drm/nvc0: implement pgraph engine hooks

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>

1 files changed, 708 insertions, 1 deletions

diff --git a/drivers/gpu/drm/nouveau/nvc0_graph.c b/drivers/gpu/drm/nouveau/nvc0_graph.c
index 717a517..cf2f6aa 100644
--- a/drivers/gpu/drm/nouveau/nvc0_graph.c
+++ b/drivers/gpu/drm/nouveau/nvc0_graph.c
@@ -22,9 +22,16 @@
  * Authors: Ben Skeggs
  */
 
+#include <linux/firmware.h>
+
 #include "drmP.h"
 
 #include "nouveau_drv.h"
+#include "nouveau_mm.h"
+#include "nvc0_graph.h"
+
+static void nvc0_graph_isr(struct drm_device *);
+static int  nvc0_graph_unload_context_to(struct drm_device *dev, u64 chan);
 
 void
 nvc0_graph_fifo_access(struct drm_device *dev, bool enabled)
@@ -37,39 +44,739 @@ nvc0_graph_channel(struct drm_device *dev)
 	return NULL;
 }
 
+static int
+nvc0_graph_construct_context(struct nouveau_channel *chan)
+{
+	struct drm_nouveau_private *dev_priv = chan->dev->dev_private;
+	struct nvc0_graph_priv *priv = dev_priv->engine.graph.priv;
+	struct nvc0_graph_chan *grch = chan->pgraph_ctx;
+	struct drm_device *dev = chan->dev;
+	int ret, i;
+	u32 *ctx;
+
+	ctx = kmalloc(priv->grctx_size, GFP_KERNEL);
+	if (!ctx)
+		return -ENOMEM;
+
+	nvc0_graph_load_context(chan);
+
+	nv_wo32(grch->grctx, 0x1c, 1);
+	nv_wo32(grch->grctx, 0x20, 0);
+	nv_wo32(grch->grctx, 0x28, 0);
+	nv_wo32(grch->grctx, 0x2c, 0);
+	dev_priv->engine.instmem.flush(dev);
+
+	ret = nvc0_grctx_generate(chan);
+	if (ret) {
+		kfree(ctx);
+		return ret;
+	}
+
+	ret = nvc0_graph_unload_context_to(dev, chan->ramin->vinst);
+	if (ret) {
+		kfree(ctx);
+		return ret;
+	}
+
+	for (i = 0; i < priv->grctx_size; i += 4)
+		ctx[i / 4] = nv_ro32(grch->grctx, i);
+
+	priv->grctx_vals = ctx;
+	return 0;
+}
+
+static int
+nvc0_graph_create_context_mmio_list(struct nouveau_channel *chan)
+{
+	struct drm_nouveau_private *dev_priv = chan->dev->dev_private;
+	struct nvc0_graph_priv *priv = dev_priv->engine.graph.priv;
+	struct nvc0_graph_chan *grch = chan->pgraph_ctx;
+	struct drm_device *dev = chan->dev;
+	int i = 0, gpc, tp, ret;
+	u32 magic;
+
+	ret = nouveau_gpuobj_new(dev, NULL, 0x2000, 256, NVOBJ_FLAG_VM,
+				 &grch->unk408004);
+	if (ret)
+		return ret;
+
+	ret = nouveau_gpuobj_new(dev, NULL, 0x8000, 256, NVOBJ_FLAG_VM,
+				 &grch->unk40800c);
+	if (ret)
+		return ret;
+
+	ret = nouveau_gpuobj_new(dev, NULL, 384 * 1024, 4096, NVOBJ_FLAG_VM,
+				 &grch->unk418810);
+	if (ret)
+		return ret;
+
+	ret = nouveau_gpuobj_new(dev, NULL, 0x1000, 0, NVOBJ_FLAG_VM,
+				 &grch->mmio);
+	if (ret)
+		return ret;
+
+
+	nv_wo32(grch->mmio, i++ * 4, 0x00408004);
+	nv_wo32(grch->mmio, i++ * 4, grch->unk408004->vinst >> 8);
+	nv_wo32(grch->mmio, i++ * 4, 0x00408008);
+	nv_wo32(grch->mmio, i++ * 4, 0x80000018);
+
+	nv_wo32(grch->mmio, i++ * 4, 0x0040800c);
