r600g: move queries to drivers/radeon

Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Tom Stellard <thomas.stellard@amd.com>

src/gallium/drivers/r600/Makefile.sources

@@ -4,7 +4,6 @@ C_SOURCES = \
 	r600_hw_context.c \
 	r600_isa.c \
 	r600_pipe.c \
-	r600_query.c \
 	r600_resource.c \
 	r600_shader.c \
 	r600_state.c \

src/gallium/drivers/r600/r600_blit.c

@@ -54,7 +54,7 @@ static void r600_blitter_begin(struct pipe_context *ctx, enum r600_blitter_op op
 {
 	struct r600_context *rctx = (struct r600_context *)ctx;
 
-	r600_suspend_nontimer_queries(rctx);
+	r600_suspend_nontimer_queries(&rctx->b);
 
 	util_blitter_save_vertex_buffer_slot(rctx->blitter, rctx->vertex_buffer_state.vb);
 	util_blitter_save_vertex_elements(rctx->blitter, rctx->vertex_fetch_shader.cso);
@@ -86,18 +86,18 @@ static void r600_blitter_begin(struct pipe_context *ctx, enum r600_blitter_op op
 			(struct pipe_sampler_view**)rctx->samplers[PIPE_SHADER_FRAGMENT].views.views);
 	}
 
-	if ((op & R600_DISABLE_RENDER_COND) && rctx->current_render_cond) {
+	if ((op & R600_DISABLE_RENDER_COND) && rctx->b.current_render_cond) {
            util_blitter_save_render_condition(rctx->blitter,
-                                              rctx->current_render_cond,
-                                              rctx->current_render_cond_cond,
-                                              rctx->current_render_cond_mode);
+                                              rctx->b.current_render_cond,
+                                              rctx->b.current_render_cond_cond,
+                                              rctx->b.current_render_cond_mode);
         }
 }
 
 static void r600_blitter_end(struct pipe_context *ctx)
 {
 	struct r600_context *rctx = (struct r600_context *)ctx;
-        r600_resume_nontimer_queries(rctx);
+        r600_resume_nontimer_queries(&rctx->b);
 }
 
 static unsigned u_max_sample(struct pipe_resource *r)

src/gallium/drivers/r600/r600_hw_context.c

@@ -29,94 +29,11 @@
 #include <errno.h>
 #include <unistd.h>
 
-/* Get backends mask */
-void r600_get_backend_mask(struct r600_context *ctx)
-{
-	struct radeon_winsys_cs *cs = ctx->b.rings.gfx.cs;
-	struct r600_resource *buffer;
-	uint32_t *results;
-	unsigned num_backends = ctx->screen->b.info.r600_num_backends;
-	unsigned i, mask = 0;
-	uint64_t va;
-
-	/* if backend_map query is supported by the kernel */
-	if (ctx->screen->b.info.r600_backend_map_valid) {
-		unsigned num_tile_pipes = ctx->screen->b.info.r600_num_tile_pipes;
-		unsigned backend_map = ctx->screen->b.info.r600_backend_map;
-		unsigned item_width, item_mask;
-
-		if (ctx->b.chip_class >= EVERGREEN) {
-			item_width = 4;
-			item_mask = 0x7;
-		} else {
-			item_width = 2;
-			item_mask = 0x3;
-		}
-
-		while(num_tile_pipes--) {
-			i = backend_map & item_mask;
-			mask |= (1<<i);
-			backend_map >>= item_width;
-		}
-		if (mask != 0) {
-			ctx->backend_mask = mask;
-			return;
-		}
-	}
-
-	/* otherwise backup path for older kernels */
-
-	/* create buffer for event data */
-	buffer = (struct r600_resource*)
-		pipe_buffer_create(&ctx->screen->b.b, PIPE_BIND_CUSTOM,
-				   PIPE_USAGE_STAGING, ctx->max_db*16);
-	if (!buffer)
-		goto err;
-	va = r600_resource_va(&ctx->screen->b.b, (void*)buffer);
-
-	/* initialize buffer with zeroes */
-	results = r600_buffer_map_sync_with_rings(&ctx->b, buffer, PIPE_TRANSFER_WRITE);
-	if (results) {
-		memset(results, 0, ctx->max_db * 4 * 4);
-		ctx->b.ws->buffer_unmap(buffer->cs_buf);
-
-		/* emit EVENT_WRITE for ZPASS_DONE */
-		cs->buf[cs->cdw++] = PKT3(PKT3_EVENT_WRITE, 2, 0);
-		cs->buf[cs->cdw++] = EVENT_TYPE(EVENT_TYPE_ZPASS_DONE) | EVENT_INDEX(1);
-		cs->buf[cs->cdw++] = va;
-		cs->buf[cs->cdw++] = (va >> 32UL) & 0xFF;
-
-		cs->buf[cs->cdw++] = PKT3(PKT3_NOP, 0, 0);
-		cs->buf[cs->cdw++] = r600_context_bo_reloc(&ctx->b, &ctx->b.rings.gfx, buffer, RADEON_USAGE_WRITE);
-
-		/* analyze results */
-		results = r600_buffer_map_sync_with_rings(&ctx->b, buffer, PIPE_TRANSFER_READ);
-		if (results) {
-			for(i = 0; i < ctx->max_db; i++) {
-				/* at least highest bit will be set if backend is used */
-				if (results[i*4 + 1])
-					mask |= (1<<i);
-			}
-			ctx->b.ws->buffer_unmap(buffer->cs_buf);
-		}
-	}
-
-	pipe_resource_reference((struct pipe_resource**)&buffer, NULL);
-
-	if (mask != 0) {
-		ctx->backend_mask = mask;
-		return;
-	}
-
-err:
-	/* fallback to old method - set num_backends lower bits to 1 */
-	ctx->backend_mask = (~((uint32_t)0))>>(32-num_backends);
-	return;
-}
 
 void r600_need_cs_space(struct r600_context *ctx, unsigned num_dw,
 			boolean count_draw_in)
 {
+
 	if (!ctx->b.ws->cs_memory_below_limit(ctx->b.rings.gfx.cs, ctx->b.vram, ctx->b.gtt)) {
 		ctx->b.gtt = 0;
 		ctx->b.vram = 0;
@@ -151,7 +68,7 @@ void r600_need_cs_space(struct r600_context *ctx, unsigned num_dw,
 	}
 
 	/* Count in queries_suspend. */
-	num_dw += ctx->num_cs_dw_nontimer_queries_suspend;
+	num_dw += ctx->b.num_cs_dw_nontimer_queries_suspend;
 
 	/* Count in streamout_end at the end of CS. */
 	if (ctx->b.streamout.begin_emitted) {
@@ -159,7 +76,7 @@ void r600_need_cs_space(struct r600_context *ctx, unsigned num_dw,
 	}
 
