freedreno: add transform-feedback state

Signed-off-by: Rob Clark <robclark@freedesktop.org>

src/gallium/drivers/freedreno/freedreno_context.h

@@ -82,6 +82,20 @@ struct fd_vertex_stateobj {
 	unsigned num_elements;
 };
 
+struct fd_streamout_stateobj {
+	struct pipe_stream_output_target *targets[PIPE_MAX_SO_BUFFERS];
+	unsigned num_targets;
+	/* Track offset from vtxcnt for streamout data.  This counter
+	 * is just incremented by # of vertices on each draw until
+	 * reset or new streamout buffer bound.
+	 *
+	 * When we eventually have GS, the CPU won't actually know the
+	 * number of vertices per draw, so I think we'll have to do
+	 * something more clever.
+	 */
+	unsigned offsets[PIPE_MAX_SO_BUFFERS];
+};
+
 /* group together the vertex and vertexbuf state.. for ease of passing
  * around, and because various internal operations (gmem<->mem, etc)
  * need their own vertex state:
@@ -319,6 +333,7 @@ struct fd_context {
 		FD_DIRTY_VTXBUF      = (1 << 15),
 		FD_DIRTY_INDEXBUF    = (1 << 16),
 		FD_DIRTY_SCISSOR     = (1 << 17),
+		FD_DIRTY_STREAMOUT   = (1 << 18),
 	} dirty;
 
 	struct pipe_blend_state *blend;
@@ -339,6 +354,7 @@ struct fd_context {
 	struct pipe_viewport_state viewport;
 	struct fd_constbuf_stateobj constbuf[PIPE_SHADER_TYPES];
 	struct pipe_index_buffer indexbuf;
+	struct fd_streamout_stateobj streamout;
 
 	/* GMEM/tile handling fxns: */
 	void (*emit_tile_init)(struct fd_context *ctx);

src/gallium/drivers/freedreno/freedreno_draw.c

@@ -62,7 +62,7 @@ fd_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
 	struct fd_context *ctx = fd_context(pctx);
 	struct pipe_framebuffer_state *pfb = &ctx->framebuffer;
 	struct pipe_scissor_state *scissor = fd_context_get_scissor(ctx);
-	unsigned i, buffers = 0;
+	unsigned i, prims, buffers = 0;
 
 	/* if we supported transform feedback, we'd have to disable this: */
 	if (((scissor->maxx - scissor->minx) *
@@ -144,11 +144,17 @@ fd_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
 		if (ctx->fragtex.textures[i])
 			resource_used(ctx, ctx->fragtex.textures[i]->texture, true);
 
+	/* Mark streamout buffers as being read.. actually they are written.. */
+	for (i = 0; i < ctx->streamout.num_targets; i++)
+		if (ctx->streamout.targets[i])
+			resource_used(ctx, ctx->streamout.targets[i]->buffer, false);
+
 	ctx->num_draws++;
 
+	prims = u_reduced_prims_for_vertices(info->mode, info->count);
+
 	ctx->stats.draw_calls++;
-	ctx->stats.prims_emitted +=
-		u_reduced_prims_for_vertices(info->mode, info->count);
+	ctx->stats.prims_emitted += prims;
 
 	/* any buffers that haven't been cleared yet, we need to restore: */
 	ctx->restore |= buffers & (FD_BUFFER_ALL & ~ctx->cleared);
@@ -162,6 +168,9 @@ fd_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
 	fd_hw_query_set_stage(ctx, ctx->ring, FD_STAGE_DRAW);
 	ctx->draw_vbo(ctx, info);
 
+	for (i = 0; i < ctx->streamout.num_targets; i++)
+		ctx->streamout.offsets[i] += prims;
+
 	/* if an app (or, well, piglit test) does many thousands of draws
 	 * without flush (or anything which implicitly flushes, like
 	 * changing render targets), we can exceed the ringbuffer size.

src/gallium/drivers/freedreno/freedreno_resource.c

@@ -647,6 +647,8 @@ fd_blitter_pipe_begin(struct fd_context *ctx)
 	util_blitter_save_vertex_buffer_slot(ctx->blitter, ctx->vtx.vertexbuf.vb);
 	util_blitter_save_vertex_elements(ctx->blitter, ctx->vtx.vtx);
 	util_blitter_save_vertex_shader(ctx->blitter, ctx->prog.vp);
+	util_blitter_save_so_targets(ctx->blitter, ctx->streamout.num_targets,
+			ctx->streamout.targets);
 	util_blitter_save_rasterizer(ctx->blitter, ctx->rasterizer);
 	util_blitter_save_viewport(ctx->blitter, &ctx->viewport);
 	util_blitter_save_scissor(ctx->blitter, &ctx->scissor);

src/gallium/drivers/freedreno/freedreno_state.c

@@ -300,6 +300,67 @@ fd_vertex_state_bind(struct pipe_context *pctx, void *hwcso)
 	ctx->dirty |= FD_DIRTY_VTXSTATE;
 }
 
+static struct pipe_stream_output_target *
+fd_create_stream_output_target(struct pipe_context *pctx,
+		struct pipe_resource *prsc, unsigned buffer_offset,
+		unsigned buffer_size)
+{
+	struct pipe_stream_output_target *target;
+
+	target = CALLOC_STRUCT(pipe_stream_output_target);
+	if (!target)
+		return NULL;
+
+	pipe_reference_init(&target->reference, 1);
+	pipe_resource_reference(&target->buffer, prsc);
+
+	target->context = pctx;
+	target->buffer_offset = buffer_offset;
+	target->buffer_size = buffer_size;
+
+	return target;
+}
+
+static void
+fd_stream_output_target_destroy(struct pipe_context *pctx,
+		struct pipe_stream_output_target *target)
+{
+	pipe_resource_reference(&target->buffer, NULL);
+	FREE(target);
+}
+
+static void
+fd_set_stream_output_targets(struct pipe_context *pctx,
+		unsigned num_targets, struct pipe_stream_output_target **targets,
+		const unsigned *offsets)
+{
+	struct fd_context *ctx = fd_context(pctx);
+	struct fd_streamout_stateobj *so = &ctx->streamout;
+	unsigned i;
+
+	debug_assert(num_targets <= ARRAY_SIZE(so->targets));
+
+	for (i = 0; i < num_targets; i++) {
+		boolean changed = targets[i] != so->targets[i];
+		boolean append = (offsets[i] == (unsigned)-1);
+
+		if (!changed && append)
+			continue;
+
+		so->offsets[i] = 0;
+
+		pipe_so_target_reference(&so->targets[i], targets[i]);
+	}
+
+	for (; i < so->num_targets; i++) {
+		pipe_so_target_reference(&so->targets[i], NULL);
+	}
+
+	so->num_targets = num_targets;
+
+	ctx->dirty |= FD_DIRTY_STREAMOUT;
+}
+
 void
 fd_state_init(struct pipe_context *pctx)
 {
@@ -328,4 +389,8 @@ fd_state_init(struct pipe_context *pctx)
 	pctx->create_vertex_elements_state = fd_vertex_state_create;
 	pctx->delete_vertex_elements_state = fd_vertex_state_delete;
 	pctx->bind_vertex_elements_state = fd_vertex_state_bind;
+
+	pctx->create_stream_output_target = fd_create_stream_output_target;
+	pctx->stream_output_target_destroy = fd_stream_output_target_destroy;
+	pctx->set_stream_output_targets = fd_set_stream_output_targets;
 }