add constants atom and convert to use ptr

7 files changed, 170 insertions, 77 deletions

diff --git a/lwrapper.h b/lwrapper.h
index 93bf02b..409078b 100644
--- a/lwrapper.h
+++ b/lwrapper.h
@@ -49,7 +49,7 @@ extern void idr_destroy(struct idr *idp);
 extern void idr_init(struct idr *idp);
 
 #include "list.h"
-#define dev_err(d, p, args...) fprintf(stderr, p, args)
+#define dev_err(d, p, args...) fprintf(stderr, p, ##args)
 
 struct kref {
 	int refcount;
diff --git a/r600_atom.c b/r600_atom.c
index d5ef81e..04af0f4 100644
--- a/r600_atom.c
+++ b/r600_atom.c
@@ -105,10 +105,10 @@ static int r600_cb_emit(struct radeon_device *rdev,
 	cb->pkts[273] = radeon_ib_reloc(ib, cb->vsshader1, RADEON_GEM_DOMAIN_GTT);
 	cb->pkts[289] = radeon_ib_reloc(ib, cb->psshader1, RADEON_GEM_DOMAIN_GTT);
 	cb->pkts[294] = radeon_ib_reloc(ib, cb->psshader1, RADEON_GEM_DOMAIN_GTT);
+	cb->pkts[319] = radeon_ib_reloc(ib, cb->vbo1, RADEON_GEM_DOMAIN_GTT);
+	cb->pkts[330] = radeon_ib_reloc(ib, cb->vbo1, RADEON_GEM_DOMAIN_GTT);
 	cb->pkts[337] = radeon_ib_reloc(ib, cb->vbo1, RADEON_GEM_DOMAIN_GTT);
 	cb->pkts[348] = radeon_ib_reloc(ib, cb->vbo1, RADEON_GEM_DOMAIN_GTT);
-	cb->pkts[355] = radeon_ib_reloc(ib, cb->vbo1, RADEON_GEM_DOMAIN_GTT);
-	cb->pkts[366] = radeon_ib_reloc(ib, cb->vbo1, RADEON_GEM_DOMAIN_GTT);
 	r = radeon_ib_copy(ib, cb->pkts, atom->npkts);
 	return r;
 }
@@ -733,11 +733,73 @@ out_err:
 }
 