 	/* Count in render_condition(NULL) at the end of CS. */
-	if (ctx->predicate_drawing) {
+	if (ctx->b.predicate_drawing) {
 		num_dw += 3;
 	}
 
@@ -317,13 +234,13 @@ void r600_context_flush(struct r600_context *ctx, unsigned flags)
 {
 	struct radeon_winsys_cs *cs = ctx->b.rings.gfx.cs;
 
-	ctx->nontimer_queries_suspended = false;
+	ctx->b.nontimer_queries_suspended = false;
 	ctx->b.streamout.suspended = false;
 
 	/* suspend queries */
-	if (ctx->num_cs_dw_nontimer_queries_suspend) {
-		r600_suspend_nontimer_queries(ctx);
-		ctx->nontimer_queries_suspended = true;
+	if (ctx->b.num_cs_dw_nontimer_queries_suspend) {
+		r600_suspend_nontimer_queries(&ctx->b);
+		ctx->b.nontimer_queries_suspended = true;
 	}
 
 	if (ctx->b.streamout.begin_emitted) {
@@ -421,8 +338,8 @@ void r600_begin_new_cs(struct r600_context *ctx)
 	}
 
 	/* resume queries */
-	if (ctx->nontimer_queries_suspended) {
-		r600_resume_nontimer_queries(ctx);
+	if (ctx->b.nontimer_queries_suspended) {
+		r600_resume_nontimer_queries(&ctx->b);
 	}
 
 	/* Re-emit the draw state. */

src/gallium/drivers/r600/r600_pipe.c

@@ -78,10 +78,10 @@ static void r600_flush(struct pipe_context *ctx, unsigned flags)
 
 	rctx->b.rings.gfx.flushing = true;
 	/* Disable render condition. */
-	if (rctx->current_render_cond) {
-		render_cond = rctx->current_render_cond;
-		render_cond_cond = rctx->current_render_cond_cond;
-		render_cond_mode = rctx->current_render_cond_mode;
+	if (rctx->b.current_render_cond) {
+		render_cond = rctx->b.current_render_cond;
+		render_cond_cond = rctx->b.current_render_cond_cond;
+		render_cond_mode = rctx->b.current_render_cond_mode;
 		ctx->render_condition(ctx, NULL, FALSE, 0);
 	}
 
@@ -210,10 +210,7 @@ static struct pipe_context *r600_create_context(struct pipe_screen *screen, void
 	rctx->screen = rscreen;
 	rctx->keep_tiling_flags = rscreen->b.info.drm_minor >= 12;
 
-	LIST_INITHEAD(&rctx->active_nontimer_queries);
-
 	r600_init_blit_functions(rctx);
-	r600_init_query_functions(rctx);
 	r600_init_context_resource_functions(rctx);
 
 	if (rscreen->b.info.has_uvd) {
@@ -231,7 +228,6 @@ static struct pipe_context *r600_create_context(struct pipe_screen *screen, void
 	case R700:
 		r600_init_state_functions(rctx);
 		r600_init_atom_start_cs(rctx);
-		rctx->max_db = 4;
 		rctx->custom_dsa_flush = r600_create_db_flush_dsa(rctx);
 		rctx->custom_blend_resolve = rctx->b.chip_class == R700 ? r700_create_resolve_blend(rctx)
 								      : r600_create_resolve_blend(rctx);
@@ -247,7 +243,6 @@ static struct pipe_context *r600_create_context(struct pipe_screen *screen, void
 		evergreen_init_state_functions(rctx);
 		evergreen_init_atom_start_cs(rctx);
 		evergreen_init_atom_start_compute_cs(rctx);
-		rctx->max_db = 8;
 		rctx->custom_dsa_flush = evergreen_create_db_flush_dsa(rctx);
 		rctx->custom_blend_resolve = evergreen_create_resolve_blend(rctx);
 		rctx->custom_blend_decompress = evergreen_create_decompress_blend(rctx);
@@ -298,7 +293,7 @@ static struct pipe_context *r600_create_context(struct pipe_screen *screen, void
 	rctx->blitter->draw_rectangle = r600_draw_rectangle;
 
 	r600_begin_new_cs(rctx);
-	r600_get_backend_mask(rctx); /* this emits commands and must be last */
+	r600_query_init_backend_mask(&rctx->b); /* this emits commands and must be last */
 
 	rctx->dummy_pixel_shader =
 		util_make_fragment_cloneinput_shader(&rctx->b.b, 0,

src/gallium/drivers/r600/r600_pipe.h

@@ -62,11 +62,6 @@
 #define R600_BIG_ENDIAN 0
 #endif
 
-#define R600_QUERY_DRAW_CALLS		(PIPE_QUERY_DRIVER_SPECIFIC + 0)
-#define R600_QUERY_REQUESTED_VRAM	(PIPE_QUERY_DRIVER_SPECIFIC + 1)
-#define R600_QUERY_REQUESTED_GTT	(PIPE_QUERY_DRIVER_SPECIFIC + 2)
-#define R600_QUERY_BUFFER_WAIT_TIME	(PIPE_QUERY_DRIVER_SPECIFIC + 3)
-
 struct r600_context;
 struct r600_bytecode;
 struct r600_shader_key;
@@ -364,34 +359,6 @@ struct r600_shader_state {
 	struct r600_pipe_shader_selector *shader;
 };
 
-struct r600_query_buffer {
-	/* The buffer where query results are stored. */
-	struct r600_resource			*buf;
-	/* Offset of the next free result after current query data */
-	unsigned				results_end;
-	/* If a query buffer is full, a new buffer is created and the old one
-	 * is put in here. When we calculate the result, we sum up the samples
-	 * from all buffers. */
-	struct r600_query_buffer		*previous;
-};
-
-struct r600_query {
-	/* The query buffer and how many results are in it. */
-	struct r600_query_buffer		buffer;
-	/* The type of query */
-	unsigned				type;
-	/* Size of the result in memory for both begin_query and end_query,
-	 * this can be one or two numbers, or it could even be a size of a structure. */
-	unsigned				result_size;
-	/* The number of dwords for begin_query or end_query. */
-	unsigned				num_cs_dw;
-	/* linked list of queries */
-	struct list_head			list;
-	/* for custom non-GPU queries */
-	uint64_t begin_result;
-	uint64_t end_result;
-};
-
 struct r600_context {
 	struct r600_common_context	b;
 	struct r600_screen		*screen;
@@ -404,8 +371,6 @@ struct r600_context {
 	boolean				keep_tiling_flags;
 	unsigned			default_ps_gprs, default_vs_gprs;
 	unsigned			r6xx_num_clause_temp_gprs;
-	unsigned			backend_mask;
-	unsigned			max_db; /* for OQ */
 
 	/* Miscellaneous state objects. */
 	void				*custom_dsa_flush;
@@ -478,25 +443,6 @@ struct r600_context {
 	int				last_primitive_type; /* Last primitive type used in draw_vbo. */
 	int				last_start_instance;
 
