4 files changed, 71 insertions, 48 deletions

diff --git a/src/gallium/drivers/freedreno/freedreno_context.c b/src/gallium/drivers/freedreno/freedreno_context.c
index 02613dcdbdb..8e6d43150ce 100644
--- a/src/gallium/drivers/freedreno/freedreno_context.c
+++ b/src/gallium/drivers/freedreno/freedreno_context.c
@@ -94,9 +94,7 @@ void
 fd_context_render(struct pipe_context *pctx)
 {
 	struct fd_context *ctx = fd_context(pctx);
-	struct pipe_framebuffer_state *pfb = &ctx->framebuffer;
 	struct fd_resource *rsc, *rsc_tmp;
-	int i;
 
 	DBG("needs_flush: %d", ctx->needs_flush);
 
@@ -118,21 +116,11 @@ fd_context_render(struct pipe_context *pctx)
 	ctx->gmem_reason = 0;
 	ctx->num_draws = 0;
 
-	for (i = 0; i < pfb->nr_cbufs; i++)
-		if (pfb->cbufs[i])
-			fd_resource(pfb->cbufs[i]->texture)->dirty = false;
-	if (pfb->zsbuf) {
-		rsc = fd_resource(pfb->zsbuf->texture);
-		rsc->dirty = false;
-		if (rsc->stencil)
-			rsc->stencil->dirty = false;
-	}
-
 	/* go through all the used resources and clear their reading flag */
 	LIST_FOR_EACH_ENTRY_SAFE(rsc, rsc_tmp, &ctx->used_resources, list) {
-		assert(rsc->reading || rsc->writing);
-		rsc->reading = false;
-		rsc->writing = false;
+		debug_assert(rsc->status != 0);
+		rsc->status = 0;
+		rsc->pending_ctx = NULL;
 		list_delinit(&rsc->list);
 	}
 
diff --git a/src/gallium/drivers/freedreno/freedreno_draw.c b/src/gallium/drivers/freedreno/freedreno_draw.c
index f88654063fa..a04cc879630 100644
--- a/src/gallium/drivers/freedreno/freedreno_draw.c
+++ b/src/gallium/drivers/freedreno/freedreno_draw.c
@@ -40,7 +40,8 @@
 #include "freedreno_util.h"
 
 static void
-resource_used(struct fd_context *ctx, struct pipe_resource *prsc, boolean reading)
+resource_used(struct fd_context *ctx, struct pipe_resource *prsc,
+		enum fd_resource_status status)
 {
 	struct fd_resource *rsc;
 
@@ -48,12 +49,29 @@ resource_used(struct fd_context *ctx, struct pipe_resource *prsc, boolean readin
 		return;
 
 	rsc = fd_resource(prsc);
-	if (reading)
-		rsc->reading = true;
-	else
-		rsc->writing = true;
+	rsc->status |= status;
+	if (rsc->stencil)
+		rsc->stencil->status |= status;
+
+	/* TODO resources can actually be shared across contexts,
+	 * so I'm not sure a single list-head will do the trick?
+	 */
+	debug_assert((rsc->pending_ctx == ctx) || !rsc->pending_ctx);
 	list_delinit(&rsc->list);
 	list_addtail(&rsc->list, &ctx->used_resources);
+	rsc->pending_ctx = ctx;
+}
+
+static void
+resource_read(struct fd_context *ctx, struct pipe_resource *prsc)
+{
+	resource_used(ctx, prsc, FD_PENDING_READ);
+}
+
+static void
+resource_written(struct fd_context *ctx, struct pipe_resource *prsc)
+{
+	resource_used(ctx, prsc, FD_PENDING_WRITE);
 }
 
 static void
@@ -72,6 +90,8 @@ fd_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
 