+	nv_wo32(grch->mmio, i++ * 4, grch->unk40800c->vinst >> 8);
+	nv_wo32(grch->mmio, i++ * 4, 0x00408010);
+	nv_wo32(grch->mmio, i++ * 4, 0x80000000);
+
+	nv_wo32(grch->mmio, i++ * 4, 0x00418810);
+	nv_wo32(grch->mmio, i++ * 4, 0x80000000 | grch->unk418810->vinst >> 12);
+	nv_wo32(grch->mmio, i++ * 4, 0x00419848);
+	nv_wo32(grch->mmio, i++ * 4, 0x10000000 | grch->unk418810->vinst >> 12);
+
+	nv_wo32(grch->mmio, i++ * 4, 0x00419004);
+	nv_wo32(grch->mmio, i++ * 4, grch->unk40800c->vinst >> 8);
+	nv_wo32(grch->mmio, i++ * 4, 0x00419008);
+	nv_wo32(grch->mmio, i++ * 4, 0x00000000);
+
+	nv_wo32(grch->mmio, i++ * 4, 0x00418808);
+	nv_wo32(grch->mmio, i++ * 4, grch->unk408004->vinst >> 8);
+	nv_wo32(grch->mmio, i++ * 4, 0x0041880c);
+	nv_wo32(grch->mmio, i++ * 4, 0x80000018);
+
+	magic = 0x02180000;
+	nv_wo32(grch->mmio, i++ * 4, 0x00405830);
+	nv_wo32(grch->mmio, i++ * 4, magic);
+	for (gpc = 0; gpc < priv->gpc_nr; gpc++) {
+		for (tp = 0; tp < priv->tp_nr[gpc]; tp++, magic += 0x02fc) {
+			u32 reg = 0x504520 + (gpc * 0x8000) + (tp * 0x0800);
+			nv_wo32(grch->mmio, i++ * 4, reg);
+			nv_wo32(grch->mmio, i++ * 4, magic);
+		}
+	}
+
+	grch->mmio_nr = i / 2;
+	return 0;
+}
+
 int
 nvc0_graph_create_context(struct nouveau_channel *chan)
 {
+	struct drm_nouveau_private *dev_priv = chan->dev->dev_private;
+	struct nouveau_instmem_engine *pinstmem = &dev_priv->engine.instmem;
+	struct nouveau_pgraph_engine *pgraph = &dev_priv->engine.graph;
+	struct nvc0_graph_priv *priv = pgraph->priv;
+	struct nvc0_graph_chan *grch;
+	struct drm_device *dev = chan->dev;
+	struct nouveau_gpuobj *grctx;
+	int ret, i;
+
+	chan->pgraph_ctx = kzalloc(sizeof(*grch), GFP_KERNEL);
+	if (!chan->pgraph_ctx)
+		return -ENOMEM;
+	grch = chan->pgraph_ctx;
+
+	ret = nouveau_gpuobj_new(dev, NULL, priv->grctx_size, 256,
+				 NVOBJ_FLAG_VM | NVOBJ_FLAG_ZERO_ALLOC,
+				 &grch->grctx);
+	if (ret)
+		goto error;
+	chan->ramin_grctx = grch->grctx;
+	grctx = grch->grctx;
+
+	ret = nvc0_graph_create_context_mmio_list(chan);
+	if (ret)
+		goto error;
+
+	nv_wo32(chan->ramin, 0x0210, lower_32_bits(grctx->vinst) | 4);
+	nv_wo32(chan->ramin, 0x0214, upper_32_bits(grctx->vinst));
+	pinstmem->flush(dev);
+
+	if (!priv->grctx_vals) {
+		ret = nvc0_graph_construct_context(chan);
+		if (ret)
+			goto error;
+	}
+
+	for (i = 0; i < priv->grctx_size; i += 4)
+		nv_wo32(grctx, i, priv->grctx_vals[i / 4]);
+
+        nv_wo32(grctx, 0xf4, 0);
+        nv_wo32(grctx, 0xf8, 0);
+        nv_wo32(grctx, 0x10, grch->mmio_nr);
+        nv_wo32(grctx, 0x14, lower_32_bits(grch->mmio->vinst));
+        nv_wo32(grctx, 0x18, upper_32_bits(grch->mmio->vinst));
+        nv_wo32(grctx, 0x1c, 1);
+        nv_wo32(grctx, 0x20, 0);
+        nv_wo32(grctx, 0x28, 0);
+        nv_wo32(grctx, 0x2c, 0);
