diff --git a/linux-core/Makefile b/linux-core/Makefile
index effc2061..af29bd6a 100644
--- a/linux-core/Makefile
+++ b/linux-core/Makefile
@@ -238,9 +238,7 @@ drmstat: drmstat.c
 	$(CC) $(PRGCFLAGS) $< -o $@ $(DRMSTATLIBS)
 
 install:
-	echo "Installing DRM modules to "$(O)
-	install *.ko $(O)/driver/char/drm
-	depmod -a
+	make -C $(LINUXDIR) $(GETCONFIG) SUBDIRS=`pwd` DRMSRCDIR=`pwd` modules_install
 
 else
 
diff --git a/linux-core/ati_pcigart.c b/linux-core/ati_pcigart.c
index bb30dd74..52bf8922 100644
--- a/linux-core/ati_pcigart.c
+++ b/linux-core/ati_pcigart.c
@@ -205,10 +205,18 @@ int drm_ati_pcigart_init(drm_device_t *dev, drm_ati_pcigart_info *gart_info)
 		page_base = (u32) entry->busaddr[i];
 
 		for (j = 0; j < (PAGE_SIZE / ATI_PCIGART_PAGE_SIZE); j++) {
-			if (gart_info->is_pcie)
+			switch(gart_info->gart_reg_if) {
+			case DRM_ATI_GART_IGP:
+				*pci_gart = cpu_to_le32((page_base) | 0xc);
+				break;
+			case DRM_ATI_GART_PCIE:
 				*pci_gart = cpu_to_le32((page_base >> 8) | 0xc);
-			else
+				break;
+			default:
+			case DRM_ATI_GART_PCI:
 				*pci_gart = cpu_to_le32(page_base);
+				break;
+			}
 			pci_gart++;
 			page_base += ATI_PCIGART_PAGE_SIZE;
 		}
diff --git a/linux-core/drmP.h b/linux-core/drmP.h
index 25fc8733..0d0ec992 100644
--- a/linux-core/drmP.h
+++ b/linux-core/drmP.h
@@ -75,7 +75,6 @@
 #include <linux/poll.h>
 #include <asm/pgalloc.h>
 #include "drm.h"
-#include "drm_crtc.h"
 #include <linux/slab.h>
 
 #define __OS_HAS_AGP (defined(CONFIG_AGP) || (defined(CONFIG_AGP_MODULE) && defined(MODULE)))
@@ -590,9 +589,13 @@ typedef struct drm_vbl_sig {
 #define DRM_ATI_GART_MAIN 1
 #define DRM_ATI_GART_FB   2
 
+#define DRM_ATI_GART_PCI 1
+#define DRM_ATI_GART_PCIE 2
+#define DRM_ATI_GART_IGP 3
+
 typedef struct ati_pcigart_info {
 	int gart_table_location;
-	int is_pcie;
+	int gart_reg_if;
 	void *addr;
 	dma_addr_t bus_addr;
 	drm_local_map_t mapping;
diff --git a/linux-core/drm_crtc.c b/linux-core/drm_crtc.c
index 2f140dbd..ef025922 100644
--- a/linux-core/drm_crtc.c
+++ b/linux-core/drm_crtc.c
@@ -582,7 +582,7 @@ void drm_mode_config_cleanup(drm_device_t *dev)
 {
 	struct drm_output *output, *ot;
 	struct drm_crtc *crtc, *ct;
-	struct drm_crtc *fb, *fbt;
+	struct drm_framebuffer *fb, *fbt;
 	list_for_each_entry_safe(output, ot, &dev->mode_config.output_list, head) {
 		drm_output_destroy(output);
 	}
diff --git a/shared-core/drm_pciids.txt b/shared-core/drm_pciids.txt
index 5cea11a7..72c28627 100644
--- a/shared-core/drm_pciids.txt
+++ b/shared-core/drm_pciids.txt
@@ -98,6 +98,7 @@
 0x1002 0x5653 CHIP_RV410|RADEON_IS_MOBILITY|RADEON_NEW_MEMMAP "ATI Radeon Mobility X700 M26"
 0x1002 0x5834 CHIP_RS300|RADEON_IS_IGP "ATI Radeon RS300 9100 IGP"
 0x1002 0x5835 CHIP_RS300|RADEON_IS_IGP|RADEON_IS_MOBILITY "ATI Radeon RS300 Mobility IGP"
+0x1002 0x5955 CHIP_RS400|RADEON_IS_IGP|RADEON_IS_MOBILITY|RADEON_IS_IGPGART "ATI Radeon XPRESS 200M 5955"
 0x1002 0x5960 CHIP_RV280 "ATI Radeon RV280 9250"
 0x1002 0x5961 CHIP_RV280 "ATI Radeon RV280 9200"
 0x1002 0x5962 CHIP_RV280 "ATI Radeon RV280 9200"
diff --git a/shared-core/i915_dma.c b/shared-core/i915_dma.c
index a7603405..cf729c6f 100644
--- a/shared-core/i915_dma.c
+++ b/shared-core/i915_dma.c
@@ -616,7 +616,6 @@ static int i915_batchbuffer(DRM_IOCTL_ARGS)
 {
 	DRM_DEVICE;
 	drm_i915_private_t *dev_priv = (drm_i915_private_t *) dev->dev_private;
-	u32 *hw_status = dev_priv->hw_status_page;
 	drm_i915_sarea_t *sarea_priv = (drm_i915_sarea_t *)
 	    dev_priv->sarea_priv;
 	drm_i915_batchbuffer_t batch;
@@ -642,7 +641,7 @@ static int i915_batchbuffer(DRM_IOCTL_ARGS)
 
 	ret = i915_dispatch_batchbuffer(dev, &batch);
 
