Merge remote branch 'origin/master' into pipe-video

10 files changed, 151 insertions, 46 deletions

diff --git a/src/gallium/winsys/i915/drm/i915_drm_batchbuffer.c b/src/gallium/winsys/i915/drm/i915_drm_batchbuffer.c
index 7cc5af89639..03aa1b1537a 100644
--- a/src/gallium/winsys/i915/drm/i915_drm_batchbuffer.c
+++ b/src/gallium/winsys/i915/drm/i915_drm_batchbuffer.c
@@ -63,7 +63,6 @@ i915_drm_batchbuffer_create(struct i915_winsys *iws)
    batch->base.size = 0;
 
    batch->base.relocs = 0;
-   batch->base.max_relocs = 300;/*INTEL_DEFAULT_RELOCS;*/
 
    batch->base.iws = iws;
 
@@ -104,8 +103,6 @@ i915_drm_batchbuffer_reloc(struct i915_winsys_batchbuffer *ibatch,
    unsigned offset;
    int ret = 0;
 
-   assert(batch->base.relocs < batch->base.max_relocs);
-
    switch (usage) {
    case I915_USAGE_SAMPLER:
       write_domain = 0;
diff --git a/src/gallium/winsys/i915/sw/i915_sw_batchbuffer.c b/src/gallium/winsys/i915/sw/i915_sw_batchbuffer.c
index 3d0c1fa6224..3bf54011d9e 100644
--- a/src/gallium/winsys/i915/sw/i915_sw_batchbuffer.c
+++ b/src/gallium/winsys/i915/sw/i915_sw_batchbuffer.c
@@ -49,7 +49,6 @@ i915_sw_batchbuffer_create(struct i915_winsys *iws)
    batch->base.size = 0;
 
    batch->base.relocs = 0;
-   batch->base.max_relocs = 300;/*INTEL_DEFAULT_RELOCS;*/
 
    batch->base.iws = iws;
 
@@ -75,8 +74,6 @@ i915_sw_batchbuffer_reloc(struct i915_winsys_batchbuffer *ibatch,
    struct i915_sw_batchbuffer *batch = i915_sw_batchbuffer(ibatch);
    int ret = 0;
 
-   assert(batch->base.relocs < batch->base.max_relocs);
-
    if (usage == I915_USAGE_SAMPLER) {
 
    } else if (usage == I915_USAGE_RENDER) {
diff --git a/src/gallium/winsys/r600/drm/evergreen_hw_context.c b/src/gallium/winsys/r600/drm/evergreen_hw_context.c
index aa4035a302b..66398afa698 100644
--- a/src/gallium/winsys/r600/drm/evergreen_hw_context.c
+++ b/src/gallium/winsys/r600/drm/evergreen_hw_context.c
@@ -834,25 +834,25 @@ void evergreen_context_draw(struct r600_context *ctx, const struct r600_draw *dr
 	}
 
 	/* draw packet */
-	ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_INDEX_TYPE, 0);
+	ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_INDEX_TYPE, 0, ctx->predicate_drawing);
 	ctx->pm4[ctx->pm4_cdwords++] = draw->vgt_index_type;
-	ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_NUM_INSTANCES, 0);
+	ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_NUM_INSTANCES, 0, ctx->predicate_drawing);
 	ctx->pm4[ctx->pm4_cdwords++] = draw->vgt_num_instances;
 	if (draw->indices) {
-		ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_DRAW_INDEX, 3);
+	        ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_DRAW_INDEX, 3, ctx->predicate_drawing);
 		ctx->pm4[ctx->pm4_cdwords++] = draw->indices_bo_offset + r600_bo_offset(draw->indices);
 		ctx->pm4[ctx->pm4_cdwords++] = 0;
 		ctx->pm4[ctx->pm4_cdwords++] = draw->vgt_num_indices;
 		ctx->pm4[ctx->pm4_cdwords++] = draw->vgt_draw_initiator;
-		ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_NOP, 0);
+		ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_NOP, 0, ctx->predicate_drawing);
 		ctx->pm4[ctx->pm4_cdwords++] = 0;
 		r600_context_bo_reloc(ctx, &ctx->pm4[ctx->pm4_cdwords - 1], draw->indices);
 	} else {
-		ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_DRAW_INDEX_AUTO, 1);
+		ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_DRAW_INDEX_AUTO, 1, ctx->predicate_drawing);
 		ctx->pm4[ctx->pm4_cdwords++] = draw->vgt_num_indices;
 		ctx->pm4[ctx->pm4_cdwords++] = draw->vgt_draw_initiator;
 	}
-	ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_EVENT_WRITE, 0);
+	ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_EVENT_WRITE, 0, ctx->predicate_drawing);
 	ctx->pm4[ctx->pm4_cdwords++] = EVENT_TYPE(EVENT_TYPE_CACHE_FLUSH_AND_INV_EVENT) | EVENT_INDEX(0);
 