-	/* Queries. */
-	/* The list of active queries. Only one query of each type can be active. */
-	int				num_occlusion_queries;
-	int				num_pipelinestat_queries;
-	/* Keep track of non-timer queries, because they should be suspended
-	 * during context flushing.
-	 * The timer queries (TIME_ELAPSED) shouldn't be suspended. */
-	struct list_head		active_nontimer_queries;
-	unsigned			num_cs_dw_nontimer_queries_suspend;
-	/* If queries have been suspended. */
-	bool				nontimer_queries_suspended;
-	unsigned			num_draw_calls;
-
-	/* Render condition. */
-	struct pipe_query		*current_render_cond;
-	unsigned			current_render_cond_mode;
-	boolean				current_render_cond_cond;
-	boolean				predicate_drawing;
-
 	void				*sb_context;
 	struct r600_isa		*isa;
 };
@@ -589,12 +535,6 @@ void r600_decompress_color_textures(struct r600_context *rctx,
 /* r600_pipe.c */
 const char * r600_llvm_gpu_string(enum radeon_family family);
 
-
-/* r600_query.c */
-void r600_init_query_functions(struct r600_context *rctx);
-void r600_suspend_nontimer_queries(struct r600_context *ctx);
-void r600_resume_nontimer_queries(struct r600_context *ctx);
-
 /* r600_resource.c */
 void r600_init_context_resource_functions(struct r600_context *r600);
 
@@ -628,7 +568,6 @@ boolean r600_is_format_supported(struct pipe_screen *screen,
 void r600_update_db_shader_control(struct r600_context * rctx);
 
 /* r600_hw_context.c */
-void r600_get_backend_mask(struct r600_context *ctx);
 void r600_context_flush(struct r600_context *ctx, unsigned flags);
 void r600_begin_new_cs(struct r600_context *ctx);
 void r600_flush_emit(struct r600_context *ctx);

src/gallium/drivers/r600/r600_state_common.c

@@ -1339,10 +1339,10 @@ static void r600_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info
 	}
 
 	/* Draw packets. */
-	cs->buf[cs->cdw++] = PKT3(PKT3_NUM_INSTANCES, 0, rctx->predicate_drawing);
+	cs->buf[cs->cdw++] = PKT3(PKT3_NUM_INSTANCES, 0, rctx->b.predicate_drawing);
 	cs->buf[cs->cdw++] = info.instance_count;
 	if (info.indexed) {
-		cs->buf[cs->cdw++] = PKT3(PKT3_INDEX_TYPE, 0, rctx->predicate_drawing);
+		cs->buf[cs->cdw++] = PKT3(PKT3_INDEX_TYPE, 0, rctx->b.predicate_drawing);
 		cs->buf[cs->cdw++] = ib.index_size == 4 ?
 					(VGT_INDEX_32 | (R600_BIG_ENDIAN ? VGT_DMA_SWAP_32_BIT : 0)) :
 					(VGT_INDEX_16 | (R600_BIG_ENDIAN ? VGT_DMA_SWAP_16_BIT : 0));
@@ -1350,19 +1350,19 @@ static void r600_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info
 		if (ib.user_buffer) {
 			unsigned size_bytes = info.count*ib.index_size;
 			unsigned size_dw = align(size_bytes, 4) / 4;
-			cs->buf[cs->cdw++] = PKT3(PKT3_DRAW_INDEX_IMMD, 1 + size_dw, rctx->predicate_drawing);
+			cs->buf[cs->cdw++] = PKT3(PKT3_DRAW_INDEX_IMMD, 1 + size_dw, rctx->b.predicate_drawing);
 			cs->buf[cs->cdw++] = info.count;
 			cs->buf[cs->cdw++] = V_0287F0_DI_SRC_SEL_IMMEDIATE;
 			memcpy(cs->buf+cs->cdw, ib.user_buffer, size_bytes);
 			cs->cdw += size_dw;
 		} else {
 			uint64_t va = r600_resource_va(ctx->screen, ib.buffer) + ib.offset;
-			cs->buf[cs->cdw++] = PKT3(PKT3_DRAW_INDEX, 3, rctx->predicate_drawing);
+			cs->buf[cs->cdw++] = PKT3(PKT3_DRAW_INDEX, 3, rctx->b.predicate_drawing);
 			cs->buf[cs->cdw++] = va;
 			cs->buf[cs->cdw++] = (va >> 32UL) & 0xFF;
 			cs->buf[cs->cdw++] = info.count;
 			cs->buf[cs->cdw++] = V_0287F0_DI_SRC_SEL_DMA;
-			cs->buf[cs->cdw++] = PKT3(PKT3_NOP, 0, rctx->predicate_drawing);
+			cs->buf[cs->cdw++] = PKT3(PKT3_NOP, 0, rctx->b.predicate_drawing);
 			cs->buf[cs->cdw++] = r600_context_bo_reloc(&rctx->b, &rctx->b.rings.gfx, (struct r600_resource*)ib.buffer, RADEON_USAGE_READ);
 		}
 	} else {
@@ -1383,7 +1383,7 @@ static void r600_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info
 			cs->buf[cs->cdw++] = r600_context_bo_reloc(&rctx->b, &rctx->b.rings.gfx, t->buf_filled_size, RADEON_USAGE_READ);
 		}
 
-		cs->buf[cs->cdw++] = PKT3(PKT3_DRAW_INDEX_AUTO, 1, rctx->predicate_drawing);
+		cs->buf[cs->cdw++] = PKT3(PKT3_DRAW_INDEX_AUTO, 1, rctx->b.predicate_drawing);
 		cs->buf[cs->cdw++] = info.count;
 		cs->buf[cs->cdw++] = V_0287F0_DI_SRC_SEL_AUTO_INDEX |
 					(info.count_from_stream_output ? S_0287F0_USE_OPAQUE(1) : 0);
@@ -1416,7 +1416,7 @@ static void r600_draw_vbo(struct pipe_context *ctx, const struct pipe_draw_info
 	}
 
 	pipe_resource_reference(&ib.buffer, NULL);
-	rctx->num_draw_calls++;
+	rctx->b.num_draw_calls++;
 }
 
 void r600_draw_rectangle(struct blitter_context *blitter,
@@ -2127,6 +2127,22 @@ static void r600_invalidate_buffer(struct pipe_context *ctx, struct pipe_resourc
 	/* XXX TODO: texture buffer objects */
 }
 
