add draw support

5 files changed, 59 insertions, 32 deletions

diff --git a/r600_atom.c b/r600_atom.c
index 12417d1..930490b 100644
--- a/r600_atom.c
+++ b/r600_atom.c
@@ -17,9 +17,6 @@
 #include "r600.h"
 #include "r600d.h"
 
-/* FIXME delete */
-void r600_tflat(struct r600_atom *atom);
-
 /*
  * drm_r600_framebuffer
  */
@@ -159,7 +156,6 @@ int r600_framebuffer_create(struct r600_winsys *rdev, struct r600_atom *atom, vo
 	atom->pkts[atom->npkts++] = 0x0000008E;
 	atom->pkts[atom->npkts++] = cb_target_mask;
 	atom->pkts[atom->npkts++] = cb_shader_mask;
-	r600_tflat(atom);
 fprintf(stderr, "%s %d pkts\n", __func__, atom->npkts);
 	return 0;
 }
@@ -540,8 +536,7 @@ int r600_fs_shader_emit(struct r600_winsys *rdev,
 	int r;
 
 	atom->pkts[4] = radeon_ib_reloc(ib, atom->bo[0], RADEON_GEM_DOMAIN_GTT);
-	r = radeon_ib_copy(ib, atom->pkts, atom->npkts);
-	return r;
+	return radeon_ib_copy(ib, atom->pkts, atom->npkts);
 }
 
 int r600_fs_shader_create(struct r600_winsys *rdev, struct r600_atom *atom, void *data)
@@ -592,31 +587,34 @@ fprintf(stderr, "%s %d pkts\n", __func__, atom->npkts);
 	return 0;
 }
 
-#define WPKT(a) (atom->pkts[atom->npkts++] = (a))
-#define WPKT_RELOC(s)\
-	do {\
-		WPKT(PKT3(PKT3_NOP, 0));\
-		if (!strcmp("bo", s)) {\
-		fprintf(stderr, "cb->pkts[%d] = radeon_ib_reloc(ib, cb->"s", cb->placements[0] | cb->placements[1]);\n", atom->npkts);\
-		} else {\
-		fprintf(stderr, "cb->pkts[%d] = radeon_ib_reloc(ib, cb->"s", RADEON_GEM_DOMAIN_GTT);\n", atom->npkts);\
-		}\
-		WPKT(0x00000000);\
-	} while (0)
-void r600_tflat(struct r600_atom *atom)
+int r600_draw_cmd_size(struct drm_r600_batch *batch)
+{
+	switch (G_0287F0_SOURCE_SELECT(batch->vgt_draw_initiator)) {
+	case 2:
+		return 10;
+	default:
+		return 0;
+	}
+}
+
+int r600_draw_cmd_emit(struct radeon_ib *ib, struct drm_r600_batch *batch)
 {
-    WPKT(PKT3(PKT3_SET_CONTEXT_REG, 2));
-    WPKT(0x00000094);
-    WPKT(0x80000000);
-    WPKT(0x00FA00FA);
-    WPKT(PKT3(PKT3_SET_CONFIG_REG, 1));
-    WPKT(0x00000256);
-    WPKT(0x00000005);
-    WPKT(PKT3(PKT3_INDEX_TYPE, 0));
-    WPKT(0x00000000);
-    WPKT(PKT3(PKT3_NUM_INSTANCES, 0));
-    WPKT(0x00000001);
-    WPKT(PKT3(PKT3_DRAW_INDEX_AUTO, 1));
-    WPKT(0x00000004);
-    WPKT(0x00000002);
+	u32 pkts[32], i = 0;
+
+	switch (G_0287F0_SOURCE_SELECT(batch->vgt_draw_initiator)) {
+	case 2:
+		pkts[i++] = PKT3(PKT3_SET_CONFIG_REG, 1);
+		pkts[i++] = 0x00000256;
+		pkts[i++] = batch->vgt_primitive_type;
+		pkts[i++] = PKT3(PKT3_INDEX_TYPE, 0);
+		pkts[i++] = batch->vgt_dma_index_type;
+		pkts[i++] = PKT3(PKT3_NUM_INSTANCES, 0);
+		pkts[i++] = batch->vgt_dma_num_instances;
+		pkts[i++] = PKT3(PKT3_DRAW_INDEX_AUTO, 1);
+		pkts[i++] = batch->vgt_num_indices;
+		pkts[i++] = batch->vgt_draw_initiator;
+		return radeon_ib_copy(ib, pkts, i);
+	default:
+		return 0;
+	}
 }
diff --git a/r600_batch.c b/r600_batch.c
index 38e8c83..3c4ef58 100644
--- a/r600_batch.c
+++ b/r600_batch.c
@@ -179,6 +179,7 @@ static int r600_batches_flush_locked(struct r600_winsys *rdev, struct r600_batch
 			if (r)
 				goto out_err;
 		}
+		r = r600_draw_cmd_emit(batches->ib, &batch->drm);
 		/* flush + wait until */
 		batches->ib->ptr[batches->ib->cpkts++] = PKT3(PKT3_EVENT_WRITE, 0);
 		batches->ib->ptr[batches->ib->cpkts++] = 0x00000016;
@@ -230,6 +231,7 @@ int r600_batches_queue(struct r600_winsys *rdev, struct r600_request *rq)
 	rbatch->atoms[i++] = batch->vs_constants;
 	rbatch->atoms[i++] = batch->ps_constants;
 	rbatch->atoms[i++] = batch->framebuffer;
+	memcpy(&rbatch->drm, batch, sizeof(struct drm_r600_batch));
 reprocess:
 	r600_atom_flush_cleanup(&rbatch->pre_flushes);
 	r600_atom_flush_cleanup(&rbatch->post_flushes);
@@ -237,6 +239,7 @@ reprocess:
 	rbatch->npkts = 0;
 	/* flush + wait until = 5dw */
 	rbatch->npkts += 5;
+	rbatch->npkts += r600_draw_cmd_size(batch);
 	for (i = 0; i < rbatch->inputs.drm.nbuffers; i++) {
 		if (!r600_batches_fs_resource_is_present(rdev, batches, &rbatch->inputs.drm.buffers[i])) {
 			rbatch->nfs_resources += 1;
diff --git a/r600_winsys.h b/r600_winsys.h
index 9cbc503..bf162e5 100644
--- a/r600_winsys.h
+++ b/r600_winsys.h
@@ -226,6 +226,11 @@ struct drm_r600_batch {
 	struct r600_atom		*vs_shader;
 	struct r600_atom		*fs_shader;
 	struct drm_r600_vs_input	inputs;
+	u32				vgt_primitive_type;
+	u32				vgt_dma_index_type;
+	u32				vgt_dma_num_instances;
+	u32				vgt_num_indices;
+	u32				vgt_draw_initiator;
 };
 
