmesagallium/generated/api_exec_es2.c
+++ b/lib/libGL/gallium/libmesagallium/generated/api_exec_es2.c
@@ -1,17 +1,23 @@
 /* DO NOT EDIT *************************************************
  * THIS FILE AUTOMATICALLY GENERATED BY THE es_generator.py SCRIPT
- * API specification file:   /share/OpenBSD/xenocara/lib/libGL/libmesa/generated/../../../../dist/Mesa/src/mesa/main/APIspec.xml
+ * API specification file:   /usr/xenocara/lib/libGL/gallium/libmesagallium/generated/../../../../../dist/Mesa/src/mesa/main/APIspec.xml
  * GLES version:             GLES2.0
- * date:                     2011-05-15 15:41:34
+ * date:                     2012-07-13 17:41:05
  */
 
 
 #include "GLES2/gl2.h"
 #include "GLES2/gl2ext.h"
 #include "main/mfeatures.h"
+#include "main/compiler.h"
+#include "main/api_exec.h"
 
 #if FEATURE_ES2
 
+#ifndef GLAPIENTRYP
+#define GLAPIENTRYP GL_APIENTRYP
+#endif
+
 
 /* These types are needed for the Mesa veneer, but are not defined in
  * the standard GLES headers.
@@ -23,49 +29,6 @@ typedef double GLclampd;
 extern void *_mesa_get_current_context(void);
 extern void _mesa_error(void *ctx, GLenum error, const char *fmtString, ... );
 
-#include "main/compiler.h"
-#include "main/api_exec.h"
-#include "main/remap.h"
-
-/* cannot include main/dispatch.h here */
-#ifdef IN_DRI_DRIVER
-#define _GLAPI_USE_REMAP_TABLE
-#endif
-/* glapi uses GLAPIENTRY while GLES headers define GL_APIENTRY */
-#ifndef GLAPIENTRY
-#define GLAPIENTRY GL_APIENTRY
-#endif
-#include "es2api/glapi/glapitable.h"
-#include "es2api/glapi/glapioffsets.h"
-#include "es2api/glapi/glapidispatch.h"
-
-#if FEATURE_remap_table
-
-#if !FEATURE_GL
-int driDispatchRemapTable[driDispatchRemapTable_size];
-#endif
-
-#define need_MESA_remap_table
-
-#include "es2api/main/remap_helper.h"
-
-void
-_mesa_init_remap_table_es2(void)
-{
-   _mesa_do_init_remap_table(_mesa_function_pool,
-                             driDispatchRemapTable_size,
-                             MESA_remap_table_functions);
-}
-
-void
-_mesa_map_static_functions_es2(void)
-{
-}
-
-#endif
-
-typedef void (*_glapi_proc)(void); /* generic function pointer */
-
 /*************************************************************
  * Generated functions begin here
  */
@@ -1768,12 +1731,52 @@ static void GL_APIENTRY _es_TexImage2D(GLenum target, GLint level, GLint interna
     }
     switch(internalFormat) {
         case GL_ALPHA:
+            if (format != GL_ALPHA) {
+                _mesa_error(_mesa_get_current_context(), GL_INVALID_VALUE, "glTexImage2D(internalFormat=0x%x)", internalFormat);
+                return;
+            }
+            break;
         case GL_RGB:
+            if (format != GL_RGB) {
+                _mesa_error(_mesa_get_current_context(), GL_INVALID_VALUE, "glTexImage2D(internalFormat=0x%x)", internalFormat);
+                return;
+            }
+            break;
         case GL_RGBA:
+            if (format != GL_RGBA) {
+                _mesa_error(_mesa_get_current_context(), GL_INVALID_VALUE, "glTexImage2D(internalFormat=0x%x)", internalFormat);
+                return;
+            }
+            break;
         case GL_LUMINANCE:
+            if (format != GL_LUMINANCE) {
+                _mesa_error(_mesa_get_current_context(), GL_INVALID_VALUE, "glTexImage2D(internalFormat=0x%x)", internalFormat);
+                return;
+            }
+            break;
         case GL_LUMINANCE_ALPHA:
+            if (format != GL_LUMINANCE_ALPHA) {
+                _mesa_error(_mesa_get_current_context(), GL_INVALID_VALUE, "glTexImage2D(internalFormat=0x%x)", internalFormat);
+                return;
+            }
+            break;
         case GL_DEPTH_COMPONENT:
+            if (format != GL_DEPTH_COMPONENT) {
+                _mesa_error(_mesa_get_current_context(), GL_INVALID_VALUE, "glTexImage2D(internalFormat=0x%x)", internalFormat);
+                return;
+            }
+            break;
         case GL_DEPTH_STENCIL_OES:
+            if (format != GL_DEPTH_STENCIL_OES) {
+                _mesa_error(_mesa_get_current_context(), GL_INVALID_VALUE, "glTexImage2D(internalFormat=0x%x)", internalFormat);
+                return;
+            }
+            break;
+        case GL_BGRA_EXT:
+            if (format != GL_BGRA_EXT) {
+                _mesa_error(_mesa_get_current_context(), GL_INVALID_VALUE, "glTexImage2D(internalFormat=0x%x)", internalFormat);
+                return;
+            }
             break;
         default:
             _mesa_error(_mesa_get_current_context(), GL_INVALID_VALUE, "glTexImage2D(internalFormat=%d)", internalFormat);
@@ -1824,6 +1827,12 @@ static void GL_APIENTRY _es_TexImage2D(GLenum target, GLint level, GLint interna
                 return;
             }
             break;
+        case GL_BGRA_EXT:
+            if (type != GL_UNSIGNED_BYTE) {
+                _mesa_error(_mesa_get_current_context(), GL_INVALID_OPERATION, "glTexImage2D(format=0x%x)", format);
+                return;
+            }
+            break;
         default:
             _mesa_error(_mesa_get_current_context(), GL_INVALID_OPERATION, "glTexImage2D(format=0x%x)", format);
             return;
@@ -2113,6 +2122,12 @@ static void GL_APIENTRY _es_TexSubImage2D(GLenum target, GLint level, GLint xoff
                 return;
             }
             break;
+        case GL_BGRA_EXT:
+            if (type != GL_UNSIGNED_BYTE) {
+                _mesa_error(_mesa_get_current_context(), GL_INVALID_OPERATION, "glTexSubImage2D(format=0x%x)", format);
+                return;
+            }
+            break;
         default:
             _mesa_error(_mesa_get_current_context(), GL_INVALID_OPERATION, "glTexSubImage2D(format=0x%x)", format);
             return;
@@ -2243,29 +2258,37 @@ extern void GL_APIENTRY _mesa_UseProgramObjectARB(GLuint program);
 extern void GL_APIENTRY _mesa_ValidateProgramARB(GLuint program);
 #define _es_ValidateProgram _mesa_ValidateProgramARB
 