 	/* emulate unsupported primitives: */
 	if (!fd_supported_prim(ctx, info->mode)) {
+		if (ctx->streamout.num_targets > 0)
+			debug_error("stream-out with emulated prims");
 		util_primconvert_save_index_buffer(ctx->primconvert, &ctx->indexbuf);
 		util_primconvert_save_rasterizer_state(ctx->primconvert, ctx->rasterizer);
 		util_primconvert_draw_vbo(ctx->primconvert, info);
@@ -86,17 +106,13 @@ fd_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
 
 	if (fd_depth_enabled(ctx)) {
 		buffers |= FD_BUFFER_DEPTH;
-		fd_resource(pfb->zsbuf->texture)->dirty = true;
+		resource_written(ctx, pfb->zsbuf->texture);
 		ctx->gmem_reason |= FD_GMEM_DEPTH_ENABLED;
 	}
 
 	if (fd_stencil_enabled(ctx)) {
-		struct fd_resource *rsc = fd_resource(pfb->zsbuf->texture);
 		buffers |= FD_BUFFER_STENCIL;
-		if (rsc->stencil)
-			rsc->stencil->dirty = true;
-		else
-			rsc->dirty = true;
+		resource_written(ctx, pfb->zsbuf->texture);
 		ctx->gmem_reason |= FD_GMEM_STENCIL_ENABLED;
 	}
 
@@ -111,7 +127,7 @@ fd_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
 
 		surf = pfb->cbufs[i]->texture;
 
-		fd_resource(surf)->dirty = true;
+		resource_written(ctx, surf);
 		buffers |= PIPE_CLEAR_COLOR0 << i;
 
 		if (surf->nr_samples > 1)
@@ -123,31 +139,31 @@ fd_draw_vbo(struct pipe_context *pctx, const struct pipe_draw_info *info)
 
 	/* Skip over buffer 0, that is sent along with the command stream */
 	for (i = 1; i < PIPE_MAX_CONSTANT_BUFFERS; i++) {
-		resource_used(ctx, ctx->constbuf[PIPE_SHADER_VERTEX].cb[i].buffer, true);
-		resource_used(ctx, ctx->constbuf[PIPE_SHADER_FRAGMENT].cb[i].buffer, true);
+		resource_read(ctx, ctx->constbuf[PIPE_SHADER_VERTEX].cb[i].buffer);
+		resource_read(ctx, ctx->constbuf[PIPE_SHADER_FRAGMENT].cb[i].buffer);
 	}
 
 	/* Mark VBOs as being read */
 	for (i = 0; i < ctx->vtx.vertexbuf.count; i++) {
 		assert(!ctx->vtx.vertexbuf.vb[i].user_buffer);
-		resource_used(ctx, ctx->vtx.vertexbuf.vb[i].buffer, true);
+		resource_read(ctx, ctx->vtx.vertexbuf.vb[i].buffer);
 	}
 
 	/* Mark index buffer as being read */
-	resource_used(ctx, ctx->indexbuf.buffer, true);
+	resource_read(ctx, ctx->indexbuf.buffer);
 
 	/* Mark textures as being read */
 	for (i = 0; i < ctx->verttex.num_textures; i++)
 		if (ctx->verttex.textures[i])
-			resource_used(ctx, ctx->verttex.textures[i]->texture, true);
+			resource_read(ctx, ctx->verttex.textures[i]->texture);
 	for (i = 0; i < ctx->fragtex.num_textures; i++)
 		if (ctx->fragtex.textures[i])
-			resource_used(ctx, ctx->fragtex.textures[i]->texture, true);
+			resource_read(ctx, ctx->fragtex.textures[i]->texture);
 
-	/* Mark streamout buffers as being read.. actually they are written.. */
+	/* Mark streamout buffers as being written.. */
 	for (i = 0; i < ctx->streamout.num_targets; i++)
 		if (ctx->streamout.targets[i])
-			resource_used(ctx, ctx->streamout.targets[i]->buffer, false);
+			resource_written(ctx, ctx->streamout.targets[i]->buffer);
 
 	ctx->num_draws++;
 
@@ -228,15 +244,10 @@ fd_clear(struct pipe_context *pctx, unsigned buffers,
 	if (buffers & PIPE_CLEAR_COLOR)
 		for (i = 0; i < pfb->nr_cbufs; i++)
 			if (buffers & (PIPE_CLEAR_COLOR0 << i))
-				fd_resource(pfb->cbufs[i]->texture)->dirty = true;
+				resource_written(ctx, pfb->cbufs[i]->texture);
 
 	if (buffers & (PIPE_CLEAR_DEPTH | PIPE_CLEAR_STENCIL)) {
-		struct fd_resource *rsc = fd_resource(pfb->zsbuf->texture);
-		if (rsc->stencil && buffers & PIPE_CLEAR_STENCIL)
-			rsc->stencil->dirty = true;
-		if (!rsc->stencil || buffers & PIPE_CLEAR_DEPTH)
-			rsc->dirty = true;
-
+		resource_written(ctx, pfb->zsbuf->texture);
 		ctx->gmem_reason |= FD_GMEM_CLEARS_DEPTH_STENCIL;
 	}
 
diff --git a/src/gallium/drivers/freedreno/freedreno_resource.c b/src/gallium/drivers/freedreno/freedreno_resource.c
index d649925af48..709ad4eb55b 100644
--- a/src/gallium/drivers/freedreno/freedreno_resource.c
+++ b/src/gallium/drivers/freedreno/freedreno_resource.c
@@ -42,6 +42,14 @@
 