 /*
+ * r600_constant
+ */
+struct r600_constants {
+	struct radeon_atom	atom;
+	u32			pkts[1024];
+	struct r600_atoms	*atoms;
+};
+
+static void r600_constants_release(struct kref *kref)
+{
+	struct radeon_atom *atom = container_of(kref, struct radeon_atom, kref);
+	struct r600_constants *constants = container_of(atom, struct r600_constants, atom);
+
+	mutex_lock(&constants->atoms->mutex);
+	list_del_init(&constants->atom.list);
+	mutex_unlock(&constants->atoms->mutex);
+	kfree(constants);
+}
+
+static int r600_constants_create(struct radeon_device *rdev,
+				struct r600_atoms *atoms,
+				struct drm_radeon_atom *patom,
+				struct radeon_atom **atom)
+{
+	struct drm_r600_constants pconstants;
+	struct r600_constants *constants;
+	int r;
+
+	constants = kmalloc(sizeof(*constants), GFP_KERNEL);
+	if (constants == NULL)
+		return -ENOMEM;
+	/* make sure structure properly initialized */
+	memset(constants, 0, sizeof(*constants));
+	r = radeon_atom_init(&constants->atom, &atoms->idr, &r600_constants_release,
+			&radeon_atom_emit_default, &r600_atom_process_default,
+			constants->pkts);
+	if (r)
+		goto out_err;
+	/* KERNEL use get user data */
+	memcpy(&pconstants, (void*)(unsigned long)patom->data, sizeof(struct drm_r600_constants));
+	if ((pconstants.nconstants & 0x3) || (pconstants.offset & 0x3)) {
+		dev_err(rdev->dev, "invalid offset or number of constants (0x%08X %d)\n",
+			pconstants.offset, pconstants.nconstants);
+			return -EINVAL;
+	}
+	/* constants */
+	constants->pkts[constants->atom.npkts++] = PKT3(PKT3_SET_ALU_CONST, pconstants.nconstants * 4);
+	constants->pkts[constants->atom.npkts++] = pconstants.offset;
+	memcpy(&constants->pkts[constants->atom.npkts],
+		(void*)(unsigned long)pconstants.constants,
+		pconstants.nconstants * 4 * 4);
+	constants->atom.npkts += pconstants.nconstants * 4;
+	*atom = &constants->atom;
+	return 0;
+out_err:
+	radeon_atom_put(&constants->atom);
+	*atom = NULL;
+	return r;
+}
+
+/*
  * r600 atom core functions
  */
 int r600_atom_create(struct radeon_device *rdev,
 			struct r600_atoms *atoms,
-			struct drm_radeon_atom *patom)
+			struct drm_radeon_atom *patom,
+			struct radeon_atom **atomptr)
 {
 	struct radeon_atom *atom;
 	int r;
@@ -787,11 +849,19 @@ int r600_atom_create(struct radeon_device *rdev,
 		atom->type = patom->type;
 		list_add_tail(&atom->list, &atoms->blend_atoms);
 		break;
+	case R600_ATOM_CONSTANTS:
+		r = r600_constants_create(rdev, atoms, patom, &atom);
+		if (r)
+			return r;
+		atom->type = patom->type;
+		list_add_tail(&atom->list, &atoms->constants_atoms);
+		break;
 	default:
 		dev_err(rdev->dev, "unknown R600 atom type 0x%08X\n", patom->type);
 		return -EINVAL;
 	}
 	patom->id = atom->id;
+	*atomptr = atom;
 	mutex_unlock(&atoms->mutex);
 	return 0;
 }
@@ -810,6 +880,7 @@ static int r600_batch_alloc(struct r600_batch **batch)
 	INIT_LIST_HEAD(&rbatch->list);
 	INIT_LIST_HEAD(&rbatch->pre_flushes);
 	INIT_LIST_HEAD(&rbatch->post_flushes);
+	rbatch->nemit_atoms = 0;
 	*batch = rbatch;
 	return 0;
 }
@@ -824,7 +895,8 @@ static void r600_batches_clear_locked(struct radeon_device *rdev, struct r600_ba
 