-extern void GL_APIENTRY _vbo_VertexAttrib1f(GLuint index, GLfloat x);
-#define _es_VertexAttrib1f _vbo_VertexAttrib1f
+/* this function is special and is defined elsewhere */
+extern void GL_APIENTRY _es_VertexAttrib1f(GLuint index, GLfloat x);
+#define _check_VertexAttrib1f _es_VertexAttrib1f
 
-extern void GL_APIENTRY _vbo_VertexAttrib1fv(GLuint index, const GLfloat * v);
-#define _es_VertexAttrib1fv _vbo_VertexAttrib1fv
+/* this function is special and is defined elsewhere */
+extern void GL_APIENTRY _es_VertexAttrib1fv(GLuint index, const GLfloat * v);
+#define _check_VertexAttrib1fv _es_VertexAttrib1fv
 
-extern void GL_APIENTRY _vbo_VertexAttrib2f(GLuint index, GLfloat x, GLfloat y);
-#define _es_VertexAttrib2f _vbo_VertexAttrib2f
+/* this function is special and is defined elsewhere */
+extern void GL_APIENTRY _es_VertexAttrib2f(GLuint index, GLfloat x, GLfloat y);
+#define _check_VertexAttrib2f _es_VertexAttrib2f
 
-extern void GL_APIENTRY _vbo_VertexAttrib2fv(GLuint index, const GLfloat * v);
-#define _es_VertexAttrib2fv _vbo_VertexAttrib2fv
+/* this function is special and is defined elsewhere */
+extern void GL_APIENTRY _es_VertexAttrib2fv(GLuint index, const GLfloat * v);
+#define _check_VertexAttrib2fv _es_VertexAttrib2fv
 
-extern void GL_APIENTRY _vbo_VertexAttrib3f(GLuint index, GLfloat x, GLfloat y, GLfloat z);
-#define _es_VertexAttrib3f _vbo_VertexAttrib3f
+/* this function is special and is defined elsewhere */
+extern void GL_APIENTRY _es_VertexAttrib3f(GLuint index, GLfloat x, GLfloat y, GLfloat z);
+#define _check_VertexAttrib3f _es_VertexAttrib3f
 
-extern void GL_APIENTRY _vbo_VertexAttrib3fv(GLuint index, const GLfloat * v);
-#define _es_VertexAttrib3fv _vbo_VertexAttrib3fv
+/* this function is special and is defined elsewhere */
+extern void GL_APIENTRY _es_VertexAttrib3fv(GLuint index, const GLfloat * v);
+#define _check_VertexAttrib3fv _es_VertexAttrib3fv
 
-extern void GL_APIENTRY _vbo_VertexAttrib4f(GLuint index, GLfloat x, GLfloat y, GLfloat z, GLfloat w);
-#define _es_VertexAttrib4f _vbo_VertexAttrib4f
+/* this function is special and is defined elsewhere */
+extern void GL_APIENTRY _es_VertexAttrib4f(GLuint index, GLfloat x, GLfloat y, GLfloat z, GLfloat w);
+#define _check_VertexAttrib4f _es_VertexAttrib4f
 
-extern void GL_APIENTRY _vbo_VertexAttrib4fv(GLuint index, const GLfloat * v);
-#define _es_VertexAttrib4fv _vbo_VertexAttrib4fv
+/* this function is special and is defined elsewhere */
+extern void GL_APIENTRY _es_VertexAttrib4fv(GLuint index, const GLfloat * v);
+#define _check_VertexAttrib4fv _es_VertexAttrib4fv
 
 extern void GL_APIENTRY _mesa_VertexAttribPointerARB(GLuint index, GLint size, GLenum type, GLboolean normalized, GLsizei stride, const GLvoid * pointer);
 static void GL_APIENTRY _es_VertexAttribPointer(GLuint index, GLint size, GLenum type, GLboolean normalized, GLsizei stride, const GLvoid *pointer)
@@ -2300,14 +2323,66 @@ extern void GL_APIENTRY _mesa_Viewport(GLint x, GLint y, GLsizei width, GLsizei
 #define _es_Viewport _mesa_Viewport
 
 
+#include "glapi/glapi.h"
+
+#if FEATURE_remap_table
+
+/* define esLocalRemapTable */
+#include "es2api/main/dispatch.h"
+
+#define need_MESA_remap_table
+#include "es2api/main/remap_helper.h"
+
+static void
+init_remap_table(void)
+{
+   _glthread_DECLARE_STATIC_MUTEX(mutex);
+   static GLboolean initialized = GL_FALSE;
+   const struct gl_function_pool_remap *remap = MESA_remap_table_functions;
+   int i;
+
+   _glthread_LOCK_MUTEX(mutex);
+   if (initialized) {
+      _glthread_UNLOCK_MUTEX(mutex);
+      return;
+   }
+
+   for (i = 0; i < esLocalRemapTable_size; i++) {
+      GLint offset;
+      const char *spec;
+
+      /* sanity check */
+      ASSERT(i == remap[i].remap_index);
+      spec = _mesa_function_pool + remap[i].pool_index;
+
+      offset = _mesa_map_function_spec(spec);
+      esLocalRemapTable[i] = offset;
+   }
+   initialized = GL_TRUE;
+   _glthread_UNLOCK_MUTEX(mutex);
+}
+
+#else /* FEATURE_remap_table */
+
+#include "es2api/main/dispatch.h"
+
+static INLINE void
+init_remap_table(void)
+{
+}
+
+#endif /* FEATURE_remap_table */
+
 struct _glapi_table *
 _mesa_create_exec_table_es2(void)
 {
    struct _glapi_table *exec;
-   exec = _mesa_alloc_dispatch_table(sizeof *exec);
+
+   exec = _mesa_alloc_dispatch_table(_gloffset_COUNT);
    if (exec == NULL)
       return NULL;
 
+   init_remap_table();
 