+static void r600_set_occlusion_query_state(struct pipe_context *ctx, bool enable)
+{
+	struct r600_context *rctx = (struct r600_context*)ctx;
+
+	if (rctx->db_misc_state.occlusion_query_enabled != enable) {
+		rctx->db_misc_state.occlusion_query_enabled = enable;
+		rctx->db_misc_state.atom.dirty = true;
+	}
+}
+
+static void r600_need_gfx_cs_space(struct pipe_context *ctx, unsigned num_dw,
+                                   bool include_draw_vbo)
+{
+	r600_need_cs_space((struct r600_context*)ctx, num_dw, include_draw_vbo);
+}
+
 /* keep this at the end of this file, please */
 void r600_init_common_state_functions(struct r600_context *rctx)
 {
@@ -2163,6 +2179,8 @@ void r600_init_common_state_functions(struct r600_context *rctx)
 	rctx->b.b.surface_destroy = r600_surface_destroy;
 	rctx->b.b.draw_vbo = r600_draw_vbo;
 	rctx->b.invalidate_buffer = r600_invalidate_buffer;
+	rctx->b.set_occlusion_query_state = r600_set_occlusion_query_state;
+	rctx->b.need_gfx_cs_space = r600_need_gfx_cs_space;
 }
 
 void r600_trace_emit(struct r600_context *rctx)

src/gallium/drivers/radeon/Makefile.sources

@@ -1,6 +1,7 @@
 C_SOURCES := \
 	r600_buffer_common.c \
 	r600_pipe_common.c \
+        r600_query.c \
 	r600_streamout.c \
         r600_texture.c \
 	radeon_uvd.c

src/gallium/drivers/radeon/r600_pipe_common.c

@@ -228,11 +228,14 @@ bool r600_common_context_init(struct r600_common_context *rctx,
 			 sizeof(struct r600_transfer), 64,
 			 UTIL_SLAB_SINGLETHREADED);
 
+	rctx->screen = rscreen;
 	rctx->ws = rscreen->ws;
 	rctx->family = rscreen->family;
 	rctx->chip_class = rscreen->chip_class;
+	rctx->max_db = rscreen->chip_class >= EVERGREEN ? 8 : 4;
 
 	r600_streamout_init(rctx);
+	r600_query_init(rctx);
 
 	rctx->allocator_so_filled_size = u_suballocator_create(&rctx->b, 4096, 4,
 							       0, PIPE_USAGE_STATIC, TRUE);

src/gallium/drivers/radeon/r600_pipe_common.h

@@ -36,6 +36,7 @@
 
 #include "../../winsys/radeon/drm/radeon_winsys.h"
 
+#include "util/u_double_list.h"
 #include "util/u_range.h"
 #include "util/u_slab.h"
 #include "util/u_suballoc.h"
@@ -45,6 +46,11 @@
 #define R600_RESOURCE_FLAG_FLUSHED_DEPTH	(PIPE_RESOURCE_FLAG_DRV_PRIV << 1)
 #define R600_RESOURCE_FLAG_FORCE_TILING		(PIPE_RESOURCE_FLAG_DRV_PRIV << 2)
 
+#define R600_QUERY_DRAW_CALLS		(PIPE_QUERY_DRIVER_SPECIFIC + 0)
+#define R600_QUERY_REQUESTED_VRAM	(PIPE_QUERY_DRIVER_SPECIFIC + 1)
+#define R600_QUERY_REQUESTED_GTT	(PIPE_QUERY_DRIVER_SPECIFIC + 2)
+#define R600_QUERY_BUFFER_WAIT_TIME	(PIPE_QUERY_DRIVER_SPECIFIC + 3)
+
 /* read caches */
 #define R600_CONTEXT_INV_VERTEX_CACHE		(1 << 0)
 #define R600_CONTEXT_INV_TEX_CACHE		(1 << 1)
@@ -225,6 +231,7 @@ struct r600_rings {
 struct r600_common_context {
 	struct pipe_context b; /* base class */
 
+	struct r600_common_screen	*screen;
 	struct radeon_winsys		*ws;
 	enum radeon_family		family;
 	enum chip_class			chip_class;
@@ -244,6 +251,29 @@ struct r600_common_context {
 	/* Additional context states. */
 	unsigned flags; /* flush flags */
 
+	/* Queries. */
+	/* The list of active queries. Only one query of each type can be active. */
+	int				num_occlusion_queries;
+	int				num_pipelinestat_queries;
+	/* Keep track of non-timer queries, because they should be suspended
+	 * during context flushing.
+	 * The timer queries (TIME_ELAPSED) shouldn't be suspended. */
+	struct list_head		active_nontimer_queries;
+	unsigned			num_cs_dw_nontimer_queries_suspend;
+	/* If queries have been suspended. */
+	bool				nontimer_queries_suspended;
+	/* Additional hardware info. */
+	unsigned			backend_mask;
+	unsigned			max_db; /* for OQ */
+	/* Misc stats. */
+	unsigned			num_draw_calls;
+
+	/* Render condition. */
+	struct pipe_query		*current_render_cond;
+	unsigned			current_render_cond_mode;
+	boolean				current_render_cond_cond;
+	boolean				predicate_drawing;
+
 	/* Copy one resource to another using async DMA.
 	 * False is returned if the copy couldn't be done. */
 	boolean (*dma_copy)(struct pipe_context *ctx,
@@ -267,6 +297,13 @@ struct r600_common_context {
 	/* Reallocate the buffer and update all resource bindings where
 	 * the buffer is bound, including all resource descriptors. */
 	void (*invalidate_buffer)(struct pipe_context *ctx, struct pipe_resource *buf);
+
+	/* Enable or disable occlusion queries. */
+	void (*set_occlusion_query_state)(struct pipe_context *ctx, bool enable);
+
+	/* This ensures there is enough space in the command stream. */
+	void (*need_gfx_cs_space)(struct pipe_context *ctx, unsigned num_dw,
+				  bool include_draw_vbo);
 };
 
 /* r600_buffer.c */
@@ -297,6 +334,12 @@ bool r600_can_dump_shader(struct r600_common_screen *rscreen,
 void r600_screen_clear_buffer(struct r600_common_screen *rscreen, struct pipe_resource *dst,
 			      unsigned offset, unsigned size, unsigned value);
 
+/* r600_query.c */
+void r600_query_init(struct r600_common_context *rctx);
+void r600_suspend_nontimer_queries(struct r600_common_context *ctx);
+void r600_resume_nontimer_queries(struct r600_common_context *ctx);
+void r600_query_init_backend_mask(struct r600_common_context *ctx);
+
 /* r600_streamout.c */
 void r600_streamout_buffers_dirty(struct r600_common_context *rctx);
 void r600_set_streamout_targets(struct pipe_context *ctx,

src/gallium/drivers/r600/r600_query.c → src/gallium/drivers/radeon/r600_query.c

@@ -1,5 +1,6 @@
 /*
  * Copyright 2010 Jerome Glisse <glisse@freedesktop.org>
+ * Copyright 2014 Marek Olšák <marek.olsak@amd.com>
  *
  * Permission is hereby granted, free of charge, to any person obtaining a
  * copy of this software and associated documentation files (the "Software"),
@@ -20,10 +21,40 @@
  * OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
  * USE OR OTHER DEALINGS IN THE SOFTWARE.
  */
-#include "r600_pipe.h"
-#include "r600d.h"
+
+#include "r600_cs.h"
 #include "util/u_memory.h"
 