-	sarea_priv->last_dispatch = (int)hw_status[5];
+	sarea_priv->last_dispatch = READ_BREADCRUMB(dev_priv);
 	return ret;
 }
 
@@ -650,7 +649,6 @@ static int i915_cmdbuffer(DRM_IOCTL_ARGS)
 {
 	DRM_DEVICE;
 	drm_i915_private_t *dev_priv = (drm_i915_private_t *) dev->dev_private;
-	u32 *hw_status = dev_priv->hw_status_page;
 	drm_i915_sarea_t *sarea_priv = (drm_i915_sarea_t *)
 	    dev_priv->sarea_priv;
 	drm_i915_cmdbuffer_t cmdbuf;
@@ -678,7 +676,7 @@ static int i915_cmdbuffer(DRM_IOCTL_ARGS)
 		return ret;
 	}
 
-	sarea_priv->last_dispatch = (int)hw_status[5];
+	sarea_priv->last_dispatch = READ_BREADCRUMB(dev_priv);
 	return 0;
 }
 
diff --git a/shared-core/nouveau_reg.h b/shared-core/nouveau_reg.h
index 3360fec2..ea4a2f6b 100644
--- a/shared-core/nouveau_reg.h
+++ b/shared-core/nouveau_reg.h
@@ -65,6 +65,10 @@
 #define NV03_PMC_ENABLE                                    0x00000200
 #    define NV_PMC_ENABLE_PFIFO                               (1<< 8)
 #    define NV_PMC_ENABLE_PGRAPH                              (1<<12)
+/* Disabling the below bit breaks newer (G7X only?) mobile chipsets,
+ * the card will hang early on in the X init process.
+ */
+#    define NV_PMC_ENABLE_UNK13                               (1<<13)
 #define NV40_PMC_1700                                      0x00001700
 #define NV40_PMC_1704                                      0x00001704
 #define NV40_PMC_1708                                      0x00001708
diff --git a/shared-core/nv04_mc.c b/shared-core/nv04_mc.c
index 2619eb71..0e23efb2 100644
--- a/shared-core/nv04_mc.c
+++ b/shared-core/nv04_mc.c
@@ -8,6 +8,11 @@ nv04_mc_init(drm_device_t *dev)
 {
 	drm_nouveau_private_t *dev_priv = dev->dev_private;
 
+	/* Power up everything, resetting each individual unit will
+	 * be done later if needed.
+	 */
+	NV_WRITE(NV03_PMC_ENABLE, 0xFFFFFFFF);
+
 	NV_WRITE(NV03_PMC_INTR_EN_0, 0);
 
 	return 0;
diff --git a/shared-core/nv10_graph.c b/shared-core/nv10_graph.c
index db2757d7..4cc44e21 100644
--- a/shared-core/nv10_graph.c
+++ b/shared-core/nv10_graph.c
@@ -527,6 +527,37 @@ NV10_PGRAPH_DEBUG_4,
 0x00400a04,
 };
 
+static int nv10_graph_ctx_regs_find_offset(drm_device_t *dev, int reg)
+{
+	drm_nouveau_private_t *dev_priv = dev->dev_private;
+	int i, j;
+	for (i = 0; i < sizeof(nv10_graph_ctx_regs)/sizeof(nv10_graph_ctx_regs[0]); i++) {
+		if (nv10_graph_ctx_regs[i] == reg)
+			return i;
+	}
+	if (dev_priv->chipset>=0x17) {
+		for (j = 0; j < sizeof(nv17_graph_ctx_regs)/sizeof(nv17_graph_ctx_regs[0]); i++,j++) {
+			if (nv17_graph_ctx_regs[j] == reg)
+				return i;
+		}
+	}
+	return -1;
+}
+
+static void restore_ctx_regs(drm_device_t *dev, int channel)
+{
+	drm_nouveau_private_t *dev_priv = dev->dev_private;
+	struct nouveau_fifo *fifo = &dev_priv->fifos[channel];
+	int i, j;
+	for (i = 0; i < sizeof(nv10_graph_ctx_regs)/sizeof(nv10_graph_ctx_regs[0]); i++)
+		NV_WRITE(nv10_graph_ctx_regs[i], fifo->pgraph_ctx[i]);
+	if (dev_priv->chipset>=0x17) {
+		for (j = 0; j < sizeof(nv17_graph_ctx_regs)/sizeof(nv17_graph_ctx_regs[0]); i++,j++)
+			NV_WRITE(nv17_graph_ctx_regs[j], fifo->pgraph_ctx[i]);
+	}
+	nouveau_wait_for_idle(dev);
+}
+
 void nouveau_nv10_context_switch(drm_device_t *dev)
 {
 	drm_nouveau_private_t *dev_priv = dev->dev_private;
@@ -559,15 +590,8 @@ void nouveau_nv10_context_switch(drm_device_t *dev)
 
 	nouveau_wait_for_idle(dev);
 	// restore PGRAPH context
-	//XXX not working yet
 #if 1
-	for (i = 0; i < sizeof(nv10_graph_ctx_regs)/sizeof(nv10_graph_ctx_regs[0]); i++)
-		NV_WRITE(nv10_graph_ctx_regs[i], dev_priv->fifos[channel].pgraph_ctx[i]);
-	if (dev_priv->chipset>=0x17) {
-		for (j = 0; j < sizeof(nv17_graph_ctx_regs)/sizeof(nv17_graph_ctx_regs[0]); i++,j++)
-			NV_WRITE(nv17_graph_ctx_regs[j], dev_priv->fifos[channel].pgraph_ctx[i]);
-	}
-	nouveau_wait_for_idle(dev);
+	restore_ctx_regs(dev, channel);
 #endif
 	