 	/* flush color buffer */
diff --git a/src/gallium/winsys/r600/drm/r600_drm.c b/src/gallium/winsys/r600/drm/r600_drm.c
index f5cd48d39c6..cd0aa318be5 100644
--- a/src/gallium/winsys/r600/drm/r600_drm.c
+++ b/src/gallium/winsys/r600/drm/r600_drm.c
@@ -45,6 +45,10 @@
 #define RADEON_INFO_CLOCK_CRYSTAL_FREQ 0x9
 #endif
 
+#ifndef RADEON_INFO_NUM_BACKENDS
+#define RADEON_INFO_NUM_BACKENDS 0xa
+#endif
+
 enum radeon_family r600_get_family(struct radeon *r600)
 {
 	return r600->family;
@@ -65,6 +69,17 @@ unsigned r600_get_clock_crystal_freq(struct radeon *radeon)
 	return radeon->clock_crystal_freq;
 }
 
+unsigned r600_get_num_backends(struct radeon *radeon)
+{
+	return radeon->num_backends;
+}
+
+unsigned r600_get_minor_version(struct radeon *radeon)
+{
+	return radeon->minor_version;
+}
+
+
 static int radeon_get_device(struct radeon *radeon)
 {
 	struct drm_radeon_info info = {};
@@ -195,6 +210,26 @@ static int radeon_get_clock_crystal_freq(struct radeon *radeon)
 	return 0;
 }
 
+
+static int radeon_get_num_backends(struct radeon *radeon)
+{
+	struct drm_radeon_info info;
+	uint32_t num_backends;
+	int r;
+
+	radeon->device = 0;
+	info.request = RADEON_INFO_NUM_BACKENDS;
+	info.value = (uintptr_t)&num_backends;
+	r = drmCommandWriteRead(radeon->fd, DRM_RADEON_INFO, &info,
+			sizeof(struct drm_radeon_info));
+	if (r)
+		return r;
+
+	radeon->num_backends = num_backends;
+	return 0;
+}
+
+
 static int radeon_init_fence(struct radeon *radeon)
 {
 	radeon->fence = 1;
@@ -211,6 +246,7 @@ static struct radeon *radeon_new(int fd, unsigned device)
 {
 	struct radeon *radeon;
 	int r;
+	drmVersionPtr version;
 
 	radeon = calloc(1, sizeof(*radeon));
 	if (radeon == NULL) {
@@ -219,13 +255,27 @@ static struct radeon *radeon_new(int fd, unsigned device)
 	radeon->fd = fd;
 	radeon->device = device;
 	radeon->refcount = 1;
-	if (fd >= 0) {
-		r = radeon_get_device(radeon);
-		if (r) {
-			fprintf(stderr, "Failed to get device id\n");
-			return radeon_decref(radeon);
-		}
+
+	version = drmGetVersion(radeon->fd);
+	if (version->version_major != 2) {
+		fprintf(stderr, "%s: DRM version is %d.%d.%d but this driver is "
+			"only compatible with 2.x.x\n", __FUNCTION__,
+			version->version_major, version->version_minor,
+			version->version_patchlevel);
+		drmFreeVersion(version);
+		exit(1);
 	}
+
+	radeon->minor_version = version->version_minor;
+
+	drmFreeVersion(version);
+
+	r = radeon_get_device(radeon);
+	if (r) {
+		fprintf(stderr, "Failed to get device id\n");
+		return radeon_decref(radeon);
+	}
+
 	radeon->family = radeon_family_from_device(radeon->device);
 	if (radeon->family == CHIP_UNKNOWN) {
 		fprintf(stderr, "Unknown chipset 0x%04X\n", radeon->device);
@@ -278,6 +328,9 @@ static struct radeon *radeon_new(int fd, unsigned device)
 	/* get the GPU counter frequency, failure is non fatal */
 	radeon_get_clock_crystal_freq(radeon);
 