     SET_ActiveTexture(exec, _es_ActiveTexture);
     SET_AttachShader(exec, _es_AttachShader);
@@ -2454,14 +2529,14 @@ _mesa_create_exec_table_es2(void)
     SET_UnmapBufferOES(exec, _es_UnmapBufferOES);
     SET_UseProgram(exec, _es_UseProgram);
     SET_ValidateProgram(exec, _es_ValidateProgram);
-    SET_VertexAttrib1f(exec, _es_VertexAttrib1f);
-    SET_VertexAttrib1fv(exec, _es_VertexAttrib1fv);
-    SET_VertexAttrib2f(exec, _es_VertexAttrib2f);
-    SET_VertexAttrib2fv(exec, _es_VertexAttrib2fv);
-    SET_VertexAttrib3f(exec, _es_VertexAttrib3f);
-    SET_VertexAttrib3fv(exec, _es_VertexAttrib3fv);
-    SET_VertexAttrib4f(exec, _es_VertexAttrib4f);
-    SET_VertexAttrib4fv(exec, _es_VertexAttrib4fv);
+    SET_VertexAttrib1f(exec, _check_VertexAttrib1f);
+    SET_VertexAttrib1fv(exec, _check_VertexAttrib1fv);
+    SET_VertexAttrib2f(exec, _check_VertexAttrib2f);
+    SET_VertexAttrib2fv(exec, _check_VertexAttrib2fv);
+    SET_VertexAttrib3f(exec, _check_VertexAttrib3f);
+    SET_VertexAttrib3fv(exec, _check_VertexAttrib3fv);
+    SET_VertexAttrib4f(exec, _check_VertexAttrib4f);
+    SET_VertexAttrib4fv(exec, _check_VertexAttrib4fv);
     SET_VertexAttribPointer(exec, _es_VertexAttribPointer);
     SET_Viewport(exec, _es_Viewport);
 
diff --git a/lib/libGL/glsl/libglsl/Makefile b/lib/libGL/glsl/libglsl/Makefile
index a404aa459..90d92558e 100644
--- a/lib/libGL/glsl/libglsl/Makefile
+++ b/lib/libGL/glsl/libglsl/Makefile
@@ -1,4 +1,4 @@
-# $OpenBSD: Makefile,v 1.4 2011/10/31 18:36:06 matthieu Exp $
+# $OpenBSD: Makefile,v 1.5 2012/08/17 13:58:00 mpi Exp $
 .include <bsd.xconf.mk>
 
 MESA=	${.CURDIR}/../../../../dist/Mesa/src/mesa
@@ -15,6 +15,7 @@ LIBGLCPP_SOURCES = \
 
 GLCPP_SOURCES = \
 	$(LIBGLCPP_SOURCES) \
+	ralloc.c \
 	glcpp.c
 
 C_SOURCES = \
@@ -41,6 +42,7 @@ CXX_SOURCES = \
 	ir_expression_flattening.cpp \
 	ir_function.cpp \
 	ir_function_can_inline.cpp \
+	ir_function_detect_recursion.cpp \
 	ir_hierarchical_visitor.cpp \
 	ir_hv_accept.cpp \
 	ir_import_prototypes.cpp \
@@ -73,6 +75,7 @@ CXX_SOURCES = \
 	opt_constant_propagation.cpp \
 	opt_constant_variable.cpp \
 	opt_copy_propagation.cpp \
+	opt_copy_propagation_elements.cpp \
 	opt_dead_code.cpp \
 	opt_dead_code_local.cpp \
 	opt_dead_functions.cpp \
@@ -90,7 +93,7 @@ LIB=	libglsl
 
 SRCS=	\
 	$(C_SOURCES) \
-	$(CXX_SOURCES)
+	$(CXX_SOURCES) \
 
 OBJS+= $(C_SOURCES:.c=.so) $(CXX_SOURCES:.cpp=.so)
 
diff --git a/lib/libGL/libmesa/Makefile b/lib/libGL/libmesa/Makefile
index 1d2fef524..3f8a22ad0 100644
--- a/lib/libGL/libmesa/Makefile
+++ b/lib/libGL/libmesa/Makefile
@@ -1,4 +1,4 @@
-# $OpenBSD: Makefile,v 1.13 2011/12/03 15:36:21 miod Exp $
+# $OpenBSD: Makefile,v 1.14 2012/08/17 13:58:00 mpi Exp $
 
 MESA=	${.CURDIR}/../../../dist/Mesa/src/mesa
 MAPI=	${.CURDIR}/../../../dist/Mesa/src/mapi
@@ -102,6 +102,7 @@ MAIN_SOURCES = \
 	multisample.c \
 	nvprogram.c \
 	pack.c \
+	pbo.c \
 	pixel.c \
 	pixelstore.c \
 	pixeltransfer.c \
@@ -113,6 +114,7 @@ MAIN_SOURCES = \
 	readpix.c \
 	remap.c \
 	renderbuffer.c \
+	samplerobj.c \
 	scissor.c \
 	shaderapi.c \
 	shaderobj.c \
@@ -121,10 +123,10 @@ MAIN_SOURCES = \
 	stencil.c \
 	syncobj.c \
 	texcompress.c \
+	texcompress_rgtc.c \
 	texcompress_s3tc.c \
 	texcompress_fxt1.c \
 	texenv.c \
-	texenvprogram.c \
 	texfetch.c \
 	texformat.c \
 	texgen.c \
@@ -133,9 +135,9 @@ MAIN_SOURCES = \
 	texobj.c \
 	texpal.c \
 	texparam.c \
-	texrender.c \
 	texstate.c \
 	texstore.c \
+	texturebarrier.c \
 	transformfeedback.c \
 	uniforms.c \
 	varray.c \
@@ -144,6 +146,9 @@ MAIN_SOURCES = \
 	vtxfmt.c \
 	$(MAIN_ES_SOURCES)
 
+MAIN_CXX_SOURCES = \
+	ff_fragment_shader.cpp
+
 # math
 MATH_SOURCES= \
 	m_debug_clip.c \
@@ -184,6 +189,7 @@ SWRAST_SOURCES= \
 	s_stencil.c \
 	s_texcombine.c \
 	s_texfilter.c \
+	s_texrender.c \
 	s_triangle.c \
 	s_zoom.c
 
@@ -319,6 +325,7 @@ MESA_SOURCES = \
 	$(ASM_C_SOURCES)
 
 MESA_CXX_SOURCES = \
+	$(MAIN_CXX_SOURCES) \
 	$(SHADER_CXX_SOURCES)
 
 SRCS=	${MESA_SOURCES}	\
diff --git a/lib/libGL/libmesa/generated/api_exec_es1.c b/lib/libGL/libmesa/generated/api_exec_es1.c
index 54b3f0d62..6b8b4f8b3 100644
--- a/lib/libGL/libmesa/generated/api_exec_es1.c
+++ b/lib/libGL/libmesa/generated/api_exec_es1.c
@@ -1,17 +1,23 @@
 /* DO NOT EDIT *************************************************
  * THIS FILE AUTOMATICALLY GENERATED BY THE es_generator.py SCRIPT
- * API specification file:   /share/OpenBSD/xenocara/lib/libGL/libmesa/generated/../../../../dist/Mesa/src/mesa/main/APIspec.xml
+ * API specification file:   /usr/xenocara/lib/libGL/libmesa/generated/../../../../dist/Mesa/src/mesa/main/APIspec.xml
  * GLES version:             GLES1.1
- * date:                     2011-05-15 15:41:33
+ * date:                     2012-07-13 17:40:03
  */
 