+	pinstmem->flush(dev);
 	return 0;
+
+error:
+	pgraph->destroy_context(chan);
+	return ret;
 }
 
 void
 nvc0_graph_destroy_context(struct nouveau_channel *chan)
 {
+	struct nvc0_graph_chan *grch;
+
+	grch = chan->pgraph_ctx;
+	chan->pgraph_ctx = NULL;
+	if (!grch)
+		return;
+
+	nouveau_gpuobj_ref(NULL, &grch->mmio);
+	nouveau_gpuobj_ref(NULL, &grch->unk418810);
+	nouveau_gpuobj_ref(NULL, &grch->unk40800c);
+	nouveau_gpuobj_ref(NULL, &grch->unk408004);
+	nouveau_gpuobj_ref(NULL, &grch->grctx);
+	chan->ramin_grctx = NULL;
 }
 
 int
 nvc0_graph_load_context(struct nouveau_channel *chan)
 {
+	struct drm_device *dev = chan->dev;
+
+	nv_wr32(dev, 0x409840, 0x00000030);
+	nv_wr32(dev, 0x409500, 0x80000000 | chan->ramin->vinst >> 12);
+	nv_wr32(dev, 0x409504, 0x00000003);
+	if (!nv_wait(dev, 0x409800, 0x00000010, 0x00000010))
+		NV_ERROR(dev, "PGRAPH: load_ctx timeout\n");
+
+	printk(KERN_ERR "load_ctx 0x%08x\n", nv_rd32(dev, 0x409b00));
+	return 0;
+}
+
+static int
+nvc0_graph_unload_context_to(struct drm_device *dev, u64 chan)
+{
+	nv_wr32(dev, 0x409840, 0x00000003);
+	nv_wr32(dev, 0x409500, 0x80000000 | chan >> 12);
+	nv_wr32(dev, 0x409504, 0x00000009);
+	if (!nv_wait(dev, 0x409800, 0x00000001, 0x00000000)) {
+		NV_ERROR(dev, "PGRAPH: unload_ctx timeout\n");
+		return -EBUSY;
+	}
+
 	return 0;
 }
 
 int
 nvc0_graph_unload_context(struct drm_device *dev)
 {
-	return 0;
+	u64 inst = (u64)(nv_rd32(dev, 0x409b00) & 0x0fffffff) << 12;
+	return nvc0_graph_unload_context_to(dev, inst);
+}
+
+static void
+nvc0_graph_destroy(struct drm_device *dev)
+{
+	struct drm_nouveau_private *dev_priv = dev->dev_private;
+	struct nouveau_pgraph_engine *pgraph = &dev_priv->engine.graph;
+	struct nvc0_graph_priv *priv;
+
+	priv = pgraph->priv;
+	if (!priv)
+		return;
+
+	nouveau_irq_unregister(dev, 12);
+
+	nouveau_gpuobj_ref(NULL, &priv->unk4188b8);
+	nouveau_gpuobj_ref(NULL, &priv->unk4188b4);
+
+	if (priv->grctx_vals)
+		kfree(priv->grctx_vals);
+	kfree(priv);
 }
 
 void
 nvc0_graph_takedown(struct drm_device *dev)
 {
+	nvc0_graph_destroy(dev);
+}
+
+static int
+nvc0_graph_create(struct drm_device *dev)
+{
+	struct drm_nouveau_private *dev_priv = dev->dev_private;
+	struct nouveau_pgraph_engine *pgraph = &dev_priv->engine.graph;
+	struct nvc0_graph_priv *priv;
+	int ret, gpc, i;
+
+	priv = kzalloc(sizeof(*priv), GFP_KERNEL);
+	if (!priv)
+		return -ENOMEM;
+	pgraph->priv = priv;
+
+	ret = nouveau_gpuobj_new(dev, NULL, 0x1000, 256, 0, &priv->unk4188b4);
+	if (ret)
+		goto error;
+
+	ret = nouveau_gpuobj_new(dev, NULL, 0x1000, 256, 0, &priv->unk4188b8);
+	if (ret)
+		goto error;
+
+	for (i = 0; i < 0x1000; i += 4) {
+		nv_wo32(priv->unk4188b4, i, 0x00000010);
+		nv_wo32(priv->unk4188b8, i, 0x00000010);
+	}
+
+	priv->gpc_nr  =  nv_rd32(dev, 0x409604) & 0x0000001f;