 	NV_WRITE(NV10_PGRAPH_CTX_CONTROL, 0x10010100);
@@ -582,20 +606,52 @@ void nouveau_nv10_context_switch(drm_device_t *dev)
 	NV_WRITE(NV04_PGRAPH_FIFO,0x1);
 }
 
+#define NV_WRITE_CTX(reg, val) do { \
+	int offset = nv10_graph_ctx_regs_find_offset(dev, reg); \
+	if (offset > 0) \
+		fifo->pgraph_ctx[offset] = val; \
+	} while (0)
 int nv10_graph_context_create(drm_device_t *dev, int channel) {
 	drm_nouveau_private_t *dev_priv = dev->dev_private;
+	struct nouveau_fifo *fifo = &dev_priv->fifos[channel];
+	uint32_t tmp, vramsz;
+
 	DRM_DEBUG("nv10_graph_context_create %d\n", channel);
 
-	memset(dev_priv->fifos[channel].pgraph_ctx, 0, sizeof(dev_priv->fifos[channel].pgraph_ctx));
+	memset(fifo->pgraph_ctx, 0, sizeof(fifo->pgraph_ctx));
 
-	//dev_priv->fifos[channel].pgraph_ctx_user = channel << 24;
-	dev_priv->fifos[channel].pgraph_ctx[0] = 0x0001ffff;
+	/* per channel init from ddx */
+	tmp = NV_READ(NV10_PGRAPH_SURFACE) & 0x0007ff00;
+	/*XXX the original ddx code, does this in 2 steps :
+	 * tmp = NV_READ(NV10_PGRAPH_SURFACE) & 0x0007ff00;
+	 * NV_WRITE(NV10_PGRAPH_SURFACE, tmp);
+	 * tmp = NV_READ(NV10_PGRAPH_SURFACE) | 0x00020100;
+	 * NV_WRITE(NV10_PGRAPH_SURFACE, tmp);
+	 */
+	tmp |= 0x00020100;
+	NV_WRITE_CTX(NV10_PGRAPH_SURFACE, tmp);
+
+	vramsz = drm_get_resource_len(dev, 0) - 1;
+	NV_WRITE_CTX(NV04_PGRAPH_BOFFSET0, 0);
+	NV_WRITE_CTX(NV04_PGRAPH_BOFFSET1, 0);
+	NV_WRITE_CTX(NV04_PGRAPH_BLIMIT0 , vramsz);
+	NV_WRITE_CTX(NV04_PGRAPH_BLIMIT1 , vramsz);
+
+	NV_WRITE_CTX(NV04_PGRAPH_PATTERN_SHAPE, 0x00000000);
+	NV_WRITE_CTX(NV04_PGRAPH_BETA_AND     , 0xFFFFFFFF);
+
+
+	NV_WRITE_CTX(NV03_PGRAPH_XY_LOGIC_MISC0, 0x0001ffff);
 	/* is it really needed ??? */
 	if (dev_priv->chipset>=0x17) {
-		dev_priv->fifos[channel].pgraph_ctx[sizeof(nv10_graph_ctx_regs) + 0] = NV_READ(NV10_PGRAPH_DEBUG_4);
-		dev_priv->fifos[channel].pgraph_ctx[sizeof(nv10_graph_ctx_regs) + 1] = NV_READ(0x004006b0);
+		NV_WRITE_CTX(NV10_PGRAPH_DEBUG_4, NV_READ(NV10_PGRAPH_DEBUG_4));
+		NV_WRITE_CTX(0x004006b0, NV_READ(0x004006b0));
 	}
 
+	/* for the first channel init the regs */
+	if (dev_priv->fifo_alloc_count == 0)
+		restore_ctx_regs(dev, channel);
+
 
 	//XXX should be saved/restored for each fifo
 	//we supposed here we have X fifo and only one 3D fifo.
@@ -606,7 +662,6 @@ int nv10_graph_context_create(drm_device_t *dev, int channel) {
 
 int nv10_graph_init(drm_device_t *dev) {
 	drm_nouveau_private_t *dev_priv = dev->dev_private;
-	uint32_t tmp, vramsz;
 	int i;
 
 	NV_WRITE(NV03_PMC_ENABLE, NV_READ(NV03_PMC_ENABLE) &
@@ -637,23 +692,7 @@ int nv10_graph_init(drm_device_t *dev) {
 	NV_WRITE(NV10_PGRAPH_STATE      , 0xFFFFFFFF);
 	NV_WRITE(NV04_PGRAPH_FIFO       , 0x00000001);
 
-	/* the below don't belong here, per-channel context state */
-	tmp = NV_READ(NV10_PGRAPH_SURFACE) & 0x0007ff00;
-	NV_WRITE(NV10_PGRAPH_SURFACE, tmp);
-	tmp = NV_READ(NV10_PGRAPH_SURFACE) | 0x00020100;
-	NV_WRITE(NV10_PGRAPH_SURFACE, tmp);
-
-	vramsz = drm_get_resource_len(dev, 0) - 1;
-	NV_WRITE(NV04_PGRAPH_BOFFSET0, 0);
-	NV_WRITE(NV04_PGRAPH_BOFFSET1, 0);
-	NV_WRITE(NV04_PGRAPH_BLIMIT0 , vramsz);
-	NV_WRITE(NV04_PGRAPH_BLIMIT1 , vramsz);
-
-	NV_WRITE(NV04_PGRAPH_PATTERN_SHAPE, 0x00000000);
-	NV_WRITE(NV04_PGRAPH_BETA_AND     , 0xFFFFFFFF);
-
 	return 0;
-
 }
 