 
 #include "GLES/gl.h"
 #include "GLES/glext.h"
 #include "main/mfeatures.h"
+#include "main/compiler.h"
+#include "main/api_exec.h"
 
 #if FEATURE_ES1
 
+#ifndef GLAPIENTRYP
+#define GLAPIENTRYP GL_APIENTRYP
+#endif
+
 
 /* These types are needed for the Mesa veneer, but are not defined in
  * the standard GLES headers.
@@ -23,49 +29,6 @@ typedef double GLclampd;
 extern void *_mesa_get_current_context(void);
 extern void _mesa_error(void *ctx, GLenum error, const char *fmtString, ... );
 
-#include "main/compiler.h"
-#include "main/api_exec.h"
-#include "main/remap.h"
-
-/* cannot include main/dispatch.h here */
-#ifdef IN_DRI_DRIVER
-#define _GLAPI_USE_REMAP_TABLE
-#endif
-/* glapi uses GLAPIENTRY while GLES headers define GL_APIENTRY */
-#ifndef GLAPIENTRY
-#define GLAPIENTRY GL_APIENTRY
-#endif
-#include "es1api/glapi/glapitable.h"
-#include "es1api/glapi/glapioffsets.h"
-#include "es1api/glapi/glapidispatch.h"
-
-#if FEATURE_remap_table
-
-#if !FEATURE_GL
-int driDispatchRemapTable[driDispatchRemapTable_size];
-#endif
-
-#define need_MESA_remap_table
-
-#include "es1api/main/remap_helper.h"
-
-void
-_mesa_init_remap_table_es1(void)
-{
-   _mesa_do_init_remap_table(_mesa_function_pool,
-                             driDispatchRemapTable_size,
-                             MESA_remap_table_functions);
-}
-
-void
-_mesa_map_static_functions_es1(void)
-{
-}
-
-#endif
-
-typedef void (*_glapi_proc)(void); /* generic function pointer */
-
 /*************************************************************
  * Generated functions begin here
  */
@@ -586,10 +549,11 @@ static void GL_APIENTRY _es_ClipPlanex(GLenum plane, const GLfixed *equation)
     _mesa_ClipPlane(plane, converted_equation);
 }
 
-extern void GL_APIENTRY _vbo_Color4f(GLfloat red, GLfloat green, GLfloat blue, GLfloat alpha);
-#define _es_Color4f _vbo_Color4f
+/* this function is special and is defined elsewhere */
+extern void GL_APIENTRY _es_Color4f(GLfloat red, GLfloat green, GLfloat blue, GLfloat alpha);
+#define _check_Color4f _es_Color4f
 
-extern void GL_APIENTRY _vbo_Color4f(GLfloat  red, GLfloat  green, GLfloat  blue, GLfloat  alpha);
+extern void GL_APIENTRY _es_Color4f(GLfloat  red, GLfloat  green, GLfloat  blue, GLfloat  alpha);
 static void GL_APIENTRY _es_Color4ub(GLubyte red, GLubyte green, GLubyte blue, GLubyte alpha)
 {
     GLfloat converted_red;
@@ -602,10 +566,10 @@ static void GL_APIENTRY _es_Color4ub(GLubyte red, GLubyte green, GLubyte blue, G
     converted_blue = (GLfloat) (blue / 255.0f);
     converted_alpha = (GLfloat) (alpha / 255.0f);
 
-    _vbo_Color4f(converted_red, converted_green, converted_blue, converted_alpha);
+    _es_Color4f(converted_red, converted_green, converted_blue, converted_alpha);
 }
 
-extern void GL_APIENTRY _vbo_Color4f(GLfloat  red, GLfloat  green, GLfloat  blue, GLfloat  alpha);
+extern void GL_APIENTRY _es_Color4f(GLfloat  red, GLfloat  green, GLfloat  blue, GLfloat  alpha);
 static void GL_APIENTRY _es_Color4x(GLfixed red, GLfixed green, GLfixed blue, GLfixed alpha)
 {
     GLfloat converted_red;
@@ -618,7 +582,7 @@ static void GL_APIENTRY _es_Color4x(GLfixed red, GLfixed green, GLfixed blue, GL
     converted_blue = (GLfloat) (blue / 65536.0f);
     converted_alpha = (GLfloat) (alpha / 65536.0f);
 
-    _vbo_Color4f(converted_red, converted_green, converted_blue, converted_alpha);
+    _es_Color4f(converted_red, converted_green, converted_blue, converted_alpha);
 }
 
 extern void GL_APIENTRY _mesa_ColorMask(GLboolean red, GLboolean green, GLboolean blue, GLboolean alpha);
@@ -2548,8 +2512,9 @@ static void * GL_APIENTRY _es_MapBufferOES(GLenum target, GLenum access)
     return _mesa_MapBufferARB(target, access);
 }
 
-extern void GL_APIENTRY _vbo_Materialf(GLenum face, GLenum pname, GLfloat param);
-static void GL_APIENTRY _es_Materialf(GLenum face, GLenum pname, GLfloat param)
+/* this function is special and is defined elsewhere */
+extern void GL_APIENTRY _es_Materialf(GLenum face, GLenum pname, GLfloat param);
+static void GL_APIENTRY _check_Materialf(GLenum face, GLenum pname, GLfloat param)
 {
     switch(face) {
         case GL_FRONT_AND_BACK:
@@ -2566,11 +2531,12 @@ static void GL_APIENTRY _es_Materialf(GLenum face, GLenum pname, GLfloat param)
             return;
     }
 
-    _vbo_Materialf(face, pname, param);
+    _es_Materialf(face, pname, param);
 }
 
-extern void GL_APIENTRY _vbo_Materialfv(GLenum face, GLenum pname, const GLfloat * params);
-static void GL_APIENTRY _es_Materialfv(GLenum face, GLenum pname, const GLfloat *params)
+/* this function is special and is defined elsewhere */
+extern void GL_APIENTRY _es_Materialfv(GLenum face, GLenum pname, const GLfloat * params);
+static void GL_APIENTRY _check_Materialfv(GLenum face, GLenum pname, const GLfloat *params)
 {
     switch(face) {
         case GL_FRONT_AND_BACK:
@@ -2592,10 +2558,10 @@ static void GL_APIENTRY _es_Materialfv(GLenum face, GLenum pname, const GLfloat
             return;
     }
 