+	priv->rop_nr = (nv_rd32(dev, 0x409604) & 0x001f0000) >> 16;
+	for (gpc = 0; gpc < priv->gpc_nr; gpc++) {
+		priv->tp_nr[gpc] = nv_rd32(dev, GPC_UNIT(gpc, 0x2608));
+		priv->tp_total += priv->tp_nr[gpc];
+	}
+
+	/*XXX: these need figuring out... */
+	switch (dev_priv->chipset) {
+	case 0xc0:
+		if (priv->tp_total == 11) { /* 465, 3/4/4/0, 4 */
+			priv->magic_not_rop_nr = 0x07;
+			priv->magic419bd0      = 0x0a360000;
+			priv->magic419be4      = 0x04c33a54;
+			/* filled values up to tp_total, the rest 0 */
+			priv->magicgpc980[0]   = 0x22111000;
+			priv->magicgpc980[1]   = 0x00000233;
+			priv->magicgpc980[2]   = 0x00000000;
+			priv->magicgpc980[3]   = 0x00000000;
+			priv->magicgpc918      = 0x000ba2e9;
+		} else
+		if (priv->tp_total == 14) { /* 470, 3/3/4/4, 5 */
+			priv->magic_not_rop_nr = 0x05;
+			priv->magic419bd0      = 0x043c0000;
+			priv->magic419be4      = 0x09041208;
+			priv->magicgpc980[0]   = 0x11110000;
+			priv->magicgpc980[1]   = 0x00233222;
+			priv->magicgpc980[2]   = 0x00000000;
+			priv->magicgpc980[3]   = 0x00000000;
+			priv->magicgpc918      = 0x00092493;
+		} else
+		if (priv->tp_total == 15) { /* 480, 3/4/4/4, 6 */
+			priv->magic_not_rop_nr = 0x06;
+			priv->magic419bd0      = 0x023e0000;
+			priv->magic419be4      = 0x10414104;
+			priv->magicgpc980[0]   = 0x11110000;
+			priv->magicgpc980[1]   = 0x03332222;
+			priv->magicgpc980[2]   = 0x00000000;
+			priv->magicgpc980[3]   = 0x00000000;
+			priv->magicgpc918      = 0x00088889;
+		}
+		break;
+	case 0xc3: /* 450, 4/0/0/0, 2 */
+		priv->magic_not_rop_nr = 0x03;
+		priv->magic419bd0      = 0x00500000;
+		priv->magic419be4      = 0x00000000;
+		priv->magicgpc980[0]   = 0x00003210;
+		priv->magicgpc980[1]   = 0x00000000;
+		priv->magicgpc980[2]   = 0x00000000;
+		priv->magicgpc980[3]   = 0x00000000;
+		priv->magicgpc918      = 0x00200000;
+		break;
+	case 0xc4: /* 460, 3/4/0/0, 4 */
+		priv->magic_not_rop_nr = 0x01;
+		priv->magic419bd0      = 0x045c0000;
+		priv->magic419be4      = 0x09041208;
+		priv->magicgpc980[0]   = 0x02321100;
+		priv->magicgpc980[1]   = 0x00000000;
+		priv->magicgpc980[2]   = 0x00000000;
+		priv->magicgpc980[3]   = 0x00000000;
+		priv->magicgpc918      = 0x00124925;
+		break;
+	}
+
+	if (!priv->magic419bd0) {
+		NV_ERROR(dev, "PGRAPH: unknown config: %d/%d/%d/%d, %d\n",
+			 priv->tp_nr[0], priv->tp_nr[1], priv->tp_nr[2],
+			 priv->tp_nr[3], priv->rop_nr);
+		/* use 0xc3's values... */
+		priv->magic_not_rop_nr = 0x03;
+		priv->magic419bd0      = 0x00500000;
+		priv->magic419be4      = 0x00000000;
+		priv->magicgpc980[0]   = 0x00003210;
+		priv->magicgpc980[1]   = 0x00000000;
+		priv->magicgpc980[2]   = 0x00000000;
+		priv->magicgpc980[3]   = 0x00000000;
+		priv->magicgpc918      = 0x00200000;
+	}
+
+	nouveau_irq_register(dev, 12, nvc0_graph_isr);
