radeonsi: integrate shared streamout state

Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>

src/gallium/drivers/radeonsi/r600_blit.c

 		util_blitter_save_viewport(rctx->blitter, &rctx->queued.named.viewport->viewport);
 	}
 	util_blitter_save_vertex_buffer_slot(rctx->blitter, rctx->vertex_buffer);
-	util_blitter_save_so_targets(rctx->blitter, rctx->num_so_targets,
-				     (struct pipe_stream_output_target**)rctx->so_targets);
+	util_blitter_save_so_targets(rctx->blitter, rctx->b.streamout.num_targets,
+				     (struct pipe_stream_output_target**)rctx->b.streamout.targets);
 
 	if (op & R600_SAVE_FRAMEBUFFER)
 		util_blitter_save_framebuffer(rctx->blitter, &rctx->framebuffer);

src/gallium/drivers/radeonsi/r600_hw_context.c

 	num_dw += ctx->num_cs_dw_nontimer_queries_suspend;
 
 	/* Count in streamout_end at the end of CS. */
-	num_dw += ctx->num_cs_dw_streamout_end;
+	if (ctx->b.streamout.begin_emitted) {
+		num_dw += ctx->b.streamout.num_dw_for_end;
+	}
 
 	/* Count in render_condition(NULL) at the end of CS. */
 	if (ctx->predicate_drawing) {
 	struct radeon_winsys_cs *cs = ctx->b.rings.gfx.cs;
 	bool queries_suspended = false;
 
-#if 0
-	bool streamout_suspended = false;
-#endif
-
 	if (!cs->cdw)
 		return;
 
 		queries_suspended = true;
 	}
 
-#if 0
-	if (ctx->num_cs_dw_streamout_end) {
-		r600_context_streamout_end(ctx);
-		streamout_suspended = true;
+	ctx->b.streamout.suspended = false;
+
+	if (ctx->b.streamout.begin_emitted) {
+		r600_emit_streamout_end(&ctx->b);
+		ctx->b.streamout.suspended = true;
 	}
-#endif
 
 	ctx->b.flags |= R600_CONTEXT_FLUSH_AND_INV_CB |
 			R600_CONTEXT_FLUSH_AND_INV_CB_META |
 	si_pm4_emit(ctx, ctx->queued.named.init);
 	ctx->emitted.named.init = ctx->queued.named.init;
 
-#if 0
-	if (streamout_suspended) {
-		ctx->streamout_start = TRUE;
-		ctx->streamout_append_bitmask = ~0;
+	if (ctx->b.streamout.suspended) {
+		ctx->b.streamout.append_bitmask = ctx->b.streamout.enabled_mask;
+		r600_streamout_buffers_dirty(&ctx->b);
 	}
-#endif
 
 	/* resume queries */
 	if (queries_suspended) {

src/gallium/drivers/radeonsi/radeonsi_pipe.c

 	rctx->cache_flush = si_atom_cache_flush;
 	rctx->atoms.cache_flush = &rctx->cache_flush;
 
+	rctx->atoms.streamout_begin = &rctx->b.streamout.begin_atom;
+
 	switch (rctx->b.chip_class) {
 	case SI:
 	case CIK:

src/gallium/drivers/radeonsi/radeonsi_pipe.h

 			/* Caches must be flushed after resource descriptors are
 			 * updated in memory. */
 			struct r600_atom *cache_flush;
+			struct r600_atom *streamout_begin;
 		};
 		struct r600_atom *array[0];
 	} atoms;
 	/* The list of active queries. Only one query of each type can be active. */
 	struct list_head	active_nontimer_query_list;
 	unsigned		num_cs_dw_nontimer_queries_suspend;
-	unsigned		num_cs_dw_streamout_end;
 
 	unsigned		backend_mask;
 	unsigned                max_db; /* for OQ */
 	boolean                 predicate_drawing;
 
-	unsigned		num_so_targets;
-	struct r600_so_target	*so_targets[PIPE_MAX_SO_BUFFERS];
-	boolean			streamout_start;
-	unsigned		streamout_append_bitmask;
-	unsigned		*vs_so_stride_in_dw;
-	unsigned		*vs_shader_so_strides;
-
 	/* Vertex and index buffers. */
 	bool			vertex_buffers_dirty;
 	struct pipe_index_buffer index_buffer;

src/gallium/drivers/radeonsi/radeonsi_shader.h

 	unsigned			cb_shader_mask;
 	bool				cb0_is_integer;
 	unsigned			sprite_coord_enable;
-	unsigned			so_strides[4];
 	union si_shader_key		key;
 };
 

src/gallium/drivers/radeonsi/si_state.c

 
 	rctx->vs_shader = sel;
 
-	if (sel && sel->current)
+	if (sel && sel->current) {
 		si_pm4_bind_state(rctx, vs, sel->current->pm4);
-	else
+		rctx->b.streamout.stride_in_dw = sel->so.stride;
+	} else {
 		si_pm4_bind_state(rctx, vs, rctx->dummy_pixel_shader->pm4);
+	}
 
 	rctx->b.flags |= R600_CONTEXT_INV_SHADER_CACHE;
 }

src/gallium/drivers/radeonsi/si_state_draw.c

 		}
 	}
 
-	rctx->vs_shader_so_strides = rctx->vs_shader->current->so_strides;
-
 	if (!si_update_draw_info_state(rctx, info))
 		return;
 
 	}
 #endif
 
-#if 0
-	/* Enable stream out if needed. */
-	if (rctx->streamout_start) {
-		r600_context_streamout_begin(rctx);
-		rctx->streamout_start = FALSE;
-	}
-#endif
-
 	/* Set the depth buffer as dirty. */
 	if (rctx->framebuffer.zsbuf) {
 		struct pipe_surface *surf = rctx->framebuffer.zsbuf;