 void nv10_graph_takedown(drm_device_t *dev)
diff --git a/shared-core/nv40_graph.c b/shared-core/nv40_graph.c
index 5d5d0dc7..792734ed 100644
--- a/shared-core/nv40_graph.c
+++ b/shared-core/nv40_graph.c
@@ -10,6 +10,7 @@
  */
 #define NV40_GRCTX_SIZE (175*1024)
 #define NV43_GRCTX_SIZE (70*1024)
+#define NV46_GRCTX_SIZE (70*1024) /* probably ~64KiB */
 #define NV4A_GRCTX_SIZE (64*1024)
 #define NV4C_GRCTX_SIZE (25*1024)
 #define NV4E_GRCTX_SIZE (25*1024)
@@ -274,6 +275,156 @@ nv43_graph_context_init(drm_device_t *dev, struct mem_block *ctx)
 		INSTANCE_WR(ctx, i/4, 0x3f800000);
 };
 
+static void nv46_graph_context_init(drm_device_t *dev, struct mem_block *ctx)
+{
+	drm_nouveau_private_t *dev_priv = dev->dev_private;
+	int i;
+
+	INSTANCE_WR(ctx, 0x00000/4, nouveau_chip_instance_get(dev, ctx));
+	INSTANCE_WR(ctx, 0x00040/4, 0x0000ffff);
+	INSTANCE_WR(ctx, 0x00044/4, 0x0000ffff);
+	INSTANCE_WR(ctx, 0x0004c/4, 0x00000001);
+	INSTANCE_WR(ctx, 0x00138/4, 0x20010001);
+	INSTANCE_WR(ctx, 0x0013c/4, 0x0f73ef00);
+	INSTANCE_WR(ctx, 0x00144/4, 0x02008821);
+	INSTANCE_WR(ctx, 0x00174/4, 0x00000001);
+	INSTANCE_WR(ctx, 0x00178/4, 0x00000001);
+	INSTANCE_WR(ctx, 0x0017c/4, 0x00000001);
+	INSTANCE_WR(ctx, 0x00180/4, 0x00000001);
+	INSTANCE_WR(ctx, 0x00184/4, 0x00000001);
+	INSTANCE_WR(ctx, 0x00188/4, 0x00000001);
+	INSTANCE_WR(ctx, 0x0018c/4, 0x00000001);
+	INSTANCE_WR(ctx, 0x00190/4, 0x00000001);
+	INSTANCE_WR(ctx, 0x00194/4, 0x00000040);
+	INSTANCE_WR(ctx, 0x00198/4, 0x00000040);
+	INSTANCE_WR(ctx, 0x0019c/4, 0x00000040);
+	INSTANCE_WR(ctx, 0x001a4/4, 0x00000040);
+	INSTANCE_WR(ctx, 0x001ec/4, 0x0b0b0b0c);
+	INSTANCE_WR(ctx, 0x0035c/4, 0x00040000);
+	INSTANCE_WR(ctx, 0x0036c/4, 0x55555555);
+	INSTANCE_WR(ctx, 0x00370/4, 0x55555555);
+	INSTANCE_WR(ctx, 0x00374/4, 0x55555555);
+	INSTANCE_WR(ctx, 0x00378/4, 0x55555555);
+	INSTANCE_WR(ctx, 0x003a4/4, 0x00000008);
+	INSTANCE_WR(ctx, 0x003b8/4, 0x00003010);
+	INSTANCE_WR(ctx, 0x003dc/4, 0x00000111);
+	INSTANCE_WR(ctx, 0x003e0/4, 0x00000111);
+	INSTANCE_WR(ctx, 0x003e4/4, 0x00000111);
+	INSTANCE_WR(ctx, 0x003e8/4, 0x00000111);
+	INSTANCE_WR(ctx, 0x003ec/4, 0x00000111);
+	INSTANCE_WR(ctx, 0x003f0/4, 0x00000111);
+	INSTANCE_WR(ctx, 0x003f4/4, 0x00000111);
+	INSTANCE_WR(ctx, 0x003f8/4, 0x00000111);
+	INSTANCE_WR(ctx, 0x003fc/4, 0x00000111);
+	INSTANCE_WR(ctx, 0x00400/4, 0x00000111);
+	INSTANCE_WR(ctx, 0x00404/4, 0x00000111);
+	INSTANCE_WR(ctx, 0x00408/4, 0x00000111);
+	INSTANCE_WR(ctx, 0x0040c/4, 0x00000111);
+	INSTANCE_WR(ctx, 0x00410/4, 0x00000111);
+	INSTANCE_WR(ctx, 0x00414/4, 0x00000111);
+	INSTANCE_WR(ctx, 0x00418/4, 0x00000111);
+	INSTANCE_WR(ctx, 0x004b0/4, 0x00000111);
+	INSTANCE_WR(ctx, 0x004b4/4, 0x00080060);
+	INSTANCE_WR(ctx, 0x004d0/4, 0x00000080);
+	INSTANCE_WR(ctx, 0x004d4/4, 0xffff0000);
+	INSTANCE_WR(ctx, 0x004d8/4, 0x00000001);
+	INSTANCE_WR(ctx, 0x004ec/4, 0x46400000);
+	INSTANCE_WR(ctx, 0x004fc/4, 0xffff0000);
+	INSTANCE_WR(ctx, 0x00500/4, 0x88888888);
+	INSTANCE_WR(ctx, 0x00504/4, 0x88888888);
+	INSTANCE_WR(ctx, 0x00508/4, 0x88888888);
+	INSTANCE_WR(ctx, 0x0050c/4, 0x88888888);
+	INSTANCE_WR(ctx, 0x00510/4, 0x88888888);
+	INSTANCE_WR(ctx, 0x00514/4, 0x88888888);
+	INSTANCE_WR(ctx, 0x00518/4, 0x88888888);
+	INSTANCE_WR(ctx, 0x0051c/4, 0x88888888);
+	INSTANCE_WR(ctx, 0x00520/4, 0x88888888);
+	INSTANCE_WR(ctx, 0x00524/4, 0x88888888);
+	INSTANCE_WR(ctx, 0x00528/4, 0x88888888);
+	INSTANCE_WR(ctx, 0x0052c/4, 0x88888888);
+	INSTANCE_WR(ctx, 0x00530/4, 0x88888888);
+	INSTANCE_WR(ctx, 0x00534/4, 0x88888888);
+	INSTANCE_WR(ctx, 0x00538/4, 0x88888888);
+	INSTANCE_WR(ctx, 0x0053c/4, 0x88888888);
+	INSTANCE_WR(ctx, 0x00550/4, 0x0fff0000);