+	NVOBJ_CLASS(dev, 0x902d, GR); /* 2D */
+	NVOBJ_CLASS(dev, 0x9039, GR); /* M2MF */
+	NVOBJ_CLASS(dev, 0x9097, GR); /* 3D */
+	NVOBJ_CLASS(dev, 0x90c0, GR); /* COMPUTE */
+	return 0;
+
+error:
+	nvc0_graph_destroy(dev);
+	return ret;
+}
+
+static void
+nvc0_graph_init_obj418880(struct drm_device *dev)
+{
+	struct drm_nouveau_private *dev_priv = dev->dev_private;
+	struct nouveau_pgraph_engine *pgraph = &dev_priv->engine.graph;
+	struct nvc0_graph_priv *priv = pgraph->priv;
+	int i;
+
+	nv_wr32(dev, GPC_BCAST(0x0880), 0x00000000);
+	nv_wr32(dev, GPC_BCAST(0x08a4), 0x00000000);
+	for (i = 0; i < 4; i++)
+		nv_wr32(dev, GPC_BCAST(0x0888) + (i * 4), 0x00000000);
+	nv_wr32(dev, GPC_BCAST(0x08b4), priv->unk4188b4->vinst >> 8);
+	nv_wr32(dev, GPC_BCAST(0x08b8), priv->unk4188b8->vinst >> 8);
+}
+
+static void
+nvc0_graph_init_regs(struct drm_device *dev)
+{
+	nv_wr32(dev, 0x400080, 0x003083c2);
+	nv_wr32(dev, 0x400088, 0x00006fe7);
+	nv_wr32(dev, 0x40008c, 0x00000000);
+	nv_wr32(dev, 0x400090, 0x00000030);
+	nv_wr32(dev, 0x40013c, 0x013901f7);
+	nv_wr32(dev, 0x400140, 0x00000100);
+	nv_wr32(dev, 0x400144, 0x00000000);
+	nv_wr32(dev, 0x400148, 0x00000110);
+	nv_wr32(dev, 0x400138, 0x00000000);
+	nv_wr32(dev, 0x400130, 0x00000000);
+	nv_wr32(dev, 0x400134, 0x00000000);
+	nv_wr32(dev, 0x400124, 0x00000002);
+}
+
+static void
+nvc0_graph_init_gpc_0(struct drm_device *dev)
+{
+	struct drm_nouveau_private *dev_priv = dev->dev_private;
+	struct nvc0_graph_priv *priv = dev_priv->engine.graph.priv;
+	int gpc;
+	
+	//      TP      ROP UNKVAL(magic_not_rop_nr)
+	// 450: 4/0/0/0 2        3
+	// 460: 3/4/0/0 4        1
+	// 465: 3/4/4/0 4        7
+	// 470: 3/3/4/4 5        5
+	// 480: 3/4/4/4 6        6
+
+	// magicgpc918
+	// 450: 00200000 00000000001000000000000000000000
+	// 460: 00124925 00000000000100100100100100100101
+	// 465: 000ba2e9 00000000000010111010001011101001
+	// 470: 00092493 00000000000010010010010010010011
+	// 480: 00088889 00000000000010001000100010001001
+
+	/* filled values up to tp_total, remainder 0 */
+	// 450: 00003210 00000000 00000000 00000000
+	// 460: 02321100 00000000 00000000 00000000
+	// 465: 22111000 00000233 00000000 00000000
+	// 470: 11110000 00233222 00000000 00000000
+	// 480: 11110000 03332222 00000000 00000000
+	
+	nv_wr32(dev, GPC_BCAST(0x0980), priv->magicgpc980[0]);
+	nv_wr32(dev, GPC_BCAST(0x0984), priv->magicgpc980[1]);
+	nv_wr32(dev, GPC_BCAST(0x0988), priv->magicgpc980[2]);
+	nv_wr32(dev, GPC_BCAST(0x098c), priv->magicgpc980[3]);
+
+	for (gpc = 0; gpc < priv->gpc_nr; gpc++) {
+		nv_wr32(dev, GPC_UNIT(gpc, 0x0914), priv->magic_not_rop_nr << 8 |
+						  priv->tp_nr[gpc]);
+		nv_wr32(dev, GPC_UNIT(gpc, 0x0910), 0x00040000 | priv->tp_total);
+		nv_wr32(dev, GPC_UNIT(gpc, 0x0918), priv->magicgpc918);