-    _vbo_Materialfv(face, pname, params);
+    _es_Materialfv(face, pname, params);
 }
 
-extern void GL_APIENTRY _vbo_Materialf(GLenum face, GLenum pname, GLfloat  param);
+extern void GL_APIENTRY _es_Materialf(GLenum face, GLenum pname, GLfloat  param);
 static void GL_APIENTRY _es_Materialx(GLenum face, GLenum pname, GLfixed param)
 {
     GLfloat converted_param;
@@ -2617,10 +2583,10 @@ static void GL_APIENTRY _es_Materialx(GLenum face, GLenum pname, GLfixed param)
 
     converted_param = (GLfloat) (param / 65536.0f);
 
-    _vbo_Materialf(face, pname, converted_param);
+    _es_Materialf(face, pname, converted_param);
 }
 
-extern void GL_APIENTRY _vbo_Materialfv(GLenum face, GLenum pname, GLfloat const * params);
+extern void GL_APIENTRY _es_Materialfv(GLenum face, GLenum pname, GLfloat const * params);
 static void GL_APIENTRY _es_Materialxv(GLenum face, GLenum pname, const GLfixed *params)
 {
     register unsigned int i;
@@ -2654,7 +2620,7 @@ static void GL_APIENTRY _es_Materialxv(GLenum face, GLenum pname, const GLfixed
         converted_params[i] = (GLfloat) (params[i] / 65536.0f);
     }
 
-    _vbo_Materialfv(face, pname, converted_params);
+    _es_Materialfv(face, pname, converted_params);
 }
 
 extern void GL_APIENTRY _mesa_MatrixMode(GLenum mode);
@@ -2740,8 +2706,9 @@ static void GL_APIENTRY _es_MultiDrawElementsEXT(GLenum mode, const GLsizei *cou
     _mesa_MultiDrawElementsEXT(mode, count, type, indices, primcount);
 }
 
-extern void GL_APIENTRY _vbo_MultiTexCoord4f(GLenum texture, GLfloat s, GLfloat t, GLfloat r, GLfloat q);
-static void GL_APIENTRY _es_MultiTexCoord4f(GLenum texture, GLfloat s, GLfloat t, GLfloat r, GLfloat q)
+/* this function is special and is defined elsewhere */
+extern void GL_APIENTRY _es_MultiTexCoord4f(GLenum texture, GLfloat s, GLfloat t, GLfloat r, GLfloat q);
+static void GL_APIENTRY _check_MultiTexCoord4f(GLenum texture, GLfloat s, GLfloat t, GLfloat r, GLfloat q)
 {
     switch(texture) {
         case GL_TEXTURE0:
@@ -2782,10 +2749,10 @@ static void GL_APIENTRY _es_MultiTexCoord4f(GLenum texture, GLfloat s, GLfloat t
             return;
     }
 
-    _vbo_MultiTexCoord4f(texture, s, t, r, q);
+    _es_MultiTexCoord4f(texture, s, t, r, q);
 }
 
-extern void GL_APIENTRY _vbo_MultiTexCoord4f(GLenum texture, GLfloat  s, GLfloat  t, GLfloat  r, GLfloat  q);
+extern void GL_APIENTRY _es_MultiTexCoord4f(GLenum texture, GLfloat  s, GLfloat  t, GLfloat  r, GLfloat  q);
 static void GL_APIENTRY _es_MultiTexCoord4x(GLenum texture, GLfixed s, GLfixed t, GLfixed r, GLfixed q)
 {
     GLfloat converted_s;
@@ -2837,13 +2804,14 @@ static void GL_APIENTRY _es_MultiTexCoord4x(GLenum texture, GLfixed s, GLfixed t
     converted_r = (GLfloat) (r / 65536.0f);
     converted_q = (GLfloat) (q / 65536.0f);
 
-    _vbo_MultiTexCoord4f(texture, converted_s, converted_t, converted_r, converted_q);
+    _es_MultiTexCoord4f(texture, converted_s, converted_t, converted_r, converted_q);
 }
 
-extern void GL_APIENTRY _vbo_Normal3f(GLfloat nx, GLfloat ny, GLfloat nz);
-#define _es_Normal3f _vbo_Normal3f
+/* this function is special and is defined elsewhere */
+extern void GL_APIENTRY _es_Normal3f(GLfloat nx, GLfloat ny, GLfloat nz);
+#define _check_Normal3f _es_Normal3f
 
-extern void GL_APIENTRY _vbo_Normal3f(GLfloat  nx, GLfloat  ny, GLfloat  nz);
+extern void GL_APIENTRY _es_Normal3f(GLfloat  nx, GLfloat  ny, GLfloat  nz);
 static void GL_APIENTRY _es_Normal3x(GLfixed nx, GLfixed ny, GLfixed nz)
 {
     GLfloat converted_nx;
@@ -2854,7 +2822,7 @@ static void GL_APIENTRY _es_Normal3x(GLfixed nx, GLfixed ny, GLfixed nz)
     converted_ny = (GLfloat) (ny / 65536.0f);
     converted_nz = (GLfloat) (nz / 65536.0f);
 
-    _vbo_Normal3f(converted_nx, converted_ny, converted_nz);
+    _es_Normal3f(converted_nx, converted_ny, converted_nz);
 }
 
 extern void GL_APIENTRY _mesa_NormalPointer(GLenum type, GLsizei stride, const GLvoid * pointer);
@@ -4131,10 +4099,34 @@ static void GL_APIENTRY _es_TexImage2D(GLenum target, GLint level, GLint interna
     }
     switch(internalFormat) {
         case GL_ALPHA:
+            if (format != GL_ALPHA) {
+                _mesa_error(_mesa_get_current_context(), GL_INVALID_VALUE, "glTexImage2D(internalFormat=0x%x)", internalFormat);
+                return;
+            }
+            break;
         case GL_RGB:
+            if (format != GL_RGB) {
+                _mesa_error(_mesa_get_current_context(), GL_INVALID_VALUE, "glTexImage2D(internalFormat=0x%x)", internalFormat);
+                return;
+            }
+            break;
         case GL_RGBA:
+            if (format != GL_RGBA) {
+                _mesa_error(_mesa_get_current_context(), GL_INVALID_VALUE, "glTexImage2D(internalFormat=0x%x)", internalFormat);
+                return;
+            }
+            break;
         case GL_LUMINANCE:
+            if (format != GL_LUMINANCE) {
+                _mesa_error(_mesa_get_current_context(), GL_INVALID_VALUE, "glTexImage2D(internalFormat=0x%x)", internalFormat);
+                return;
+            }
+            break;
         case GL_LUMINANCE_ALPHA:
+            if (format != GL_LUMINANCE_ALPHA) {
+                _mesa_error(_mesa_get_current_context(), GL_INVALID_VALUE, "glTexImage2D(internalFormat=0x%x)", internalFormat);
+                return;
+            }
             break;
         default:
             _mesa_error(_mesa_get_current_context(), GL_INVALID_VALUE, "glTexImage2D(internalFormat=%d)", internalFormat);
@@ -4617,14 +4609,66 @@ extern void GL_APIENTRY _mesa_Viewport(GLint x, GLint y, GLsizei width, GLsizei
 #define _es_Viewport _mesa_Viewport
 