+	if (radeon->minor_version >= 9)
+		radeon_get_num_backends(radeon);
+
 	radeon->bomgr = r600_bomgr_create(radeon, 1000000);
 	if (radeon->bomgr == NULL) {
 		return NULL;
diff --git a/src/gallium/winsys/r600/drm/r600_hw_context.c b/src/gallium/winsys/r600/drm/r600_hw_context.c
index f170640407d..6b3baa2c909 100644
--- a/src/gallium/winsys/r600/drm/r600_hw_context.c
+++ b/src/gallium/winsys/r600/drm/r600_hw_context.c
@@ -107,7 +107,7 @@ int r600_context_add_block(struct r600_context *ctx, const struct r600_reg *reg,
 
 		/* initialize block */
 		block->start_offset = reg[i].offset;
-		block->pm4[block->pm4_ndwords++] = PKT3(reg[i].opcode, n);
+		block->pm4[block->pm4_ndwords++] = PKT3(reg[i].opcode, n, 0);
 		block->pm4[block->pm4_ndwords++] = (block->start_offset - reg[i].offset_base) >> 2;
 		block->reg = &block->pm4[block->pm4_ndwords];
 		block->pm4_ndwords += n;
@@ -119,7 +119,7 @@ int r600_context_add_block(struct r600_context *ctx, const struct r600_reg *reg,
 				block->nbo++;
 				assert(block->nbo < R600_BLOCK_MAX_BO);
 				block->pm4_bo_index[j] = block->nbo;
-				block->pm4[block->pm4_ndwords++] = PKT3(PKT3_NOP, 0);
+				block->pm4[block->pm4_ndwords++] = PKT3(PKT3_NOP, 0, 0);
 				block->pm4[block->pm4_ndwords++] = 0x00000000;
 				block->reloc[block->nbo].flush_flags = reg[i+j].flush_flags;
 				block->reloc[block->nbo].flush_mask = reg[i+j].flush_mask;
@@ -771,12 +771,12 @@ void r600_context_bo_flush(struct r600_context *ctx, unsigned flush_flags,
 		bo->last_flush &= flush_mask;
 		return;
 	}
-	ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_SURFACE_SYNC, 3);
+	ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_SURFACE_SYNC, 3, ctx->predicate_drawing);
 	ctx->pm4[ctx->pm4_cdwords++] = flush_flags;
 	ctx->pm4[ctx->pm4_cdwords++] = (bo->size + 255) >> 8;
 	ctx->pm4[ctx->pm4_cdwords++] = 0x00000000;
 	ctx->pm4[ctx->pm4_cdwords++] = 0x0000000A;
-	ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_NOP, 0);
+	ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_NOP, 0, ctx->predicate_drawing);
 	ctx->pm4[ctx->pm4_cdwords++] = bo->reloc_id;
 	bo->last_flush = (bo->last_flush | flush_flags) & flush_mask;
 }
@@ -1048,25 +1048,25 @@ void r600_context_draw(struct r600_context *ctx, const struct r600_draw *draw)
 	}
 