 struct r600_request {
diff --git a/r600d.h b/r600d.h
index 77ad2ec..7436cee 100644
--- a/r600d.h
+++ b/r600d.h
@@ -722,5 +722,21 @@
 #define   S_028208_BR_Y(x)                             (((x) & 0x3FFF) << 16)
 #define   G_028208_BR_Y(x)                             (((x) >> 16) & 0x3FFF)
 #define   C_028208_BR_Y                                0xC000FFFF
+#define R_0287F0_VGT_DRAW_INITIATOR                  0x0287F0
+#define   S_0287F0_SOURCE_SELECT(x)                    (((x) & 0x3) << 0)
+#define   G_0287F0_SOURCE_SELECT(x)                    (((x) >> 0) & 0x3)
+#define   C_0287F0_SOURCE_SELECT                       0xFFFFFFFC
+#define   S_0287F0_MAJOR_MODE(x)                       (((x) & 0x3) << 2)
+#define   G_0287F0_MAJOR_MODE(x)                       (((x) >> 2) & 0x3)
+#define   C_0287F0_MAJOR_MODE                          0xFFFFFFF3
+#define   S_0287F0_SPRITE_EN(x)                        (((x) & 0x1) << 4)
+#define   G_0287F0_SPRITE_EN(x)                        (((x) >> 4) & 0x1)
+#define   C_0287F0_SPRITE_EN                           0xFFFFFFEF
+#define   S_0287F0_NOT_EOP(x)                          (((x) & 0x1) << 5)
+#define   G_0287F0_NOT_EOP(x)                          (((x) >> 5) & 0x1)
+#define   C_0287F0_NOT_EOP                             0xFFFFFFDF
+#define   S_0287F0_USE_OPAQUE(x)                       (((x) & 0x1) << 6)
+#define   G_0287F0_USE_OPAQUE(x)                       (((x) >> 6) & 0x1)
+#define   C_0287F0_USE_OPAQUE                          0xFFFFFFBF
 
 #endif
diff --git a/test.c b/test.c
index ac92cd2..825d2f1 100644
--- a/test.c
+++ b/test.c
@@ -318,6 +318,11 @@ int r600_tri_flat(struct radeon *radeon)
 	batch.inputs.elements[1].sq_vtx_word0 = 0x7C000000;
 	batch.inputs.elements[1].sq_vtx_word1 = 0x18ED1000;
 	batch.inputs.elements[1].sq_vtx_word2 = 0x00080000;
+	batch.vgt_primitive_type = 5;
+	batch.vgt_dma_index_type = 0;
+	batch.vgt_dma_num_instances = 1;
+	batch.vgt_num_indices = 4;
+	batch.vgt_draw_initiator = 2;
 	rq.type = 0;
 	rq.data = &batch;
 	rq.bo[0] = vbo1;