 
+#include "glapi/glapi.h"
+
+#if FEATURE_remap_table
+
+/* define esLocalRemapTable */
+#include "es1api/main/dispatch.h"
+
+#define need_MESA_remap_table
+#include "es1api/main/remap_helper.h"
+
+static void
+init_remap_table(void)
+{
+   _glthread_DECLARE_STATIC_MUTEX(mutex);
+   static GLboolean initialized = GL_FALSE;
+   const struct gl_function_pool_remap *remap = MESA_remap_table_functions;
+   int i;
+
+   _glthread_LOCK_MUTEX(mutex);
+   if (initialized) {
+      _glthread_UNLOCK_MUTEX(mutex);
+      return;
+   }
+
+   for (i = 0; i < esLocalRemapTable_size; i++) {
+      GLint offset;
+      const char *spec;
+
+      /* sanity check */
+      ASSERT(i == remap[i].remap_index);
+      spec = _mesa_function_pool + remap[i].pool_index;
+
+      offset = _mesa_map_function_spec(spec);
+      esLocalRemapTable[i] = offset;
+   }
+   initialized = GL_TRUE;
+   _glthread_UNLOCK_MUTEX(mutex);
+}
+
+#else /* FEATURE_remap_table */
+
+#include "es1api/main/dispatch.h"
+
+static INLINE void
+init_remap_table(void)
+{
+}
+
+#endif /* FEATURE_remap_table */
+
 struct _glapi_table *
 _mesa_create_exec_table_es1(void)
 {
    struct _glapi_table *exec;
-   exec = _mesa_alloc_dispatch_table(sizeof *exec);
+
+   exec = _mesa_alloc_dispatch_table(_gloffset_COUNT);
    if (exec == NULL)
       return NULL;
 
+   init_remap_table();
 
     SET_ActiveTexture(exec, _es_ActiveTexture);
     SET_AlphaFunc(exec, _es_AlphaFunc);
@@ -4649,7 +4693,7 @@ _mesa_create_exec_table_es1(void)
     SET_ClientActiveTexture(exec, _es_ClientActiveTexture);
     SET_ClipPlanef(exec, _es_ClipPlanef);
     SET_ClipPlanex(exec, _es_ClipPlanex);
-    SET_Color4f(exec, _es_Color4f);
+    SET_Color4f(exec, _check_Color4f);
     SET_Color4ub(exec, _es_Color4ub);
     SET_Color4x(exec, _es_Color4x);
     SET_ColorMask(exec, _es_ColorMask);
@@ -4746,8 +4790,8 @@ _mesa_create_exec_table_es1(void)
     SET_LoadMatrixx(exec, _es_LoadMatrixx);
     SET_LogicOp(exec, _es_LogicOp);
     SET_MapBufferOES(exec, _es_MapBufferOES);
-    SET_Materialf(exec, _es_Materialf);
-    SET_Materialfv(exec, _es_Materialfv);
+    SET_Materialf(exec, _check_Materialf);
+    SET_Materialfv(exec, _check_Materialfv);
     SET_Materialx(exec, _es_Materialx);
     SET_Materialxv(exec, _es_Materialxv);
     SET_MatrixMode(exec, _es_MatrixMode);
@@ -4755,9 +4799,9 @@ _mesa_create_exec_table_es1(void)
     SET_MultMatrixx(exec, _es_MultMatrixx);
     SET_MultiDrawArraysEXT(exec, _es_MultiDrawArraysEXT);
     SET_MultiDrawElementsEXT(exec, _es_MultiDrawElementsEXT);
-    SET_MultiTexCoord4f(exec, _es_MultiTexCoord4f);
+    SET_MultiTexCoord4f(exec, _check_MultiTexCoord4f);
     SET_MultiTexCoord4x(exec, _es_MultiTexCoord4x);
-    SET_Normal3f(exec, _es_Normal3f);
+    SET_Normal3f(exec, _check_Normal3f);
     SET_Normal3x(exec, _es_Normal3x);
     SET_NormalPointer(exec, _es_NormalPointer);
     SET_Orthof(exec, _es_Orthof);
diff --git a/lib/libGL/libmesa/generated/api_exec_es2.c b/lib/libGL/libmesa/generated/api_exec_es2.c
index d28291b0e..114515b30 100644
--- a/lib/libGL/libmesa/generated/api_exec_es2.c
+++ b/lib/libGL/libmesa/generated/api_exec_es2.c
@@ -1,17 +1,23 @@
 /* DO NOT EDIT *************************************************
  * THIS FILE AUTOMATICALLY GENERATED BY THE es_generator.py SCRIPT
- * API specification file:   /share/OpenBSD/xenocara/lib/libGL/libmesa/generated/../../../../dist/Mesa/src/mesa/main/APIspec.xml
+ * API specification file:   /usr/xenocara/lib/libGL/libmesa/generated/../../../../dist/Mesa/src/mesa/main/APIspec.xml
  * GLES version:             GLES2.0
- * date:                     2011-05-15 15:41:34
+ * date:                     2012-07-13 17:40:05
  */
 
 
 #include "GLES2/gl2.h"
 #include "GLES2/gl2ext.h"
 #include "main/mfeatures.h"
+#include "main/compiler.h"
+#include "main/api_exec.h"
 
 #if FEATURE_ES2
 
+#ifndef GLAPIENTRYP
+#define GLAPIENTRYP GL_APIENTRYP
+#endif
+
 
 /* These types are needed for the Mesa veneer, but are not defined in
  * the standard GLES headers.
@@ -23,49 +29,6 @@ typedef double GLclampd;
 extern void *_mesa_get_current_context(void);
 extern void _mesa_error(void *ctx, GLenum error, const char *fmtString, ... );
 