+	INSTANCE_WR(ctx, 0x00554/4, 0x0fff0000);
+	INSTANCE_WR(ctx, 0x0055c/4, 0x00011100);
+	for (i=0x00578; i<0x005b4; i+=4)
+		INSTANCE_WR(ctx, i/4, 0x07ff0000);
+	INSTANCE_WR(ctx, 0x005c0/4, 0x4b7fffff);
+	INSTANCE_WR(ctx, 0x005e8/4, 0x30201000);
+	INSTANCE_WR(ctx, 0x005ec/4, 0x70605040);
+	INSTANCE_WR(ctx, 0x005f0/4, 0xb8a89888);
+	INSTANCE_WR(ctx, 0x005f4/4, 0xf8e8d8c8);
+	INSTANCE_WR(ctx, 0x00608/4, 0x40100000);
+	INSTANCE_WR(ctx, 0x00624/4, 0x0000ffff);
+	INSTANCE_WR(ctx, 0x00658/4, 0x435185d6);
+	INSTANCE_WR(ctx, 0x0065c/4, 0x2155b699);
+	INSTANCE_WR(ctx, 0x00660/4, 0xfedcba98);
+	INSTANCE_WR(ctx, 0x00664/4, 0x00000098);
+	INSTANCE_WR(ctx, 0x00674/4, 0xffffffff);
+	INSTANCE_WR(ctx, 0x00678/4, 0x00ff7000);
+	INSTANCE_WR(ctx, 0x0067c/4, 0x0000ffff);
+	INSTANCE_WR(ctx, 0x0068c/4, 0x00ff0000);
+	INSTANCE_WR(ctx, 0x006c8/4, 0x00ffff00);
+	for (i=0x0070c; i<=0x00748; i+=4)
+		INSTANCE_WR(ctx, i/4, 0x00018488);
+	for (i=0x0074c; i<=0x00788; i+=4)
+		INSTANCE_WR(ctx, i/4, 0x00028202);
+	for (i=0x007cc; i<=0x00808; i+=4)
+		INSTANCE_WR(ctx, i/4, 0x0000aae4);
+	for (i=0x0080c; i<=0x00848; i+=4)
+		INSTANCE_WR(ctx, i/4, 0x01012000);
+	for (i=0x0084c; i<=0x00888; i+=4)
+		INSTANCE_WR(ctx, i/4, 0x00080008);
+	for (i=0x008cc; i<=0x00908; i+=4)
+		INSTANCE_WR(ctx, i/4, 0x00100008);
+	for (i=0x0095c; i<=0x00968; i+=4)
+		INSTANCE_WR(ctx, i/4, 0x0001bc80);
+	for (i=0x0096c; i<=0x00978; i+=4)
+		INSTANCE_WR(ctx, i/4, 0x00000202);
+	for (i=0x0098c; i<=0x00998; i+=4)
+		INSTANCE_WR(ctx, i/4, 0x00000008);
+	for (i=0x009ac; i<=0x009b8; i+=4)
+		INSTANCE_WR(ctx, i/4, 0x00080008);
+	INSTANCE_WR(ctx, 0x009cc/4, 0x00000002);
+	INSTANCE_WR(ctx, 0x00a00/4, 0x00000421);
+	INSTANCE_WR(ctx, 0x00a04/4, 0x030c30c3);
+	INSTANCE_WR(ctx, 0x00a08/4, 0x00011001);
+	INSTANCE_WR(ctx, 0x00a14/4, 0x3e020200);
+	INSTANCE_WR(ctx, 0x00a18/4, 0x00ffffff);
+	INSTANCE_WR(ctx, 0x00a1c/4, 0x0c103f00);
+	INSTANCE_WR(ctx, 0x00a28/4, 0x00040000);
+	INSTANCE_WR(ctx, 0x00a60/4, 0x00008100);
+	INSTANCE_WR(ctx, 0x00aec/4, 0x00000001);
+	INSTANCE_WR(ctx, 0x00b30/4, 0x00001001);
+	INSTANCE_WR(ctx, 0x00b38/4, 0x00000003);
+	INSTANCE_WR(ctx, 0x00b3c/4, 0x00888001);
+	INSTANCE_WR(ctx, 0x00bc0/4, 0x00000005);
+	INSTANCE_WR(ctx, 0x00bcc/4, 0x0000ffff);
+	INSTANCE_WR(ctx, 0x00be8/4, 0x00005555);
+	INSTANCE_WR(ctx, 0x00bec/4, 0x00005555);
+	INSTANCE_WR(ctx, 0x00bf0/4, 0x00005555);
+	INSTANCE_WR(ctx, 0x00bf4/4, 0x00000001);
+	INSTANCE_WR(ctx, 0x00c2c/4, 0x00000001);
+	INSTANCE_WR(ctx, 0x00c30/4, 0x08e00001);
+	INSTANCE_WR(ctx, 0x00c34/4, 0x000e3000);
+	for (i=0x017f8; i<=0x01870; i+=8)
+		INSTANCE_WR(ctx, i/4, 0x3f800000);
+	for (i=0x035b8; i<=0x057a8; i+=24)
+		INSTANCE_WR(ctx, i/4, 0x00000001);
+	for (i=0x057b8; i<=0x05ba8; i+=16)
+		INSTANCE_WR(ctx, i/4, 0x3f800000);
+	for (i=0x07f38; i<=0x0a128; i+=24)
+		INSTANCE_WR(ctx, i/4, 0x00000001);
+	for (i=0x0a138; i<=0x0a528; i+=16)
+		INSTANCE_WR(ctx, i/4, 0x3f800000);
+	for (i=0x0c8b8; i<=0x0eaa8; i+=24)
+		INSTANCE_WR(ctx, i/4, 0x00000001);
+	for (i=0x0eab8; i<=0x0eea8; i+=16)
+		INSTANCE_WR(ctx, i/4, 0x3f800000);
+}
+
 static void nv4a_graph_context_init(drm_device_t *dev, struct mem_block *ctx)
 {
 	drm_nouveau_private_t *dev_priv = dev->dev_private;
@@ -622,6 +773,10 @@ nv40_graph_context_create(drm_device_t *dev, int channel)
 		ctx_size = NV43_GRCTX_SIZE;
 		ctx_init = nv43_graph_context_init;
 		break;
+	case 0x46:
+		ctx_size = NV46_GRCTX_SIZE;
+		ctx_init = nv46_graph_context_init;
+		break;
 	case 0x4a:
 		ctx_size = NV4A_GRCTX_SIZE;
 		ctx_init = nv4a_graph_context_init;
@@ -821,6 +976,37 @@ static uint32_t nv43_ctx_voodoo[] = {
 	~0
 };
 