+	}
+
+	nv_wr32(dev, GPC_BCAST(0x1bd4), priv->magicgpc918);
+	nv_wr32(dev, GPC_BCAST(0x08ac), priv->rop_nr);
+}
+
+static void
+nvc0_graph_init_units(struct drm_device *dev)
+{
+	nv_wr32(dev, 0x409c24, 0x000f0000);
+	nv_wr32(dev, 0x404000, 0xc0000000); /* DISPATCH */
+	nv_wr32(dev, 0x404600, 0xc0000000); /* M2MF */
+	nv_wr32(dev, 0x408030, 0xc0000000);
+	nv_wr32(dev, 0x40601c, 0xc0000000);
+	nv_wr32(dev, 0x404490, 0xc0000000); /* MACRO */
+	nv_wr32(dev, 0x406018, 0xc0000000);
+	nv_wr32(dev, 0x405840, 0xc0000000);
+	nv_wr32(dev, 0x405844, 0x00ffffff);
+	nv_mask(dev, 0x419cc0, 0x00000008, 0x00000008);
+	nv_mask(dev, 0x419eb4, 0x00001000, 0x00001000);
+}
+
+static void
+nvc0_graph_init_gpc_1(struct drm_device *dev)
+{
+	struct drm_nouveau_private *dev_priv = dev->dev_private;
+	struct nvc0_graph_priv *priv = dev_priv->engine.graph.priv;
+	int gpc, tp;
+
+	for (gpc = 0; gpc < priv->gpc_nr; gpc++) {
+		nv_wr32(dev, GPC_UNIT(gpc, 0x0420), 0xc0000000);
+		nv_wr32(dev, GPC_UNIT(gpc, 0x0900), 0xc0000000);
+		nv_wr32(dev, GPC_UNIT(gpc, 0x1028), 0xc0000000);
+		nv_wr32(dev, GPC_UNIT(gpc, 0x0824), 0xc0000000);
+		for (tp = 0; tp < priv->tp_nr[gpc]; tp++) {
+			nv_wr32(dev, TP_UNIT(gpc, tp, 0x508), 0xffffffff);
+			nv_wr32(dev, TP_UNIT(gpc, tp, 0x50c), 0xffffffff);
+			nv_wr32(dev, TP_UNIT(gpc, tp, 0x224), 0xc0000000);
+			nv_wr32(dev, TP_UNIT(gpc, tp, 0x48c), 0xc0000000);
+			nv_wr32(dev, TP_UNIT(gpc, tp, 0x084), 0xc0000000);
+			nv_wr32(dev, TP_UNIT(gpc, tp, 0xe44), 0x001ffffe);
+			nv_wr32(dev, TP_UNIT(gpc, tp, 0xe4c), 0x0000000f);
+		}
+		nv_wr32(dev, GPC_UNIT(gpc, 0x2c90), 0xffffffff);
+		nv_wr32(dev, GPC_UNIT(gpc, 0x2c94), 0xffffffff);
+	}
+}
+
+static void
+nvc0_graph_init_rop(struct drm_device *dev)
+{
+	struct drm_nouveau_private *dev_priv = dev->dev_private;
+	struct nvc0_graph_priv *priv = dev_priv->engine.graph.priv;
+	int rop;
+
+	for (rop = 0; rop < priv->rop_nr; rop++) {
+		nv_wr32(dev, ROP_UNIT(rop, 0x144), 0xc0000000);
+		nv_wr32(dev, ROP_UNIT(rop, 0x070), 0xc0000000);
+		nv_wr32(dev, ROP_UNIT(rop, 0x204), 0xffffffff);
+		nv_wr32(dev, ROP_UNIT(rop, 0x208), 0xffffffff);
+	}
+}
+
+static int
+nvc0_fuc_load_fw(struct drm_device *dev, u32 fuc_base,
+		 const char *code_fw, const char *data_fw)
+{
+	const struct firmware *fw;
+	char name[32];
+	int ret, i;
+
+	snprintf(name, sizeof(name), "nouveau/%s", data_fw);
+	ret = request_firmware(&fw, name, &dev->pdev->dev);
+	if (ret) {
+		NV_ERROR(dev, "failed to load %s\n", data_fw);
+		return ret;
+	}
+
+	nv_wr32(dev, fuc_base + 0x01c0, 0x01000000);
+	for (i = 0; i < fw->size / 4; i++)
+		nv_wr32(dev, fuc_base + 0x01c4, ((u32 *)fw->data)[i]);
+	release_firmware(fw);
+
+	snprintf(name, sizeof(name), "nouveau/%s", code_fw);
+	ret = request_firmware(&fw, name, &dev->pdev->dev);
+	if (ret) {