 #include <errno.h>
 
+
+static bool
+pending(struct fd_resource *rsc, enum fd_resource_status status)
+{
+	return (rsc->status & status) ||
+		(rsc->stencil && (rsc->stencil->status & status));
+}
+
 static void
 fd_invalidate_resource(struct fd_context *ctx, struct pipe_resource *prsc)
 {
@@ -97,7 +105,8 @@ realloc_bo(struct fd_resource *rsc, uint32_t size)
 
 	rsc->bo = fd_bo_new(screen->dev, size, flags);
 	rsc->timestamp = 0;
-	rsc->dirty = rsc->reading = rsc->writing = false;
+	rsc->status = 0;
+	rsc->pending_ctx = NULL;
 	list_delinit(&rsc->list);
 	util_range_set_empty(&rsc->valid_buffer_range);
 }
@@ -238,9 +247,9 @@ fd_resource_transfer_map(struct pipe_context *pctx,
 		/* If the GPU is writing to the resource, or if it is reading from the
 		 * resource and we're trying to write to it, flush the renders.
 		 */
-		if (rsc->dirty || (rsc->stencil && rsc->stencil->dirty) ||
-			((ptrans->usage & PIPE_TRANSFER_WRITE) && rsc->reading) ||
-			((ptrans->usage & PIPE_TRANSFER_READ) && rsc->writing))
+		if (((ptrans->usage & PIPE_TRANSFER_WRITE) &&
+					pending(rsc, FD_PENDING_READ | FD_PENDING_WRITE)) ||
+				pending(rsc, FD_PENDING_WRITE))
 			fd_context_render(pctx);
 
 		/* The GPU keeps track of how the various bo's are being used, and
@@ -678,7 +687,7 @@ fd_flush_resource(struct pipe_context *pctx, struct pipe_resource *prsc)
 {
 	struct fd_resource *rsc = fd_resource(prsc);
 
-	if (rsc->dirty || (rsc->stencil && rsc->stencil->dirty))
+	if (pending(rsc, FD_PENDING_WRITE | FD_PENDING_READ))
 		fd_context_render(pctx);
 }
 
diff --git a/src/gallium/drivers/freedreno/freedreno_resource.h b/src/gallium/drivers/freedreno/freedreno_resource.h
index e7f127edca4..7549becaa1f 100644
--- a/src/gallium/drivers/freedreno/freedreno_resource.h
+++ b/src/gallium/drivers/freedreno/freedreno_resource.h
@@ -60,6 +60,15 @@ struct fd_resource_slice {
 	uint32_t size0;          /* size of first layer in slice */
 };
 
+/* status of queued up but not flushed reads and write operations.
+ * In _transfer_map() we need to know if queued up rendering needs
+ * to be flushed to preserve the order of cpu and gpu access.
+ */
+enum fd_resource_status {
+	FD_PENDING_WRITE = 0x01,
+	FD_PENDING_READ  = 0x02,
+};
+
 struct fd_resource {
 	struct u_resource base;
 	struct fd_bo *bo;
@@ -68,14 +77,20 @@ struct fd_resource {
 	uint32_t layer_size;
 	struct fd_resource_slice slices[MAX_MIP_LEVELS];
 	uint32_t timestamp;
-	bool dirty, reading, writing;
 	/* buffer range that has been initialized */
 	struct util_range valid_buffer_range;
 
 	/* reference to the resource holding stencil data for a z32_s8 texture */
+	/* TODO rename to secondary or auxiliary? */
 	struct fd_resource *stencil;
 
+	/* pending read/write state: */
+	enum fd_resource_status status;
+	/* resources accessed by queued but not flushed draws are tracked
+	 * in the used_resources list.
+	 */
 	struct list_head list;
+	struct fd_context *pending_ctx;
 };
 
 static inline struct fd_resource *