+static uint32_t nv46_ctx_voodoo[] = {
+	0x00400889, 0x00200000, 0x0060000a, 0x00200000, 0x00300000, 0x00800001,
+	0x00700009, 0x0060000e, 0x00400d64, 0x00400d05, 0x00408f65, 0x00409306,
+	0x0040a068, 0x0040198f, 0x00200001, 0x0060000a, 0x00700080, 0x00104042,
+	0x00200001, 0x0060000a, 0x00700000, 0x001040c5, 0x00401826, 0x00401968,
+	0x0060000d, 0x00200000, 0x0060000a, 0x00700000, 0x00106000, 0x00700080,
+	0x004020e6, 0x007000a0, 0x00500060, 0x00200008, 0x0060000a, 0x0011814d,
+	0x00110158, 0x00105401, 0x0020003a, 0x00100051, 0x001040c5, 0x0010c1c4,
+	0x001041c9, 0x0010c1dc, 0x00150210, 0x0012c225, 0x00108238, 0x0010823e,
+	0x001242c0, 0x00200040, 0x00100280, 0x00128100, 0x00128120, 0x00128143,
+	0x0011415f, 0x0010815c, 0x0010c140, 0x00104029, 0x00110400, 0x00104d10,
+	0x00500060, 0x00403f87, 0x0060000d, 0x004079e6, 0x002000f7, 0x0060000a,
+	0x00200045, 0x00100620, 0x00104668, 0x0017466d, 0x0011068b, 0x00168691,
+	0x001046ae, 0x001046b0, 0x001206b4, 0x001046c4, 0x001146c6, 0x00200022,
+	0x001006cc, 0x001246f0, 0x002000c0, 0x00100700, 0x0010c3d7, 0x001043e1,
+	0x00500060, 0x0020027f, 0x0060000a, 0x00104800, 0x00108901, 0x00104910,
+	0x00124920, 0x0020001f, 0x00100940, 0x00140965, 0x00148a00, 0x00108a14,
+	0x00160b00, 0x00134b2c, 0x0010cd00, 0x0010cd04, 0x0010cd08, 0x00104d80,
+	0x00104e00, 0x0012d600, 0x00105c00, 0x00104f06, 0x00105406, 0x00105709,
+	0x00200316, 0x0060000a, 0x00300000, 0x00200080, 0x00407200, 0x00200084,
+	0x00800001, 0x0020055e, 0x0060000a, 0x002037e0, 0x0040788a, 0x00201320,
+	0x00800029, 0x00408900, 0x00600006, 0x004085e6, 0x00700080, 0x00200081,
+	0x0060000a, 0x00104280, 0x00200316, 0x0060000a, 0x00200004, 0x00800001,
+	0x00700000, 0x00200000, 0x0060000a, 0x00106002, 0x0040a068, 0x00700000,
+	0x00200000, 0x0060000a, 0x00106002, 0x00700080, 0x00400a68, 0x00500060,
+	0x00600007, 0x00409388, 0x0060000f, 0x00500060, 0x00200000, 0x0060000a,
+	0x00700000, 0x00106001, 0x00910880, 0x00901ffe, 0x01940000, 0x00200020,
+	0x0060000b, 0x00500069, 0x0060000c, 0x00402168, 0x0040a206, 0x0040a305,
+	0x00600009, 0x00700005, 0x00700006, 0x0060000e, ~0
+};
+
 static uint32_t nv4a_ctx_voodoo[] = {
 	0x00400889, 0x00200000, 0x0060000a, 0x00200000, 0x00300000, 0x00800001, 
 	0x00700009, 0x0060000e, 0x00400d64, 0x00400d05, 0x00409965, 0x00409e06, 
@@ -891,7 +1077,7 @@ static uint32_t nv4e_ctx_voodoo[] = {
  * G70		0x47
  * G71		0x49
  * NV45		0x48
- * G72		0x46
+ * G72[M]	0x46
  * G73		0x4b
  * C51_G7X	0x4c
  * C51		0x4e
@@ -913,6 +1099,7 @@ nv40_graph_init(drm_device_t *dev)
 	switch (dev_priv->chipset) {
 	case 0x40: ctx_voodoo = nv40_ctx_voodoo; break;
 	case 0x43: ctx_voodoo = nv43_ctx_voodoo; break;
+	case 0x46: ctx_voodoo = nv46_ctx_voodoo; break;
 	case 0x4a: ctx_voodoo = nv4a_ctx_voodoo; break;
 	case 0x4e: ctx_voodoo = nv4e_ctx_voodoo; break;
 	default:
diff --git a/shared-core/nv40_mc.c b/shared-core/nv40_mc.c
index 554a2241..8dbd96fd 100644
--- a/shared-core/nv40_mc.c
+++ b/shared-core/nv40_mc.c
@@ -9,6 +9,11 @@ nv40_mc_init(drm_device_t *dev)
 	drm_nouveau_private_t *dev_priv = dev->dev_private;
 	uint32_t tmp;
 