 	list_for_each_entry_safe(batch, n, &batches->batches, list) {
 		for (i = 0; i < R600_BATCH_NATOMS; i++) {
-			radeon_atom_put(batch->atoms[i]);
+			if (batch->atoms[i])
+				radeon_atom_put(batch->atoms[i]);
 		}
 		radeon_atom_flush_cleanup(&batch->pre_flushes);
 		radeon_atom_flush_cleanup(&batch->post_flushes);
@@ -862,8 +934,8 @@ static int r600_batches_flush_locked(struct radeon_device *rdev, struct r600_bat
 		list_for_each_entry(flush, &batch->pre_flushes, list) {
 			r600_emit_flush(rdev, batches->ib, flush->bo, flush->flags);
 		}
-		for (i = 0; i < R600_BATCH_NATOMS; i++) {
-			r = batch->atoms[i]->emit(rdev, batch->atoms[i], batch, batches->ib);
+		for (i = 0; i < batch->nemit_atoms; i++) {
+			r = batch->emit_atoms[i]->emit(rdev, batch->emit_atoms[i], batch, batches->ib);
 			if (r)
 				goto out_err;
 		}
@@ -910,30 +982,30 @@ int r600_batches_queue(struct radeon_device *rdev,
 	rbatch->cb_fog_green = batch->cb_fog_green;
 	rbatch->cb_fog_red = batch->cb_fog_red;
 	i = 0;
-	r = radeon_atom_find_locked(&atoms->blend_atoms, batch->blend_id,
-					R600_ATOM_BLEND, &rbatch->atoms[i++]);
-	if (r)
-		goto out_err;
-	r = radeon_atom_find_locked(&atoms->cb_cntl_atoms, batch->cb_cntl_id,
-					R600_ATOM_CB_CNTL, &rbatch->atoms[i++]);
-	if (r)
-		goto out_err;
-	r = radeon_atom_find_locked(&atoms->pa_atoms, batch->pa_id,
-					R600_ATOM_PA, &rbatch->atoms[i++]);
-	if (r)
-		goto out_err;
-	r = radeon_atom_find_locked(&atoms->tp_atoms, batch->tp_id,
-					R600_ATOM_TP, &rbatch->atoms[i++]);
-	if (r)
-		goto out_err;
-	r = radeon_atom_find_locked(&atoms->vport_atoms, batch->vport_id,
-					R600_ATOM_VPORT, &rbatch->atoms[i++]);
-	if (r)
-		goto out_err;
-	r = radeon_atom_find_locked(&atoms->cb_atoms, batch->cb_id,
-					R600_ATOM_CB, &rbatch->atoms[i++]);
-	if (r)
-		goto out_err;
+	if (batch->blend == NULL || batch->cb_cntl == NULL ||
+		batch->pa == NULL || batch->vport == NULL ||
+		batch->tp == NULL || batch->cb == NULL) {
+		dev_err(rdev->dev, "invalid batch\n");
+		return -EINVAL;
+	}
+	rbatch->atoms[i++] = batch->blend;
+	kref_get(&batch->blend->kref);
+	rbatch->atoms[i++] = batch->cb_cntl;
+	kref_get(&batch->cb_cntl->kref);
+	rbatch->atoms[i++] = batch->pa;
+	kref_get(&batch->pa->kref);
+	rbatch->atoms[i++] = batch->vport;
+	kref_get(&batch->vport->kref);
+	rbatch->atoms[i++] = batch->tp;
+	kref_get(&batch->tp->kref);
+	rbatch->atoms[i++] = batch->vs_constants;
+	if (batch->vs_constants)
+		kref_get(&batch->vs_constants->kref);
+	rbatch->atoms[i++] = batch->ps_constants;
+	if (batch->ps_constants)
+		kref_get(&batch->ps_constants->kref);
+	rbatch->atoms[i++] = batch->cb;
+	kref_get(&batch->cb->kref);
 reprocess:
 	radeon_atom_flush_cleanup(&rbatch->pre_flushes);
 	radeon_atom_flush_cleanup(&rbatch->post_flushes);
@@ -943,10 +1015,14 @@ reprocess:
 	/* flush + wait until = 5dw */
 	rbatch->npkts += 5;
 	for (i = 0; i < R600_BATCH_NATOMS; i++) {
-		r = rbatch->atoms[i]->process(rdev, rbatch->atoms[i], batches->last_id[i], rbatch);
-		if (r)
-			goto out_err;
-		rbatch->nflushes += rbatch->atoms[i]->nflushes;
+		if (rbatch->atoms[i]) {
+			r = rbatch->atoms[i]->process(rdev, rbatch->atoms[i], batches->last_id[i], rbatch);
+			if (r)
+				goto out_err;
+			rbatch->nflushes += rbatch->atoms[i]->nflushes;
+			if (batches->last_id[i] != rbatch->atoms[i]->id)
+				rbatch->emit_atoms[rbatch->nemit_atoms++] = rbatch->atoms[i];
+		}
 	}
 	/* add flush */
 	rbatch->npkts += rbatch->nflushes * 7;
@@ -967,7 +1043,9 @@ reprocess:
 	}
 	/* batch is queued */
 	for (i = 0; i < R600_BATCH_NATOMS; i++) {
-		batches->last_id[i] = rbatch->atoms[i]->id;
+		if (rbatch->atoms[i]) {
+			batches->last_id[i] = rbatch->atoms[i]->id;
+		}
 	}
 	batches->npkts += rbatch->npkts;
 	list_add_tail(&rbatch->list, &batches->batches);
@@ -1017,6 +1095,7 @@ int r600_atoms_init(struct radeon_device *rdev, struct r600_atoms *atoms)
 	INIT_LIST_HEAD(&atoms->blend_atoms);
 	INIT_LIST_HEAD(&atoms->cb_atoms);
 	INIT_LIST_HEAD(&atoms->cb_cntl_atoms);
+	INIT_LIST_HEAD(&atoms->constants_atoms);
 	INIT_LIST_HEAD(&atoms->pa_atoms);
 	INIT_LIST_HEAD(&atoms->sq_atoms);
 	INIT_LIST_HEAD(&atoms->tp_atoms);
@@ -1328,24 +1407,6 @@ void r600_tflat(struct radeon_atom *atom)
     WPKT(PKT3(PKT3_SET_LOOP_CONST, 1));
     WPKT(0x00000000);
     WPKT(0x01000FFF);
-    WPKT(PKT3(PKT3_SET_ALU_CONST, 16));
-    WPKT(0x00000400);
-    WPKT(0x3C03126F);
-    WPKT(0x00000000);
-    WPKT(0x00000000);
-    WPKT(0xBF800000);
-    WPKT(0x00000000);
-    WPKT(0x3C03126F);
-    WPKT(0x00000000);
-    WPKT(0xBF800000);
-    WPKT(0x00000000);
-    WPKT(0x00000000);
-    WPKT(0xBF800000);
-    WPKT(0x00000000);
-    WPKT(0x00000000);
-    WPKT(0x00000000);
-    WPKT(0x00000000);
-    WPKT(0x3F800000);
     WPKT(PKT3(PKT3_SET_CTL_CONST, 1));
     WPKT(0x00000000);
     WPKT(0x00000000);
diff --git a/r600_atom_api.h b/r600_atom_api.h
index 16938c7..ecc3e9c 100644
--- a/r600_atom_api.h
+++ b/r600_atom_api.h
@@ -33,6 +33,7 @@ struct drm_radeon_atom {
 #define R600_ATOM_CB_CNTL	4
 #define R600_ATOM_VPORT		5
 #define R600_ATOM_BLEND		6
+#define R600_ATOM_CONSTANTS	7
 