 	/* draw packet */
-	ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_INDEX_TYPE, 0);
+	ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_INDEX_TYPE, 0, ctx->predicate_drawing);
 	ctx->pm4[ctx->pm4_cdwords++] = draw->vgt_index_type;
-	ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_NUM_INSTANCES, 0);
+	ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_NUM_INSTANCES, 0, ctx->predicate_drawing);
 	ctx->pm4[ctx->pm4_cdwords++] = draw->vgt_num_instances;
 	if (draw->indices) {
-		ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_DRAW_INDEX, 3);
+		ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_DRAW_INDEX, 3, ctx->predicate_drawing);
 		ctx->pm4[ctx->pm4_cdwords++] = draw->indices_bo_offset + r600_bo_offset(draw->indices);
 		ctx->pm4[ctx->pm4_cdwords++] = 0;
 		ctx->pm4[ctx->pm4_cdwords++] = draw->vgt_num_indices;
 		ctx->pm4[ctx->pm4_cdwords++] = draw->vgt_draw_initiator;
-		ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_NOP, 0);
+		ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_NOP, 0, ctx->predicate_drawing);
 		ctx->pm4[ctx->pm4_cdwords++] = 0;
 		r600_context_bo_reloc(ctx, &ctx->pm4[ctx->pm4_cdwords - 1], draw->indices);
 	} else {
-		ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_DRAW_INDEX_AUTO, 1);
+		ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_DRAW_INDEX_AUTO, 1, ctx->predicate_drawing);
 		ctx->pm4[ctx->pm4_cdwords++] = draw->vgt_num_indices;
 		ctx->pm4[ctx->pm4_cdwords++] = draw->vgt_draw_initiator;
 	}
-	ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_EVENT_WRITE, 0);
+	ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_EVENT_WRITE, 0, ctx->predicate_drawing);
 	ctx->pm4[ctx->pm4_cdwords++] = EVENT_TYPE(EVENT_TYPE_CACHE_FLUSH_AND_INV_EVENT) | EVENT_INDEX(0);
 
 	/* flush color buffer */
@@ -1101,15 +1101,15 @@ void r600_context_flush(struct r600_context *ctx)
 	r600_context_queries_suspend(ctx);
 
 	/* emit fence */
-	ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_EVENT_WRITE, 0);
+	ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_EVENT_WRITE, 0, 0);
 	ctx->pm4[ctx->pm4_cdwords++] = EVENT_TYPE(EVENT_TYPE_PS_PARTIAL_FLUSH) | EVENT_INDEX(4);
-	ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_EVENT_WRITE_EOP, 4);
+	ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_EVENT_WRITE_EOP, 4, 0);
 	ctx->pm4[ctx->pm4_cdwords++] = EVENT_TYPE(EVENT_TYPE_CACHE_FLUSH_AND_INV_TS_EVENT) | EVENT_INDEX(5);
 	ctx->pm4[ctx->pm4_cdwords++] = 0;
 	ctx->pm4[ctx->pm4_cdwords++] = (1 << 29) | (0 << 24);
 	ctx->pm4[ctx->pm4_cdwords++] = ctx->radeon->fence;
 	ctx->pm4[ctx->pm4_cdwords++] = 0;
-	ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_NOP, 0);
+	ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_NOP, 0, 0);
 	ctx->pm4[ctx->pm4_cdwords++] = 0;
 	r600_context_bo_reloc(ctx, &ctx->pm4[ctx->pm4_cdwords - 1], ctx->radeon->fence_bo);
 
@@ -1282,6 +1282,7 @@ static boolean r600_query_result(struct r600_context *ctx, struct r600_query *qu
 void r600_query_begin(struct r600_context *ctx, struct r600_query *query)
 {
 	unsigned required_space;
+	int num_backends = r600_get_num_backends(ctx->radeon);
 
 	/* query request needs 6/8 dwords for begin + 6/8 dwords for end */
 	if (query->type == PIPE_QUERY_TIME_ELAPSED)
@@ -1300,21 +1301,39 @@ void r600_query_begin(struct r600_context *ctx, struct r600_query *query)
 		r600_query_result(ctx, query, TRUE);
 	}
 