+		NV_ERROR(dev, "failed to load %s\n", code_fw);
+		return ret;
+	}
+
+	nv_wr32(dev, fuc_base + 0x0180, 0x01000000);
+	for (i = 0; i < fw->size / 4; i++) {
+		if ((i & 0x3f) == 0)
+			nv_wr32(dev, fuc_base + 0x0188, i >> 6);
+		nv_wr32(dev, fuc_base + 0x0184, ((u32 *)fw->data)[i]);
+	}
+	release_firmware(fw);
+
+	return 0;
+}
+
+static int
+nvc0_graph_init_ctxctl(struct drm_device *dev)
+{
+	struct drm_nouveau_private *dev_priv = dev->dev_private;
+	struct nvc0_graph_priv *priv = dev_priv->engine.graph.priv;
+	u32 r000260;
+	int ret;
+
+	/* load fuc microcode */
+	r000260 = nv_mask(dev, 0x000260, 0x00000001, 0x00000000);
+	ret = nvc0_fuc_load_fw(dev, 0x409000, "fuc409c", "fuc409d");
+	if (ret == 0)
+		nvc0_fuc_load_fw(dev, 0x41a000, "fuc41ac", "fuc41ad");
+	nv_wr32(dev, 0x000260, r000260);
+
+	if (ret)
+		return ret;
+
+	/* start both of them running */
+	nv_wr32(dev, 0x409840, 0xffffffff);
+	nv_wr32(dev, 0x41a10c, 0x00000000);
+	nv_wr32(dev, 0x40910c, 0x00000000);
+	nv_wr32(dev, 0x41a100, 0x00000002);
+	nv_wr32(dev, 0x409100, 0x00000002);
+	if (!nv_wait(dev, 0x409800, 0x00000001, 0x00000001))
+		NV_INFO(dev, "0x409800 wait failed\n");
+
+	nv_wr32(dev, 0x409840, 0xffffffff);
+	nv_wr32(dev, 0x409500, 0x7fffffff);
+	nv_wr32(dev, 0x409504, 0x00000021);
+
+	nv_wr32(dev, 0x409840, 0xffffffff);
+	nv_wr32(dev, 0x409500, 0x00000000);
+	nv_wr32(dev, 0x409504, 0x00000010);
+	if (!nv_wait_ne(dev, 0x409800, 0xffffffff, 0x00000000)) {
+		NV_ERROR(dev, "fuc09 req 0x10 timeout\n");
+		return -EBUSY;
+	}
+	priv->grctx_size = nv_rd32(dev, 0x409800);
+
+	nv_wr32(dev, 0x409840, 0xffffffff);
+	nv_wr32(dev, 0x409500, 0x00000000);
+	nv_wr32(dev, 0x409504, 0x00000016);
+	if (!nv_wait_ne(dev, 0x409800, 0xffffffff, 0x00000000)) {
+		NV_ERROR(dev, "fuc09 req 0x16 timeout\n");
+		return -EBUSY;
+	}
+
+	nv_wr32(dev, 0x409840, 0xffffffff);
+	nv_wr32(dev, 0x409500, 0x00000000);
+	nv_wr32(dev, 0x409504, 0x00000025);
+	if (!nv_wait_ne(dev, 0x409800, 0xffffffff, 0x00000000)) {
+		NV_ERROR(dev, "fuc09 req 0x25 timeout\n");
+		return -EBUSY;
+	}
+
+	return 0;
 }
 
 int
 nvc0_graph_init(struct drm_device *dev)
 {
 	struct drm_nouveau_private *dev_priv = dev->dev_private;
+	struct nouveau_pgraph_engine *pgraph = &dev_priv->engine.graph;
+	struct nvc0_graph_priv *priv;
+	int ret;
+
 	dev_priv->engine.graph.accel_blocked = true;
+
+	switch (dev_priv->chipset) {
+	case 0xc0:
+	case 0xc3:
+	case 0xc4:
+		break;
+	default:
+		NV_ERROR(dev, "PGRAPH: unsupported chipset, please report!\n");
+		if (nouveau_noaccel != 0)
+			return 0;
+		break;
+	}
+
+	nv_mask(dev, 0x000200, 0x18001000, 0x00000000);
+	nv_mask(dev, 0x000200, 0x18001000, 0x18001000);
+
+	if (!pgraph->priv) {
+		ret = nvc0_graph_create(dev);
+		if (ret)
+			return ret;
+	}
+	priv = pgraph->priv;
+