+
+struct r600_query_buffer {
+	/* The buffer where query results are stored. */
+	struct r600_resource			*buf;
+	/* Offset of the next free result after current query data */
+	unsigned				results_end;
+	/* If a query buffer is full, a new buffer is created and the old one
+	 * is put in here. When we calculate the result, we sum up the samples
+	 * from all buffers. */
+	struct r600_query_buffer		*previous;
+};
+
+struct r600_query {
+	/* The query buffer and how many results are in it. */
+	struct r600_query_buffer		buffer;
+	/* The type of query */
+	unsigned				type;
+	/* Size of the result in memory for both begin_query and end_query,
+	 * this can be one or two numbers, or it could even be a size of a structure. */
+	unsigned				result_size;
+	/* The number of dwords for begin_query or end_query. */
+	unsigned				num_cs_dw;
+	/* linked list of queries */
+	struct list_head			list;
+	/* for custom non-GPU queries */
+	uint64_t begin_result;
+	uint64_t end_result;
+};
+
+
 static bool r600_is_timer_query(unsigned type)
 {
 	return type == PIPE_QUERY_TIME_ELAPSED ||
@@ -37,7 +68,7 @@ static bool r600_query_needs_begin(unsigned type)
 	       type != PIPE_QUERY_TIMESTAMP;
 }
 
-static struct r600_resource *r600_new_query_buffer(struct r600_context *ctx, unsigned type)
+static struct r600_resource *r600_new_query_buffer(struct r600_common_context *ctx, unsigned type)
 {
 	unsigned j, i, num_results, buf_size = 4096;
 	uint32_t *results;
@@ -56,13 +87,13 @@ static struct r600_resource *r600_new_query_buffer(struct r600_context *ctx, uns
 	 * usage pattern.
 	 */
 	struct r600_resource *buf = (struct r600_resource*)
-		pipe_buffer_create(&ctx->screen->b.b, PIPE_BIND_CUSTOM,
+		pipe_buffer_create(ctx->b.screen, PIPE_BIND_CUSTOM,
 				   PIPE_USAGE_STAGING, buf_size);
 
 	switch (type) {
 	case PIPE_QUERY_OCCLUSION_COUNTER:
 	case PIPE_QUERY_OCCLUSION_PREDICATE:
-		results = r600_buffer_map_sync_with_rings(&ctx->b, buf, PIPE_TRANSFER_WRITE);
+		results = r600_buffer_map_sync_with_rings(ctx, buf, PIPE_TRANSFER_WRITE);
 		memset(results, 0, buf_size);
 
 		/* Set top bits for unused backends. */
@@ -76,7 +107,7 @@ static struct r600_resource *r600_new_query_buffer(struct r600_context *ctx, uns
 			}
 			results += 4 * ctx->max_db;
 		}
-		ctx->b.ws->buffer_unmap(buf->cs_buf);
+		ctx->ws->buffer_unmap(buf->cs_buf);
 		break;
 	case PIPE_QUERY_TIME_ELAPSED:
 	case PIPE_QUERY_TIMESTAMP:
@@ -86,9 +117,9 @@ static struct r600_resource *r600_new_query_buffer(struct r600_context *ctx, uns
 	case PIPE_QUERY_SO_STATISTICS:
 	case PIPE_QUERY_SO_OVERFLOW_PREDICATE:
 	case PIPE_QUERY_PIPELINE_STATISTICS:
-		results = r600_buffer_map_sync_with_rings(&ctx->b, buf, PIPE_TRANSFER_WRITE);
+		results = r600_buffer_map_sync_with_rings(ctx, buf, PIPE_TRANSFER_WRITE);
 		memset(results, 0, buf_size);
-		ctx->b.ws->buffer_unmap(buf->cs_buf);
+		ctx->ws->buffer_unmap(buf->cs_buf);
 		break;
 	default:
 		assert(0);
@@ -96,11 +127,12 @@ static struct r600_resource *r600_new_query_buffer(struct r600_context *ctx, uns
 	return buf;
 }
 
-static void r600_update_occlusion_query_state(struct r600_context *rctx,
+static void r600_update_occlusion_query_state(struct r600_common_context *rctx,
 					      unsigned type, int diff)
 {
 	if (type == PIPE_QUERY_OCCLUSION_COUNTER ||
 	    type == PIPE_QUERY_OCCLUSION_PREDICATE) {
+		bool old_enable = rctx->num_occlusion_queries != 0;
 		bool enable;
 
 		rctx->num_occlusion_queries += diff;
@@ -108,20 +140,19 @@ static void r600_update_occlusion_query_state(struct r600_context *rctx,
 
 		enable = rctx->num_occlusion_queries != 0;
 
-		if (rctx->db_misc_state.occlusion_query_enabled != enable) {
-			rctx->db_misc_state.occlusion_query_enabled = enable;
-			rctx->db_misc_state.atom.dirty = true;
+		if (enable != old_enable) {
+			rctx->set_occlusion_query_state(&rctx->b, enable);
 		}
 	}
 }
 
-static void r600_emit_query_begin(struct r600_context *ctx, struct r600_query *query)
+static void r600_emit_query_begin(struct r600_common_context *ctx, struct r600_query *query)
 {
-	struct radeon_winsys_cs *cs = ctx->b.rings.gfx.cs;
+	struct radeon_winsys_cs *cs = ctx->rings.gfx.cs;
 	uint64_t va;
 
 	r600_update_occlusion_query_state(ctx, query->type, 1);
-	r600_need_cs_space(ctx, query->num_cs_dw * 2, TRUE);
+	ctx->need_gfx_cs_space(&ctx->b, query->num_cs_dw * 2, TRUE);
 
 	/* Get a new query buffer if needed. */
 	if (query->buffer.results_end + query->result_size > query->buffer.buf->b.b.width0) {
@@ -133,7 +164,7 @@ static void r600_emit_query_begin(struct r600_context *ctx, struct r600_query *q
 	}
 
 	/* emit begin query */
-	va = r600_resource_va(&ctx->screen->b.b, (void*)query->buffer.buf);
+	va = r600_resource_va(ctx->b.screen, (void*)query->buffer.buf);
 	va += query->buffer.results_end;
 
 	switch (query->type) {
@@ -176,24 +207,24 @@ static void r600_emit_query_begin(struct r600_context *ctx, struct r600_query *q
 		assert(0);
 	}
 	cs->buf[cs->cdw++] = PKT3(PKT3_NOP, 0, 0);
-	cs->buf[cs->cdw++] = r600_context_bo_reloc(&ctx->b, &ctx->b.rings.gfx, query->buffer.buf, RADEON_USAGE_WRITE);
+	cs->buf[cs->cdw++] = r600_context_bo_reloc(ctx, &ctx->rings.gfx, query->buffer.buf, RADEON_USAGE_WRITE);
 
 	if (!r600_is_timer_query(query->type)) {
 		ctx->num_cs_dw_nontimer_queries_suspend += query->num_cs_dw;
 	}
 }
 