 struct drm_r600_cb {
 	u32			pitch;
@@ -128,13 +129,14 @@ struct drm_r600_blend {
 	u32			cb_blend_control;
 };
 
+/* constant */
+struct drm_r600_constants {
+	u32			offset;
+	u32			nconstants;
+	u64			constants;
+};
+
 struct drm_r600_batch {
-	u32			blend_id;
-	u32			cb_id;
-	u32			cb_cntl_id;
-	u32			pa_id;
-	u32			tp_id;
-	u32			vport_id;
 	u32			cb_blend_alpha;
 	u32			cb_blend_blue;
 	u32			cb_blend_green;
@@ -146,6 +148,14 @@ struct drm_r600_batch {
 	u32			cb_fog_blue;
 	u32			cb_fog_green;
 	u32			cb_fog_red;
+	struct radeon_atom	*vs_constants;
+	struct radeon_atom	*ps_constants;
+	struct radeon_atom	*blend;
+	struct radeon_atom	*cb;
+	struct radeon_atom	*cb_cntl;
+	struct radeon_atom	*pa;
+	struct radeon_atom	*tp;
+	struct radeon_atom	*vport;
 };
 
 
diff --git a/r600_winsys.h b/r600_winsys.h
index 708b46f..616be8a 100644
--- a/r600_winsys.h
+++ b/r600_winsys.h
@@ -21,9 +21,10 @@
 
 /* opaque structure */
 struct radeon_device *rdev;
+struct radeon_atom *atom;
 