-#include "main/compiler.h"
-#include "main/api_exec.h"
-#include "main/remap.h"
-
-/* cannot include main/dispatch.h here */
-#ifdef IN_DRI_DRIVER
-#define _GLAPI_USE_REMAP_TABLE
-#endif
-/* glapi uses GLAPIENTRY while GLES headers define GL_APIENTRY */
-#ifndef GLAPIENTRY
-#define GLAPIENTRY GL_APIENTRY
-#endif
-#include "es2api/glapi/glapitable.h"
-#include "es2api/glapi/glapioffsets.h"
-#include "es2api/glapi/glapidispatch.h"
-
-#if FEATURE_remap_table
-
-#if !FEATURE_GL
-int driDispatchRemapTable[driDispatchRemapTable_size];
-#endif
-
-#define need_MESA_remap_table
-
-#include "es2api/main/remap_helper.h"
-
-void
-_mesa_init_remap_table_es2(void)
-{
-   _mesa_do_init_remap_table(_mesa_function_pool,
-                             driDispatchRemapTable_size,
-                             MESA_remap_table_functions);
-}
-
-void
-_mesa_map_static_functions_es2(void)
-{
-}
-
-#endif
-
-typedef void (*_glapi_proc)(void); /* generic function pointer */
-
 /*************************************************************
  * Generated functions begin here
  */
@@ -1768,12 +1731,52 @@ static void GL_APIENTRY _es_TexImage2D(GLenum target, GLint level, GLint interna
     }
     switch(internalFormat) {
         case GL_ALPHA:
+            if (format != GL_ALPHA) {
+                _mesa_error(_mesa_get_current_context(), GL_INVALID_VALUE, "glTexImage2D(internalFormat=0x%x)", internalFormat);
+                return;
+            }
+            break;
         case GL_RGB:
+            if (format != GL_RGB) {
+                _mesa_error(_mesa_get_current_context(), GL_INVALID_VALUE, "glTexImage2D(internalFormat=0x%x)", internalFormat);
+                return;
+            }
+            break;
         case GL_RGBA:
+            if (format != GL_RGBA) {
+                _mesa_error(_mesa_get_current_context(), GL_INVALID_VALUE, "glTexImage2D(internalFormat=0x%x)", internalFormat);
+                return;
+            }
+            break;
         case GL_LUMINANCE:
+            if (format != GL_LUMINANCE) {
+                _mesa_error(_mesa_get_current_context(), GL_INVALID_VALUE, "glTexImage2D(internalFormat=0x%x)", internalFormat);
+                return;
+            }
+            break;
         case GL_LUMINANCE_ALPHA:
+            if (format != GL_LUMINANCE_ALPHA) {
+                _mesa_error(_mesa_get_current_context(), GL_INVALID_VALUE, "glTexImage2D(internalFormat=0x%x)", internalFormat);
+                return;
+            }
+            break;
         case GL_DEPTH_COMPONENT:
+            if (format != GL_DEPTH_COMPONENT) {
+                _mesa_error(_mesa_get_current_context(), GL_INVALID_VALUE, "glTexImage2D(internalFormat=0x%x)", internalFormat);
+                return;
+            }
+            break;
         case GL_DEPTH_STENCIL_OES:
+            if (format != GL_DEPTH_STENCIL_OES) {
+                _mesa_error(_mesa_get_current_context(), GL_INVALID_VALUE, "glTexImage2D(internalFormat=0x%x)", internalFormat);
+                return;
+            }
+            break;
+        case GL_BGRA_EXT:
+            if (format != GL_BGRA_EXT) {
+                _mesa_error(_mesa_get_current_context(), GL_INVALID_VALUE, "glTexImage2D(internalFormat=0x%x)", internalFormat);
+                return;
+            }
             break;
         default:
             _mesa_error(_mesa_get_current_context(), GL_INVALID_VALUE, "glTexImage2D(internalFormat=%d)", internalFormat);
@@ -1824,6 +1827,12 @@ static void GL_APIENTRY _es_TexImage2D(GLenum target, GLint level, GLint interna
                 return;
             }
             break;
+        case GL_BGRA_EXT:
+            if (type != GL_UNSIGNED_BYTE) {
+                _mesa_error(_mesa_get_current_context(), GL_INVALID_OPERATION, "glTexImage2D(format=0x%x)", format);
+                return;
+            }
+            break;
         default:
             _mesa_error(_mesa_get_current_context(), GL_INVALID_OPERATION, "glTexImage2D(format=0x%x)", format);
             return;
@@ -2113,6 +2122,12 @@ static void GL_APIENTRY _es_TexSubImage2D(GLenum target, GLint level, GLint xoff
                 return;
             }
             break;
+        case GL_BGRA_EXT:
+            if (type != GL_UNSIGNED_BYTE) {
+                _mesa_error(_mesa_get_current_context(), GL_INVALID_OPERATION, "glTexSubImage2D(format=0x%x)", format);
+                return;
+            }
+            break;
         default:
             _mesa_error(_mesa_get_current_context(), GL_INVALID_OPERATION, "glTexSubImage2D(format=0x%x)", format);
             return;
@@ -2243,29 +2258,37 @@ extern void GL_APIENTRY _mesa_UseProgramObjectARB(GLuint program);
 extern void GL_APIENTRY _mesa_ValidateProgramARB(GLuint program);
 #define _es_ValidateProgram _mesa_ValidateProgramARB
 
-extern void GL_APIENTRY _vbo_VertexAttrib1f(GLuint index, GLfloat x);
-#define _es_VertexAttrib1f _vbo_VertexAttrib1f
+/* this function is special and is defined elsewhere */
+extern void GL_APIENTRY _es_VertexAttrib1f(GLuint index, GLfloat x);
+#define _check_VertexAttrib1f _es_VertexAttrib1f
 
-extern void GL_APIENTRY _vbo_VertexAttrib1fv(GLuint index, const GLfloat * v);
-#define _es_VertexAttrib1fv _vbo_VertexAttrib1fv
+/* this function is special and is defined elsewhere */
+extern void GL_APIENTRY _es_VertexAttrib1fv(GLuint index, const GLfloat * v);
+#define _check_VertexAttrib1fv _es_VertexAttrib1fv
 
-extern void GL_APIENTRY _vbo_VertexAttrib2f(GLuint index, GLfloat x, GLfloat y);
-#define _es_VertexAttrib2f _vbo_VertexAttrib2f
+/* this function is special and is defined elsewhere */
+extern void GL_APIENTRY _es_VertexAttrib2f(GLuint index, GLfloat x, GLfloat y);
+#define _check_VertexAttrib2f _es_VertexAttrib2f
 