-static void r600_emit_query_end(struct r600_context *ctx, struct r600_query *query)
+static void r600_emit_query_end(struct r600_common_context *ctx, struct r600_query *query)
 {
-	struct radeon_winsys_cs *cs = ctx->b.rings.gfx.cs;
+	struct radeon_winsys_cs *cs = ctx->rings.gfx.cs;
 	uint64_t va;
 
 	/* The queries which need begin already called this in begin_query. */
 	if (!r600_query_needs_begin(query->type)) {
-		r600_need_cs_space(ctx, query->num_cs_dw, FALSE);
+		ctx->need_gfx_cs_space(&ctx->b, query->num_cs_dw, FALSE);
 	}
 
-	va = r600_resource_va(&ctx->screen->b.b, (void*)query->buffer.buf);
+	va = r600_resource_va(ctx->b.screen, (void*)query->buffer.buf);
 	/* emit end query */
 	switch (query->type) {
 	case PIPE_QUERY_OCCLUSION_COUNTER:
@@ -242,7 +273,7 @@ static void r600_emit_query_end(struct r600_context *ctx, struct r600_query *que
 		assert(0);
 	}
 	cs->buf[cs->cdw++] = PKT3(PKT3_NOP, 0, 0);
-	cs->buf[cs->cdw++] = r600_context_bo_reloc(&ctx->b, &ctx->b.rings.gfx, query->buffer.buf, RADEON_USAGE_WRITE);
+	cs->buf[cs->cdw++] = r600_context_bo_reloc(ctx, &ctx->rings.gfx, query->buffer.buf, RADEON_USAGE_WRITE);
 
 	query->buffer.results_end += query->result_size;
 
@@ -255,13 +286,13 @@ static void r600_emit_query_end(struct r600_context *ctx, struct r600_query *que
 	r600_update_occlusion_query_state(ctx, query->type, -1);
 }
 
-static void r600_emit_query_predication(struct r600_context *ctx, struct r600_query *query,
+static void r600_emit_query_predication(struct r600_common_context *ctx, struct r600_query *query,
 					int operation, bool flag_wait)
 {
-	struct radeon_winsys_cs *cs = ctx->b.rings.gfx.cs;
+	struct radeon_winsys_cs *cs = ctx->rings.gfx.cs;
 
 	if (operation == PREDICATION_OP_CLEAR) {
-		r600_need_cs_space(ctx, 3, FALSE);
+		ctx->need_gfx_cs_space(&ctx->b, 3, FALSE);
 
 		cs->buf[cs->cdw++] = PKT3(PKT3_SET_PREDICATION, 1, 0);
 		cs->buf[cs->cdw++] = 0;
@@ -277,7 +308,7 @@ static void r600_emit_query_predication(struct r600_context *ctx, struct r600_qu
 			count += qbuf->results_end / query->result_size;
 		}
 
-		r600_need_cs_space(ctx, 5 * count, TRUE);
+		ctx->need_gfx_cs_space(&ctx->b, 5 * count, TRUE);
 
 		op = PRED_OP(operation) | PREDICATION_DRAW_VISIBLE |
 				(flag_wait ? PREDICATION_HINT_WAIT : PREDICATION_HINT_NOWAIT_DRAW);
@@ -285,14 +316,14 @@ static void r600_emit_query_predication(struct r600_context *ctx, struct r600_qu
 		/* emit predicate packets for all data blocks */
 		for (qbuf = &query->buffer; qbuf; qbuf = qbuf->previous) {
 			unsigned results_base = 0;
-			uint64_t va = r600_resource_va(&ctx->screen->b.b, &qbuf->buf->b.b);
+			uint64_t va = r600_resource_va(ctx->b.screen, &qbuf->buf->b.b);
 
 			while (results_base < qbuf->results_end) {
 				cs->buf[cs->cdw++] = PKT3(PKT3_SET_PREDICATION, 1, 0);
 				cs->buf[cs->cdw++] = (va + results_base) & 0xFFFFFFFFUL;
 				cs->buf[cs->cdw++] = op | (((va + results_base) >> 32UL) & 0xFF);
 				cs->buf[cs->cdw++] = PKT3(PKT3_NOP, 0, 0);
-				cs->buf[cs->cdw++] = r600_context_bo_reloc(&ctx->b, &ctx->b.rings.gfx, qbuf->buf, RADEON_USAGE_READ);
+				cs->buf[cs->cdw++] = r600_context_bo_reloc(ctx, &ctx->rings.gfx, qbuf->buf, RADEON_USAGE_READ);
 				results_base += query->result_size;
 
 				/* set CONTINUE bit for all packets except the first */
@@ -304,7 +335,7 @@ static void r600_emit_query_predication(struct r600_context *ctx, struct r600_qu
 
 static struct pipe_query *r600_create_query(struct pipe_context *ctx, unsigned query_type)
 {
-	struct r600_context *rctx = (struct r600_context *)ctx;
+	struct r600_common_context *rctx = (struct r600_common_context *)ctx;
 	struct r600_query *query;
 	bool skip_allocation = false;
 
@@ -338,7 +369,7 @@ static struct pipe_query *r600_create_query(struct pipe_context *ctx, unsigned q
 		break;
 	case PIPE_QUERY_PIPELINE_STATISTICS:
 		/* 11 values on EG, 8 on R600. */
-		query->result_size = (rctx->b.chip_class >= EVERGREEN ? 11 : 8) * 16;
+		query->result_size = (rctx->chip_class >= EVERGREEN ? 11 : 8) * 16;
 		query->num_cs_dw = 8;
 		break;
 	/* Non-GPU queries. */
@@ -383,7 +414,7 @@ static void r600_destroy_query(struct pipe_context *ctx, struct pipe_query *quer
 
 static void r600_begin_query(struct pipe_context *ctx, struct pipe_query *query)
 {
-	struct r600_context *rctx = (struct r600_context *)ctx;
+	struct r600_common_context *rctx = (struct r600_common_context *)ctx;
 	struct r600_query *rquery = (struct r600_query *)query;
 	struct r600_query_buffer *prev = rquery->buffer.previous;
 
@@ -402,7 +433,7 @@ static void r600_begin_query(struct pipe_context *ctx, struct pipe_query *query)
 		rquery->begin_result = 0;
 		return;
 	case R600_QUERY_BUFFER_WAIT_TIME:
-		rquery->begin_result = rctx->b.ws->query_value(rctx->b.ws, RADEON_BUFFER_WAIT_TIME_NS);
+		rquery->begin_result = rctx->ws->query_value(rctx->ws, RADEON_BUFFER_WAIT_TIME_NS);
 		return;
 	}
 