+	/* Power up everything, resetting each individual unit will
+	 * be done later if needed.
+	 */
+	NV_WRITE(NV03_PMC_ENABLE, 0xFFFFFFFF);
+
 	NV_WRITE(NV03_PMC_INTR_EN_0, 0);
 
 	switch (dev_priv->chipset) {
diff --git a/shared-core/r128_cce.c b/shared-core/r128_cce.c
index 62859d5a..fc7bb8fc 100644
--- a/shared-core/r128_cce.c
+++ b/shared-core/r128_cce.c
@@ -563,7 +563,7 @@ static int r128_do_init_cce(drm_device_t * dev, drm_r128_init_t * init)
 		dev_priv->gart_info.table_size = R128_PCIGART_TABLE_SIZE;
 		dev_priv->gart_info.addr = NULL;
 		dev_priv->gart_info.bus_addr = 0;
-		dev_priv->gart_info.is_pcie = 0;
+		dev_priv->gart_info.gart_reg_if = DRM_ATI_GART_PCI;
 		if (!drm_ati_pcigart_init(dev, &dev_priv->gart_info)) {
 			DRM_ERROR("failed to init PCI GART!\n");
 			dev->dev_private = (void *)dev_priv;
diff --git a/shared-core/radeon_cp.c b/shared-core/radeon_cp.c
index e02796e7..bfc2d088 100644
--- a/shared-core/radeon_cp.c
+++ b/shared-core/radeon_cp.c
@@ -824,12 +824,21 @@ static int RADEON_READ_PLL(drm_device_t * dev, int addr)
 	return RADEON_READ(RADEON_CLOCK_CNTL_DATA);
 }
 
-static int RADEON_READ_PCIE(drm_radeon_private_t *dev_priv, int addr)
+static u32 RADEON_READ_PCIE(drm_radeon_private_t *dev_priv, int addr)
 {
 	RADEON_WRITE8(RADEON_PCIE_INDEX, addr & 0xff);
 	return RADEON_READ(RADEON_PCIE_DATA);
 }
 
+static u32 RADEON_READ_IGPGART(drm_radeon_private_t *dev_priv, int addr)
+{
+	u32 ret;
+	RADEON_WRITE(RADEON_IGPGART_INDEX, addr & 0x7f);
+	ret = RADEON_READ(RADEON_IGPGART_DATA);
+	RADEON_WRITE(RADEON_IGPGART_INDEX, 0x7f);
+	return ret;
+}
+
 #if RADEON_FIFO_DEBUG
 static void radeon_status(drm_radeon_private_t * dev_priv)
 {
@@ -1266,7 +1275,45 @@ static void radeon_test_writeback(drm_radeon_private_t * dev_priv)
 	}
 }
 
-/* Enable or disable PCI-E GART on the chip */
+/* Enable or disable IGP GART on the chip */
+static void radeon_set_igpgart(drm_radeon_private_t * dev_priv, int on)
+{
+	u32 temp, tmp;
+
+	tmp = RADEON_READ(RADEON_AIC_CNTL);
+	DRM_DEBUG("setting igpgart AIC CNTL is %08X\n", tmp);
+	if (on) {
+		DRM_DEBUG("programming igpgart %08X %08lX %08X\n",
+			 dev_priv->gart_vm_start,
+			 (long)dev_priv->gart_info.bus_addr,
+			 dev_priv->gart_size);
+
+		RADEON_WRITE_IGPGART(RADEON_IGPGART_UNK_18, 0x1000);
+		RADEON_WRITE_IGPGART(RADEON_IGPGART_ENABLE, 0x1);
+		RADEON_WRITE_IGPGART(RADEON_IGPGART_CTRL, 0x42040800);
+		RADEON_WRITE_IGPGART(RADEON_IGPGART_BASE_ADDR,
+				     dev_priv->gart_info.bus_addr);
+
+		temp = RADEON_READ_IGPGART(dev_priv, RADEON_IGPGART_UNK_39);
+		RADEON_WRITE_IGPGART(RADEON_IGPGART_UNK_39, temp);
+
+		RADEON_WRITE(RADEON_AGP_BASE, (unsigned int)dev_priv->gart_vm_start);
+		dev_priv->gart_size = 32*1024*1024;
+		RADEON_WRITE(RADEON_MC_AGP_LOCATION,
+			     (((dev_priv->gart_vm_start - 1 +
+			       dev_priv->gart_size) & 0xffff0000) |
+			     (dev_priv->gart_vm_start >> 16)));
+
+		temp = RADEON_READ_IGPGART(dev_priv, RADEON_IGPGART_ENABLE);
+		RADEON_WRITE_IGPGART(RADEON_IGPGART_ENABLE, temp);
+
+		RADEON_READ_IGPGART(dev_priv, RADEON_IGPGART_FLUSH);
+		RADEON_WRITE_IGPGART(RADEON_IGPGART_FLUSH, 0x1);
+		RADEON_READ_IGPGART(dev_priv, RADEON_IGPGART_FLUSH);
+		RADEON_WRITE_IGPGART(RADEON_IGPGART_FLUSH, 0x0);
+       }
+}
+
 static void radeon_set_pciegart(drm_radeon_private_t * dev_priv, int on)
 {
 	u32 tmp = RADEON_READ_PCIE(dev_priv, RADEON_PCIE_TX_GART_CNTL);
@@ -1301,6 +1348,11 @@ static void radeon_set_pcigart(drm_radeon_private_t * dev_priv, int on)
 {
 	u32 tmp;
 