+	nvc0_graph_init_obj418880(dev);
+	nvc0_graph_init_regs(dev);
+	//nvc0_graph_init_unitplemented_magics(dev);
+	nvc0_graph_init_gpc_0(dev);
+	//nvc0_graph_init_unitplemented_c242(dev);
+
+	nv_wr32(dev, 0x400500, 0x00010001);
+	nv_wr32(dev, 0x400100, 0xffffffff);
+	nv_wr32(dev, 0x40013c, 0xffffffff);
+
+	nvc0_graph_init_units(dev);
+	nvc0_graph_init_gpc_1(dev);
+	nvc0_graph_init_rop(dev);
+
+	nv_wr32(dev, 0x400108, 0xffffffff);
+	nv_wr32(dev, 0x400138, 0xffffffff);
+	nv_wr32(dev, 0x400118, 0xffffffff);
+	nv_wr32(dev, 0x400130, 0xffffffff);
+	nv_wr32(dev, 0x40011c, 0xffffffff);
+	nv_wr32(dev, 0x400134, 0xffffffff);
+	nv_wr32(dev, 0x400054, 0x34ce3464);
+
+	ret = nvc0_graph_init_ctxctl(dev);
+	if (ret)
+		return ret;
+
+	dev_priv->engine.graph.accel_blocked = false;
 	return 0;
 }
 
+static struct nouveau_enum nvc0_graph_data_error[] = {
+	{ 5, "INVALID_ENUM" },
+	{}
+};
+
+static int
+nvc0_graph_isr_chid(struct drm_device *dev, u64 inst)
+{
+	struct drm_nouveau_private *dev_priv = dev->dev_private;
+	struct nouveau_channel *chan;
+	unsigned long flags;
+	int i;
+
+	spin_lock_irqsave(&dev_priv->channels.lock, flags);
+	for (i = 0; i < dev_priv->engine.fifo.channels; i++) {
+		chan = dev_priv->channels.ptr[i];
+		if (!chan || !chan->ramin)
+			continue;
+
+		if (inst == chan->ramin->vinst)
+			break;
+	}
+	spin_unlock_irqrestore(&dev_priv->channels.lock, flags);
+	return i;
+}
+
+static void
+nvc0_graph_isr(struct drm_device *dev)
+{
+	u64 inst = (u64)(nv_rd32(dev, 0x409b00) & 0x0fffffff) << 12;
+	u32 chid = nvc0_graph_isr_chid(dev, inst);
+	u32 stat = nv_rd32(dev, 0x400100);
+	u32 addr = nv_rd32(dev, 0x400704);
+	u32 mthd = (addr & 0x00003ffc);
+	u32 subc = (addr & 0x00070000) >> 16;
+	u32 data = nv_rd32(dev, 0x400708);
+	u32 code = nv_rd32(dev, 0x400110);
+	u32 class = nv_rd32(dev, 0x404200 + (subc * 4));
+
+	if (stat & 0x00000010) {
+		NV_INFO(dev, "PGRAPH: ILLEGAL_MTHD ch %d [0x%010llx] subc %d "
+			     "class 0x%04x mthd 0x%04x data 0x%08x\n",
+			chid, inst, subc, class, mthd, data);
+		nv_wr32(dev, 0x400100, 0x00000010);
+		stat &= ~0x00000010;
+	}
+
+	if (stat & 0x00100000) {
+		NV_INFO(dev, "PGRAPH: DATA_ERROR [");
+		nouveau_enum_print(nvc0_graph_data_error, code);
+		printk("] ch %d [0x%010llx] subc %d class 0x%04x "
+		       "mthd 0x%04x data 0x%08x\n",
+		       chid, inst, subc, class, mthd, data);
+		nv_wr32(dev, 0x400100, 0x00100000);
+		stat &= ~0x00100000;
+	}
+
+	if (stat & 0x00080000) {
+		u32 ustat = nv_rd32(dev, 0x409c18);
+
+		NV_INFO(dev, "PGRAPH: CTXCTRL ustat 0x%08x\n", ustat);
+
+		nv_wr32(dev, 0x409c20, ustat);
+		nv_wr32(dev, 0x400100, 0x00080000);
+		stat &= ~0x00080000;
+	}
+
+	if (stat) {
+		NV_INFO(dev, "PGRAPH: unknown stat 0x%08x\n", stat);
+		nv_wr32(dev, 0x400100, stat);
+	}
+
+	nv_wr32(dev, 0x400500, 0x00010001);
+}