+	if (query->type == PIPE_QUERY_OCCLUSION_COUNTER &&
+	    num_backends > 0 && num_backends < ctx->max_db) {
+		/* as per info on ZPASS the driver must set the unusued DB top bits */
+		u32 *results;
+		int i;
+
+		results = r600_bo_map(ctx->radeon, query->buffer, PB_USAGE_DONTBLOCK | PB_USAGE_CPU_WRITE, NULL);
+		if (results) {
+			memset(results + (query->num_results * 4), 0, ctx->max_db * 4 * 4);
+			
+			for (i = num_backends; i < ctx->max_db; i++) {
+				results[(i * 4)+1] = 0x80000000;
+				results[(i * 4)+3] = 0x80000000;
+			}
+			r600_bo_unmap(ctx->radeon, query->buffer);
+		}
+	}
+	
 	/* emit begin query */
 	if (query->type == PIPE_QUERY_TIME_ELAPSED) {
-		ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_EVENT_WRITE_EOP, 4);
+		ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_EVENT_WRITE_EOP, 4, 0);
 		ctx->pm4[ctx->pm4_cdwords++] = EVENT_TYPE(EVENT_TYPE_CACHE_FLUSH_AND_INV_TS_EVENT) | EVENT_INDEX(5);
 		ctx->pm4[ctx->pm4_cdwords++] = query->num_results*4 + r600_bo_offset(query->buffer);
 		ctx->pm4[ctx->pm4_cdwords++] = (3 << 29);
 		ctx->pm4[ctx->pm4_cdwords++] = 0;
 		ctx->pm4[ctx->pm4_cdwords++] = 0;
 	} else {
-		ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_EVENT_WRITE, 2);
+		ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_EVENT_WRITE, 2, 0);
 		ctx->pm4[ctx->pm4_cdwords++] = EVENT_TYPE(EVENT_TYPE_ZPASS_DONE) | EVENT_INDEX(1);
 		ctx->pm4[ctx->pm4_cdwords++] = query->num_results*4 + r600_bo_offset(query->buffer);
 		ctx->pm4[ctx->pm4_cdwords++] = 0;
 	}
-	ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_NOP, 0);
+	ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_NOP, 0, 0);
 	ctx->pm4[ctx->pm4_cdwords++] = 0;
 	r600_context_bo_reloc(ctx, &ctx->pm4[ctx->pm4_cdwords - 1], query->buffer);
 
@@ -1327,19 +1346,19 @@ void r600_query_end(struct r600_context *ctx, struct r600_query *query)
 {
 	/* emit begin query */
 	if (query->type == PIPE_QUERY_TIME_ELAPSED) {
-		ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_EVENT_WRITE_EOP, 4);
+		ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_EVENT_WRITE_EOP, 4, 0);
 		ctx->pm4[ctx->pm4_cdwords++] = EVENT_TYPE(EVENT_TYPE_CACHE_FLUSH_AND_INV_TS_EVENT) | EVENT_INDEX(5);
 		ctx->pm4[ctx->pm4_cdwords++] = query->num_results*4 + 8 + r600_bo_offset(query->buffer);
 		ctx->pm4[ctx->pm4_cdwords++] = (3 << 29);
 		ctx->pm4[ctx->pm4_cdwords++] = 0;
 		ctx->pm4[ctx->pm4_cdwords++] = 0;
 	} else {
-		ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_EVENT_WRITE, 2);
+		ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_EVENT_WRITE, 2, 0);
 		ctx->pm4[ctx->pm4_cdwords++] = EVENT_TYPE(EVENT_TYPE_ZPASS_DONE) | EVENT_INDEX(1);
 		ctx->pm4[ctx->pm4_cdwords++] = query->num_results*4 + 8 + r600_bo_offset(query->buffer);
 		ctx->pm4[ctx->pm4_cdwords++] = 0;
 	}
-	ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_NOP, 0);
+	ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_NOP, 0, 0);
 	ctx->pm4[ctx->pm4_cdwords++] = 0;
 	r600_context_bo_reloc(ctx, &ctx->pm4[ctx->pm4_cdwords - 1], query->buffer);
 
@@ -1349,6 +1368,28 @@ void r600_query_end(struct r600_context *ctx, struct r600_query *query)
 	ctx->num_query_running--;
 }
 