@@ -415,8 +446,8 @@ static void r600_begin_query(struct pipe_context *ctx, struct pipe_query *query)
 	}
 
 	/* Obtain a new buffer if the current one can't be mapped without a stall. */
-	if (r600_rings_is_buffer_referenced(&rctx->b, rquery->buffer.buf->cs_buf, RADEON_USAGE_READWRITE) ||
-	    rctx->b.ws->buffer_is_busy(rquery->buffer.buf->buf, RADEON_USAGE_READWRITE)) {
+	if (r600_rings_is_buffer_referenced(rctx, rquery->buffer.buf->cs_buf, RADEON_USAGE_READWRITE) ||
+	    rctx->ws->buffer_is_busy(rquery->buffer.buf->buf, RADEON_USAGE_READWRITE)) {
 		pipe_resource_reference((struct pipe_resource**)&rquery->buffer.buf, NULL);
 		rquery->buffer.buf = r600_new_query_buffer(rctx, rquery->type);
 	}
@@ -433,7 +464,7 @@ static void r600_begin_query(struct pipe_context *ctx, struct pipe_query *query)
 
 static void r600_end_query(struct pipe_context *ctx, struct pipe_query *query)
 {
-	struct r600_context *rctx = (struct r600_context *)ctx;
+	struct r600_common_context *rctx = (struct r600_common_context *)ctx;
 	struct r600_query *rquery = (struct r600_query *)query;
 
 	/* Non-GPU queries. */
@@ -442,13 +473,13 @@ static void r600_end_query(struct pipe_context *ctx, struct pipe_query *query)
 		rquery->end_result = rctx->num_draw_calls;
 		return;
 	case R600_QUERY_REQUESTED_VRAM:
-		rquery->end_result = rctx->b.ws->query_value(rctx->b.ws, RADEON_REQUESTED_VRAM_MEMORY);
+		rquery->end_result = rctx->ws->query_value(rctx->ws, RADEON_REQUESTED_VRAM_MEMORY);
 		return;
 	case R600_QUERY_REQUESTED_GTT:
-		rquery->end_result = rctx->b.ws->query_value(rctx->b.ws, RADEON_REQUESTED_GTT_MEMORY);
+		rquery->end_result = rctx->ws->query_value(rctx->ws, RADEON_REQUESTED_GTT_MEMORY);
 		return;
 	case R600_QUERY_BUFFER_WAIT_TIME:
-		rquery->end_result = rctx->b.ws->query_value(rctx->b.ws, RADEON_BUFFER_WAIT_TIME_NS);
+		rquery->end_result = rctx->ws->query_value(rctx->ws, RADEON_BUFFER_WAIT_TIME_NS);
 		return;
 	}
 
@@ -477,7 +508,7 @@ static unsigned r600_query_read_result(char *map, unsigned start_index, unsigned
 	return 0;
 }
 
-static boolean r600_get_query_buffer_result(struct r600_context *ctx,
+static boolean r600_get_query_buffer_result(struct r600_common_context *ctx,
 					    struct r600_query *query,
 					    struct r600_query_buffer *qbuf,
 					    boolean wait,
@@ -496,7 +527,7 @@ static boolean r600_get_query_buffer_result(struct r600_context *ctx,
 		return TRUE;
 	}
 
-	map = r600_buffer_map_sync_with_rings(&ctx->b, qbuf->buf,
+	map = r600_buffer_map_sync_with_rings(ctx, qbuf->buf,
 						PIPE_TRANSFER_READ |
 						(wait ? 0 : PIPE_TRANSFER_DONTBLOCK));
 	if (!map)
@@ -571,7 +602,7 @@ static boolean r600_get_query_buffer_result(struct r600_context *ctx,
 		}
 		break;
 	case PIPE_QUERY_PIPELINE_STATISTICS:
-		if (ctx->b.chip_class >= EVERGREEN) {
+		if (ctx->chip_class >= EVERGREEN) {
 			while (results_base != qbuf->results_end) {
 				result->pipeline_statistics.ps_invocations +=
 					r600_query_read_result(map + results_base, 0, 22, false);
@@ -639,7 +670,7 @@ static boolean r600_get_query_buffer_result(struct r600_context *ctx,
 		assert(0);
 	}
 
-	ctx->b.ws->buffer_unmap(qbuf->buf->cs_buf);
+	ctx->ws->buffer_unmap(qbuf->buf->cs_buf);
 	return TRUE;
 }
 
@@ -647,7 +678,7 @@ static boolean r600_get_query_result(struct pipe_context *ctx,
 					struct pipe_query *query,
 					boolean wait, union pipe_query_result *result)
 {
-	struct r600_context *rctx = (struct r600_context *)ctx;
+	struct r600_common_context *rctx = (struct r600_common_context *)ctx;
 	struct r600_query *rquery = (struct r600_query *)query;
 	struct r600_query_buffer *qbuf;
 
@@ -662,7 +693,7 @@ static boolean r600_get_query_result(struct pipe_context *ctx,
 	/* Convert the time to expected units. */
 	if (rquery->type == PIPE_QUERY_TIME_ELAPSED ||
 	    rquery->type == PIPE_QUERY_TIMESTAMP) {
-		result->u64 = (1000000 * result->u64) / rctx->screen->b.info.r600_clock_crystal_freq;
+		result->u64 = (1000000 * result->u64) / rctx->screen->info.r600_clock_crystal_freq;
 	}
 	return TRUE;
 }
@@ -672,7 +703,7 @@ static void r600_render_condition(struct pipe_context *ctx,
 				  boolean condition,
 				  uint mode)
 {
-	struct r600_context *rctx = (struct r600_context *)ctx;
+	struct r600_common_context *rctx = (struct r600_common_context *)ctx;
 	struct r600_query *rquery = (struct r600_query *)query;
 	bool wait_flag = false;
 
@@ -711,7 +742,7 @@ static void r600_render_condition(struct pipe_context *ctx,
 	}
 }
 
-void r600_suspend_nontimer_queries(struct r600_context *ctx)
+void r600_suspend_nontimer_queries(struct r600_common_context *ctx)
 {
 	struct r600_query *query;
 
@@ -721,7 +752,7 @@ void r600_suspend_nontimer_queries(struct r600_context *ctx)
 	assert(ctx->num_cs_dw_nontimer_queries_suspend == 0);
 }
 
-void r600_resume_nontimer_queries(struct r600_context *ctx)
+void r600_resume_nontimer_queries(struct r600_common_context *ctx)
 {
 	struct r600_query *query;
 
@@ -732,14 +763,101 @@ void r600_resume_nontimer_queries(struct r600_context *ctx)
 	}
 }
 