-extern void GL_APIENTRY _vbo_VertexAttrib2fv(GLuint index, const GLfloat * v);
-#define _es_VertexAttrib2fv _vbo_VertexAttrib2fv
+/* this function is special and is defined elsewhere */
+extern void GL_APIENTRY _es_VertexAttrib2fv(GLuint index, const GLfloat * v);
+#define _check_VertexAttrib2fv _es_VertexAttrib2fv
 
-extern void GL_APIENTRY _vbo_VertexAttrib3f(GLuint index, GLfloat x, GLfloat y, GLfloat z);
-#define _es_VertexAttrib3f _vbo_VertexAttrib3f
+/* this function is special and is defined elsewhere */
+extern void GL_APIENTRY _es_VertexAttrib3f(GLuint index, GLfloat x, GLfloat y, GLfloat z);
+#define _check_VertexAttrib3f _es_VertexAttrib3f
 
-extern void GL_APIENTRY _vbo_VertexAttrib3fv(GLuint index, const GLfloat * v);
-#define _es_VertexAttrib3fv _vbo_VertexAttrib3fv
+/* this function is special and is defined elsewhere */
+extern void GL_APIENTRY _es_VertexAttrib3fv(GLuint index, const GLfloat * v);
+#define _check_VertexAttrib3fv _es_VertexAttrib3fv
 
-extern void GL_APIENTRY _vbo_VertexAttrib4f(GLuint index, GLfloat x, GLfloat y, GLfloat z, GLfloat w);
-#define _es_VertexAttrib4f _vbo_VertexAttrib4f
+/* this function is special and is defined elsewhere */
+extern void GL_APIENTRY _es_VertexAttrib4f(GLuint index, GLfloat x, GLfloat y, GLfloat z, GLfloat w);
+#define _check_VertexAttrib4f _es_VertexAttrib4f
 
-extern void GL_APIENTRY _vbo_VertexAttrib4fv(GLuint index, const GLfloat * v);
-#define _es_VertexAttrib4fv _vbo_VertexAttrib4fv
+/* this function is special and is defined elsewhere */
+extern void GL_APIENTRY _es_VertexAttrib4fv(GLuint index, const GLfloat * v);
+#define _check_VertexAttrib4fv _es_VertexAttrib4fv
 
 extern void GL_APIENTRY _mesa_VertexAttribPointerARB(GLuint index, GLint size, GLenum type, GLboolean normalized, GLsizei stride, const GLvoid * pointer);
 static void GL_APIENTRY _es_VertexAttribPointer(GLuint index, GLint size, GLenum type, GLboolean normalized, GLsizei stride, const GLvoid *pointer)
@@ -2300,14 +2323,66 @@ extern void GL_APIENTRY _mesa_Viewport(GLint x, GLint y, GLsizei width, GLsizei
 #define _es_Viewport _mesa_Viewport
 
 
+#include "glapi/glapi.h"
+
+#if FEATURE_remap_table
+
+/* define esLocalRemapTable */
+#include "es2api/main/dispatch.h"
+
+#define need_MESA_remap_table
+#include "es2api/main/remap_helper.h"
+
+static void
+init_remap_table(void)
+{
+   _glthread_DECLARE_STATIC_MUTEX(mutex);
+   static GLboolean initialized = GL_FALSE;
+   const struct gl_function_pool_remap *remap = MESA_remap_table_functions;
+   int i;
+
+   _glthread_LOCK_MUTEX(mutex);
+   if (initialized) {
+      _glthread_UNLOCK_MUTEX(mutex);
+      return;
+   }
+
+   for (i = 0; i < esLocalRemapTable_size; i++) {
+      GLint offset;
+      const char *spec;
+
+      /* sanity check */
+      ASSERT(i == remap[i].remap_index);
+      spec = _mesa_function_pool + remap[i].pool_index;
+
+      offset = _mesa_map_function_spec(spec);
+      esLocalRemapTable[i] = offset;
+   }
+   initialized = GL_TRUE;
+   _glthread_UNLOCK_MUTEX(mutex);
+}
+
+#else /* FEATURE_remap_table */
+
+#include "es2api/main/dispatch.h"
+
+static INLINE void
+init_remap_table(void)
+{
+}
+
+#endif /* FEATURE_remap_table */
+
 struct _glapi_table *
 _mesa_create_exec_table_es2(void)
 {
    struct _glapi_table *exec;
-   exec = _mesa_alloc_dispatch_table(sizeof *exec);
+
+   exec = _mesa_alloc_dispatch_table(_gloffset_COUNT);
    if (exec == NULL)
       return NULL;
 
+   init_remap_table();
 
     SET_ActiveTexture(exec, _es_ActiveTexture);
     SET_AttachShader(exec, _es_AttachShader);
@@ -2454,14 +2529,14 @@ _mesa_create_exec_table_es2(void)
     SET_UnmapBufferOES(exec, _es_UnmapBufferOES);
     SET_UseProgram(exec, _es_UseProgram);
     SET_ValidateProgram(exec, _es_ValidateProgram);
-    SET_VertexAttrib1f(exec, _es_VertexAttrib1f);
-    SET_VertexAttrib1fv(exec, _es_VertexAttrib1fv);
-    SET_VertexAttrib2f(exec, _es_VertexAttrib2f);
-    SET_VertexAttrib2fv(exec, _es_VertexAttrib2fv);
-    SET_VertexAttrib3f(exec, _es_VertexAttrib3f);
-    SET_VertexAttrib3fv(exec, _es_VertexAttrib3fv);
-    SET_VertexAttrib4f(exec, _es_VertexAttrib4f);
-    SET_VertexAttrib4fv(exec, _es_VertexAttrib4fv);
+    SET_VertexAttrib1f(exec, _check_VertexAttrib1f);
+    SET_VertexAttrib1fv(exec, _check_VertexAttrib1fv);
+    SET_VertexAttrib2f(exec, _check_VertexAttrib2f);
+    SET_VertexAttrib2fv(exec, _check_VertexAttrib2fv);
+    SET_VertexAttrib3f(exec, _check_VertexAttrib3f);
+    SET_VertexAttrib3fv(exec, _check_VertexAttrib3fv);
+    SET_VertexAttrib4f(exec, _check_VertexAttrib4f);
+    SET_VertexAttrib4fv(exec, _check_VertexAttrib4fv);
     SET_VertexAttribPointer(exec, _es_VertexAttribPointer);
     SET_Viewport(exec, _es_Viewport);
 
diff --git a/lib/libGL/shlib_version b/lib/libGL/shlib_version
index 56246d02b..262f3bc13 100644
--- a/lib/libGL/shlib_version
+++ b/lib/libGL/shlib_version
@@ -1,2 +1,2 @@
-major=12
+major=13
 minor=0