+void r600_query_predication(struct r600_context *ctx, struct r600_query *query, int operation,
+			    int flag_wait)
+{
+	ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_SET_PREDICATION, 1, 0);
+
+	if (operation == PREDICATION_OP_CLEAR) {
+		ctx->pm4[ctx->pm4_cdwords++] = 0;
+		ctx->pm4[ctx->pm4_cdwords++] = PRED_OP(PREDICATION_OP_CLEAR);
+	} else {
+		int results_base = query->num_results - (4 * ctx->max_db);
+
+		if (results_base < 0)
+			results_base = 0;
+
+		ctx->pm4[ctx->pm4_cdwords++] = results_base*4 + r600_bo_offset(query->buffer);
+		ctx->pm4[ctx->pm4_cdwords++] = PRED_OP(operation) | (flag_wait ? PREDICATION_HINT_WAIT : PREDICATION_HINT_NOWAIT_DRAW) | PREDICATION_DRAW_VISIBLE;
+		ctx->pm4[ctx->pm4_cdwords++] = PKT3(PKT3_NOP, 0, 0);
+		ctx->pm4[ctx->pm4_cdwords++] = 0;
+		r600_context_bo_reloc(ctx, &ctx->pm4[ctx->pm4_cdwords - 1], query->buffer);
+	}
+}
+
 struct r600_query *r600_context_query_create(struct r600_context *ctx, unsigned query_type)
 {
 	struct r600_query *query;
diff --git a/src/gallium/winsys/r600/drm/r600_priv.h b/src/gallium/winsys/r600/drm/r600_priv.h
index 2d91cd97d68..41c5ee02c38 100644
--- a/src/gallium/winsys/r600/drm/r600_priv.h
+++ b/src/gallium/winsys/r600/drm/r600_priv.h
@@ -50,6 +50,8 @@ struct radeon {
 	unsigned			*cfence;
 	struct r600_bo			*fence_bo;
 	unsigned			clock_crystal_freq;
+	unsigned			num_backends;
+	unsigned                        minor_version;
 };
 
 struct r600_reg {
diff --git a/src/gallium/winsys/r600/drm/r600d.h b/src/gallium/winsys/r600/drm/r600d.h
index 1c1ac76fe69..cb12865ff01 100644
--- a/src/gallium/winsys/r600/drm/r600d.h
+++ b/src/gallium/winsys/r600/drm/r600d.h
@@ -105,6 +105,18 @@
 		 * 5 - TS events
 		 */
 
+#define PREDICATION_OP_CLEAR 0x0
+#define PREDICATION_OP_ZPASS 0x1
+#define PREDICATION_OP_PRIMCOUNT 0x2
+
+#define PRED_OP(x) ((x) << 16)
+
+#define PREDICATION_HINT_WAIT (0 << 12)
+#define PREDICATION_HINT_NOWAIT_DRAW (1 << 12)
+
+#define PREDICATION_DRAW_NOT_VISIBLE (0 << 8)
+#define PREDICATION_DRAW_VISIBLE (1 << 8)
+
 #define PKT_TYPE_S(x)                   (((x) & 0x3) << 30)
 #define PKT_TYPE_G(x)                   (((x) >> 30) & 0x3)
 #define PKT_TYPE_C                      0x3FFFFFFF
@@ -117,8 +129,9 @@
 #define PKT3_IT_OPCODE_S(x)             (((x) & 0xFF) << 8)
 #define PKT3_IT_OPCODE_G(x)             (((x) >> 8) & 0xFF)
 #define PKT3_IT_OPCODE_C                0xFFFF00FF
+#define PKT3_PRED_S(x)               (((x) >> 0) & 0x1)
 #define PKT0(index, count) (PKT_TYPE_S(0) | PKT0_BASE_INDEX_S(index) | PKT_COUNT_S(count))
-#define PKT3(op, count) (PKT_TYPE_S(3) | PKT3_IT_OPCODE_S(op) | PKT_COUNT_S(count))
+#define PKT3(op, count, predicate) (PKT_TYPE_S(3) | PKT3_IT_OPCODE_S(op) | PKT_COUNT_S(count) | PKT3_PRED_S(predicate))
 