+	if (dev_priv->flags & RADEON_IS_IGPGART) {
+		radeon_set_igpgart(dev_priv, on);
+		return;
+	}
+
 	if (dev_priv->flags & RADEON_IS_PCIE) {
 		radeon_set_pciegart(dev_priv, on);
 		return;
@@ -1635,8 +1687,10 @@ static int radeon_do_init_cp(drm_device_t * dev, drm_radeon_init_t * init)
 			dev_priv->gart_info.addr =
 			    dev_priv->gart_info.mapping.handle;
 
-			dev_priv->gart_info.is_pcie =
-			    !!(dev_priv->flags & RADEON_IS_PCIE);
+			if (dev_priv->flags & RADEON_IS_PCIE)
+				dev_priv->gart_info.gart_reg_if = DRM_ATI_GART_PCIE;
+			else
+				dev_priv->gart_info.gart_reg_if = DRM_ATI_GART_PCI;
 			dev_priv->gart_info.gart_table_location =
 			    DRM_ATI_GART_FB;
 
@@ -1644,6 +1698,10 @@ static int radeon_do_init_cp(drm_device_t * dev, drm_radeon_init_t * init)
 				  dev_priv->gart_info.addr,
 				  dev_priv->pcigart_offset);
 		} else {
+			if (dev_priv->flags & RADEON_IS_IGPGART)
+				dev_priv->gart_info.gart_reg_if = DRM_ATI_GART_IGP;
+			else
+				dev_priv->gart_info.gart_reg_if = DRM_ATI_GART_PCI;
 			dev_priv->gart_info.gart_table_location =
 			    DRM_ATI_GART_MAIN;
 			dev_priv->gart_info.addr = NULL;
diff --git a/shared-core/radeon_drv.h b/shared-core/radeon_drv.h
index 3e56af30..00646c8f 100644
--- a/shared-core/radeon_drv.h
+++ b/shared-core/radeon_drv.h
@@ -96,10 +96,11 @@
  * 1.25- Add support for r200 vertex programs (R200_EMIT_VAP_PVS_CNTL,
  *       new packet type)
  * 1.26- Add support for variable size PCI(E) gart aperture
+ * 1.27- Add support for IGP GART
  */
 
 #define DRIVER_MAJOR		1
-#define DRIVER_MINOR		26
+#define DRIVER_MINOR		27
 #define DRIVER_PATCHLEVEL	0
 
 /*
@@ -145,6 +146,7 @@ enum radeon_chip_flags {
 	RADEON_IS_PCIE = 0x00200000UL,
 	RADEON_NEW_MEMMAP = 0x00400000UL,
 	RADEON_IS_PCI = 0x00800000UL,
+	RADEON_IS_IGPGART = 0x01000000UL,
 };
 
 #define GET_RING_HEAD(dev_priv)	(dev_priv->writeback_works ? \
@@ -432,6 +434,16 @@ extern int r300_do_cp_cmdbuf(drm_device_t *dev, DRMFILE filp,
 #define RADEON_PCIE_TX_GART_END_LO	0x16
 #define RADEON_PCIE_TX_GART_END_HI	0x17
 
+#define RADEON_IGPGART_INDEX            0x168
+#define RADEON_IGPGART_DATA             0x16c
+#define RADEON_IGPGART_UNK_18           0x18
+#define RADEON_IGPGART_CTRL             0x2b
+#define RADEON_IGPGART_BASE_ADDR        0x2c
+#define RADEON_IGPGART_FLUSH            0x2e
+#define RADEON_IGPGART_ENABLE           0x38
+#define RADEON_IGPGART_UNK_39           0x39
+
+
 #define RADEON_MPP_TB_CONFIG		0x01c0
 #define RADEON_MEM_CNTL			0x0140
 #define RADEON_MEM_SDRAM_MODE_REG	0x0158
@@ -994,6 +1006,14 @@ do {									\
 	RADEON_WRITE( RADEON_CLOCK_CNTL_DATA, (val) );			\
 } while (0)
 
+#define RADEON_WRITE_IGPGART( addr, val )				\
+do {									\
+	RADEON_WRITE( RADEON_IGPGART_INDEX,				\
+			((addr) & 0x7f) | (1 << 8));			\
+	RADEON_WRITE( RADEON_IGPGART_DATA, (val) );			\
+	RADEON_WRITE( RADEON_IGPGART_INDEX, 0x7f );			\
+} while (0)
+
 #define RADEON_WRITE_PCIE( addr, val )					\
 do {									\
 	RADEON_WRITE8( RADEON_PCIE_INDEX,				\