-void r600_init_query_functions(struct r600_context *rctx)
+/* Get backends mask */
+void r600_query_init_backend_mask(struct r600_common_context *ctx)
 {
-	rctx->b.b.create_query = r600_create_query;
-	rctx->b.b.destroy_query = r600_destroy_query;
-	rctx->b.b.begin_query = r600_begin_query;
-	rctx->b.b.end_query = r600_end_query;
-	rctx->b.b.get_query_result = r600_get_query_result;
-
-	if (rctx->screen->b.info.r600_num_backends > 0)
-	    rctx->b.b.render_condition = r600_render_condition;
+	struct radeon_winsys_cs *cs = ctx->rings.gfx.cs;
+	struct r600_resource *buffer;
+	uint32_t *results;
+	unsigned num_backends = ctx->screen->info.r600_num_backends;
+	unsigned i, mask = 0;
+	uint64_t va;
+
+	/* if backend_map query is supported by the kernel */
+	if (ctx->screen->info.r600_backend_map_valid) {
+		unsigned num_tile_pipes = ctx->screen->info.r600_num_tile_pipes;
+		unsigned backend_map = ctx->screen->info.r600_backend_map;
+		unsigned item_width, item_mask;
+
+		if (ctx->chip_class >= EVERGREEN) {
+			item_width = 4;
+			item_mask = 0x7;
+		} else {
+			item_width = 2;
+			item_mask = 0x3;
+		}
+
+		while(num_tile_pipes--) {
+			i = backend_map & item_mask;
+			mask |= (1<<i);
+			backend_map >>= item_width;
+		}
+		if (mask != 0) {
+			ctx->backend_mask = mask;
+			return;
+		}
+	}
+
+	/* otherwise backup path for older kernels */
+
+	/* create buffer for event data */
+	buffer = (struct r600_resource*)
+		pipe_buffer_create(ctx->b.screen, PIPE_BIND_CUSTOM,
+				   PIPE_USAGE_STAGING, ctx->max_db*16);
+	if (!buffer)
+		goto err;
+	va = r600_resource_va(ctx->b.screen, (void*)buffer);
+
+	/* initialize buffer with zeroes */
+	results = r600_buffer_map_sync_with_rings(ctx, buffer, PIPE_TRANSFER_WRITE);
+	if (results) {
+		memset(results, 0, ctx->max_db * 4 * 4);
+		ctx->ws->buffer_unmap(buffer->cs_buf);
+
+		/* emit EVENT_WRITE for ZPASS_DONE */
+		cs->buf[cs->cdw++] = PKT3(PKT3_EVENT_WRITE, 2, 0);
+		cs->buf[cs->cdw++] = EVENT_TYPE(EVENT_TYPE_ZPASS_DONE) | EVENT_INDEX(1);
+		cs->buf[cs->cdw++] = va;
+		cs->buf[cs->cdw++] = (va >> 32UL) & 0xFF;
+
+		cs->buf[cs->cdw++] = PKT3(PKT3_NOP, 0, 0);
+		cs->buf[cs->cdw++] = r600_context_bo_reloc(ctx, &ctx->rings.gfx, buffer, RADEON_USAGE_WRITE);
+
+		/* analyze results */
+		results = r600_buffer_map_sync_with_rings(ctx, buffer, PIPE_TRANSFER_READ);
+		if (results) {
+			for(i = 0; i < ctx->max_db; i++) {
+				/* at least highest bit will be set if backend is used */
+				if (results[i*4 + 1])
+					mask |= (1<<i);
+			}
+			ctx->ws->buffer_unmap(buffer->cs_buf);
+		}
+	}
+
+	pipe_resource_reference((struct pipe_resource**)&buffer, NULL);
+
+	if (mask != 0) {
+		ctx->backend_mask = mask;
+		return;
+	}
+
+err:
+	/* fallback to old method - set num_backends lower bits to 1 */
+	ctx->backend_mask = (~((uint32_t)0))>>(32-num_backends);
+	return;
+}
+
+void r600_query_init(struct r600_common_context *rctx)
+{
+	rctx->b.create_query = r600_create_query;
+	rctx->b.destroy_query = r600_destroy_query;
+	rctx->b.begin_query = r600_begin_query;
+	rctx->b.end_query = r600_end_query;
+	rctx->b.get_query_result = r600_get_query_result;
+
+	if (((struct r600_common_screen*)rctx->b.screen)->info.r600_num_backends > 0)
+	    rctx->b.render_condition = r600_render_condition;
+
+	LIST_INITHEAD(&rctx->active_nontimer_queries);
 }

src/gallium/drivers/radeon/r600d_common.h

@@ -40,6 +40,7 @@
 #define RADEON_CP_PACKET3_COMPUTE_MODE 0x00000002
 
 #define PKT3_NOP                               0x10
+#define PKT3_SET_PREDICATION                   0x20
 #define PKT3_STRMOUT_BUFFER_UPDATE             0x34
 #define		STRMOUT_STORE_BUFFER_FILLED_SIZE	1
 #define		STRMOUT_OFFSET_SOURCE(x)	(((x) & 0x3) << 1)
@@ -51,6 +52,7 @@
 #define PKT3_WAIT_REG_MEM                      0x3C
 #define		WAIT_REG_MEM_EQUAL		3
 #define PKT3_EVENT_WRITE                       0x46
+#define PKT3_EVENT_WRITE_EOP                   0x47
 #define PKT3_SET_CONFIG_REG		       0x68
 #define PKT3_SET_CONTEXT_REG		       0x69
 #define PKT3_STRMOUT_BASE_UPDATE	       0x72 /* r700 only */
@@ -82,6 +84,16 @@
 		 * 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_CONTINUE (1 << 31)
+#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)
+
 /* R600-R700*/
 #define R_008490_CP_STRMOUT_CNTL		     0x008490
 #define   S_008490_OFFSET_UPDATE_DONE(x)		(((x) & 0x1) << 0)

src/gallium/drivers/radeonsi/si_pipe.c

@@ -132,16 +132,14 @@ static struct pipe_context *si_create_context(struct pipe_screen *screen, void *
 	if (sctx == NULL)
 		return NULL;
 
-	if (!r600_common_context_init(&sctx->b, &sscreen->b))
-		goto fail;
-
-	sctx->b.b.screen = screen;
+	sctx->b.b.screen = screen; /* this must be set first */
 	sctx->b.b.priv = priv;
 	sctx->b.b.destroy = si_destroy_context;
 	sctx->b.b.flush = si_flush_from_st;
+	sctx->screen = sscreen; /* Easy accessing of screen/winsys. */
 
-	/* Easy accessing of screen/winsys. */
-	sctx->screen = sscreen;
+	if (!r600_common_context_init(&sctx->b, &sscreen->b))
+		goto fail;
 
 	si_init_blit_functions(sctx);
 	si_init_query_functions(sctx);