 /* Registers */
 #define R_0280A0_CB_COLOR0_INFO                      0x0280A0
diff --git a/src/gallium/winsys/radeon/drm/radeon_drm_common.c b/src/gallium/winsys/radeon/drm/radeon_drm_common.c
index 4676c2a1ea9..72c2ff11125 100644
--- a/src/gallium/winsys/radeon/drm/radeon_drm_common.c
+++ b/src/gallium/winsys/radeon/drm/radeon_drm_common.c
@@ -96,10 +96,13 @@ static void do_ioctls(struct radeon_drm_winsys *winsys)
      * we don't actually use the info for anything yet. */
 
     version = drmGetVersion(winsys->fd);
-    if (version->version_major != 2) {
+    if (version->version_major != 2 ||
+        version->version_minor < 3) {
         fprintf(stderr, "%s: DRM version is %d.%d.%d but this driver is "
-                "only compatible with 2.x.x\n", __FUNCTION__,
-                version->version_major, version->version_minor,
+                "only compatible with 2.3.x (kernel 2.6.34) and later.\n",
+                __FUNCTION__,
+                version->version_major,
+                version->version_minor,
                 version->version_patchlevel);
         drmFreeVersion(version);
         exit(1);
@@ -157,6 +160,8 @@ static void do_ioctls(struct radeon_drm_winsys *winsys)
     winsys->vram_size = gem_info.vram_size;
 
     drmFreeVersion(version);
+
+    winsys->num_cpus = sysconf(_SC_NPROCESSORS_ONLN);
 }
 
 static void radeon_winsys_destroy(struct r300_winsys_screen *rws)
@@ -190,10 +195,6 @@ static uint32_t radeon_get_value(struct r300_winsys_screen *rws,
         return ws->drm_minor;
     case R300_VID_DRM_PATCHLEVEL:
         return ws->drm_patchlevel;
-    case R300_VID_DRM_2_1_0:
-        return ws->drm_major*100 + ws->drm_minor >= 201;
-    case R300_VID_DRM_2_3_0:
-        return ws->drm_major*100 + ws->drm_minor >= 203;
     case R300_VID_DRM_2_6_0:
         return ws->drm_major*100 + ws->drm_minor >= 206;
     case R300_VID_DRM_2_8_0:
diff --git a/src/gallium/winsys/radeon/drm/radeon_drm_cs.c b/src/gallium/winsys/radeon/drm/radeon_drm_cs.c
index a38b01048b2..9ee800f5950 100644
--- a/src/gallium/winsys/radeon/drm/radeon_drm_cs.c
+++ b/src/gallium/winsys/radeon/drm/radeon_drm_cs.c
@@ -384,7 +384,7 @@ void radeon_drm_cs_flush(struct r300_winsys_cs *rcs)
         for (i = 0; i < crelocs; i++)
             p_atomic_inc(&cs->csc->relocs_bo[i]->num_active_ioctls);
 
-        if (debug_get_option_thread()) {
+        if (cs->ws->num_cpus > 1 && debug_get_option_thread()) {
             cs->thread = pipe_thread_create(radeon_drm_cs_emit_ioctl, cs->csc);
             assert(cs->thread);
         } else {
diff --git a/src/gallium/winsys/radeon/drm/radeon_winsys.h b/src/gallium/winsys/radeon/drm/radeon_winsys.h
index f8a89abcfe4..9ecbb074572 100644
--- a/src/gallium/winsys/radeon/drm/radeon_winsys.h
+++ b/src/gallium/winsys/radeon/drm/radeon_winsys.h
@@ -46,6 +46,7 @@ struct radeon_drm_winsys {
     uint32_t z_pipes;       /* Z pipe count (rv530 only) */
     uint32_t gart_size;     /* GART size. */
     uint32_t vram_size;     /* VRAM size. */
+    uint32_t num_cpus;      /* Number of CPUs. */
 
     unsigned drm_major;
     unsigned drm_minor;