 /* atom */
-extern int radeon_atom_create(struct radeon_device*, struct drm_radeon_atom*);
+extern int radeon_atom_create(struct radeon_device*, struct drm_radeon_atom*, struct radeon_atom **atom);
 extern int radeon_batches_queue(struct radeon_device *rdev, void *batch);
 extern int radeon_batches_flush(struct radeon_device *rdev);
 
diff --git a/radeon_atom.c b/radeon_atom.c
index ca6923c..989d18c 100644
--- a/radeon_atom.c
+++ b/radeon_atom.c
@@ -100,11 +100,13 @@ int radeon_atom_emit_default(struct radeon_device *rdev,
 	return radeon_ib_copy(ib, atom->pkts, atom->npkts);
 }
 
-int radeon_atom_create(struct radeon_device *rdev, struct drm_radeon_atom *patom)
+int radeon_atom_create(struct radeon_device *rdev,
+			struct drm_radeon_atom *patom,
+			struct radeon_atom **atom)
 {
 	struct r600_atoms *atoms = &rdev->asic.r700.atoms;
 
-	return r600_atom_create(rdev, atoms, patom);
+	return r600_atom_create(rdev, atoms, patom, atom);
 }
 
 int radeon_batches_queue(struct radeon_device *rdev, void *batch)
diff --git a/radeon_atom.h b/radeon_atom.h
index e3befa2..3f16d15 100644
--- a/radeon_atom.h
+++ b/radeon_atom.h
@@ -53,12 +53,14 @@ struct radeon_atom {
 };
 
 /* R600 */
-#define R600_BATCH_NATOMS	6
+#define R600_BATCH_NATOMS	8
 struct r600_batch {
 	struct list_head	list;
 	struct list_head	pre_flushes;
 	struct list_head	post_flushes;
 	struct radeon_atom	*atoms[R600_BATCH_NATOMS];
+	struct radeon_atom	*emit_atoms[R600_BATCH_NATOMS];
+	u32			nemit_atoms;
 	u32			nflushes;
 	u32			npkts;
 	u32			cb_blend_alpha;
@@ -85,6 +87,7 @@ struct r600_atoms {
 	struct list_head	blend_atoms;
 	struct list_head	cb_atoms;
 	struct list_head	cb_cntl_atoms;
+	struct list_head	constants_atoms;
 	struct list_head	pa_atoms;
 	struct list_head	sq_atoms;
 	struct list_head	tp_atoms;
@@ -122,7 +125,8 @@ extern int r600_atoms_init(struct radeon_device *rdev, struct r600_atoms *atoms)
 extern void r600_atoms_release(struct radeon_device *rdev, struct r600_atoms *atoms);
 extern int r600_atom_create(struct radeon_device *rdev,
 				struct r600_atoms *atoms,
-				struct drm_radeon_atom *patom);
+				struct drm_radeon_atom *patom,
+				struct radeon_atom **atomptr);
 extern int r600_batches_queue(struct radeon_device *rdev,
 				struct r600_atoms *atoms,
 				struct drm_r600_batch *batch);
diff --git a/test.c b/test.c
index 5bda074..33424fe 100644
--- a/test.c
+++ b/test.c
@@ -44,6 +44,13 @@ int main(void)
 	return 0;
 }
 
+static u32 vsconstants[16] = {
+	0x3C03126F, 0x00000000, 0x00000000, 0xBF800000,
+	0x00000000, 0x3C03126F, 0x00000000, 0xBF800000,
+	0x00000000, 0x00000000, 0xBF800000, 0x00000000,
+	0x00000000, 0x00000000, 0x00000000, 0x3F800000,
+};
+
 int r600_tri_flat(struct radeon *radeon)
 {
 	struct radeon_device *rdev;
@@ -53,6 +60,7 @@ int r600_tri_flat(struct radeon *radeon)
 	struct drm_r600_pa pa;
 	struct drm_r600_tp tp;
 	struct drm_r600_vport vport;
+	struct drm_r600_constants vs_constants;
 	struct drm_r600_batch batch;
 	struct drm_radeon_atom atom;
 	int r;
@@ -71,19 +79,17 @@ int r600_tri_flat(struct radeon *radeon)
 	atom.type = R600_ATOM_CB;
 	atom.id = 0;
 	atom.data = (uint64_t)(uintptr_t)&cb;
-	r = radeon_atom_create(rdev, &atom);
+	r = radeon_atom_create(rdev, &atom, &batch.cb);
 	if (r)
 		return r;
-	batch.cb_id = atom.id;
 	/* build tp */
 	tp.ta_cntl_aux = 0x07000002;
 	atom.type = R600_ATOM_TP;
 	atom.id = 0;
 	atom.data = (uint64_t)(uintptr_t)&tp;
-	r = radeon_atom_create(rdev, &atom);
+	r = radeon_atom_create(rdev, &atom, &batch.tp);
 	if (r)
 		return r;
-	batch.tp_id = atom.id;
 	/* build pa */
 	pa.pa_sc_mpass_ps_cntl = 0x00000000;
 	pa.pa_sc_mode_cntl = 0x00514000;
@@ -113,10 +119,9 @@ int r600_tri_flat(struct radeon *radeon)
 	atom.type = R600_ATOM_PA;
 	atom.id = 0;
 	atom.data = (uint64_t)(uintptr_t)&pa;
-	r = radeon_atom_create(rdev, &atom);
+	r = radeon_atom_create(rdev, &atom, &batch.pa);
 	if (r)
 		return r;
-	batch.pa_id = atom.id;
 	/* cb control */
 	cb_cntl.cb_target_mask = 0x0000000f;
 	cb_cntl.cb_shader_mask = 0x0000000f;
@@ -128,10 +133,9 @@ int r600_tri_flat(struct radeon *radeon)
 	atom.type = R600_ATOM_CB_CNTL;
 	atom.id = 0;
 	atom.data = (uint64_t)(uintptr_t)&cb_cntl;
-	r = radeon_atom_create(rdev, &atom);
+	r = radeon_atom_create(rdev, &atom, &batch.cb_cntl);
 	if (r)
 		return r;
-	batch.cb_cntl_id = atom.id;
 	/* viewport */
 	vport.pa_sc_vport_zmin_0 = 0x00000000;
 	vport.pa_sc_vport_zmax_0 = 0x3f800000;
@@ -160,10 +164,9 @@ int r600_tri_flat(struct radeon *radeon)
 	atom.type = R600_ATOM_VPORT;
 	atom.id = 0;
 	atom.data = (uint64_t)(uintptr_t)&vport;
-	r = radeon_atom_create(rdev, &atom);
+	r = radeon_atom_create(rdev, &atom, &batch.vport);
 	if (r)
 		return r;
-	batch.vport_id = atom.id;
 	/* blend */
 	blend.cb_blend0_control = 0x00010001;
 	blend.cb_blend1_control = 0x00000000;
@@ -177,10 +180,22 @@ int r600_tri_flat(struct radeon *radeon)
 	atom.type = R600_ATOM_BLEND;
 	atom.id = 0;
 	atom.data = (uint64_t)(uintptr_t)&blend;
-	r = radeon_atom_create(rdev, &atom);
+	r = radeon_atom_create(rdev, &atom, &batch.blend);
+	if (r)
+		return r;
+	/* ps constant */
+	batch.ps_constants = NULL;
+	/* vs constant */
+	vs_constants.nconstants = 4;
+	vs_constants.offset = 0x400;
+	vs_constants.constants = (uintptr_t)vsconstants;
+	batch.vs_constants = NULL;
+	atom.type = R600_ATOM_CONSTANTS;
+	atom.id = 0;
+	atom.data = (uint64_t)(uintptr_t)&vs_constants;
+	r = radeon_atom_create(rdev, &atom, &batch.vs_constants);
 	if (r)
 		return r;
-	batch.blend_id = atom.id;
 
 	/* batch */
 	batch.cb_clear_alpha = 0x00000000;