[Mesa-dev] [PATCH 06/12] freedreno: dynamically sized/growable cmd buffers

[Rob Clark]

From: Rob Clark <robclark at freedesktop.org>

Signed-off-by: Rob Clark <robdclark at gmail.com>
---
 src/gallium/drivers/freedreno/freedreno_batch.c | 25 ++++++++++++--------
 src/gallium/drivers/freedreno/freedreno_util.h  | 31 ++++++++++++++-----------
 2 files changed, 33 insertions(+), 23 deletions(-)

diff --git a/src/gallium/drivers/freedreno/freedreno_batch.c b/src/gallium/drivers/freedreno/freedreno_batch.c
index 51a61d9..6d17a42 100644
--- a/src/gallium/drivers/freedreno/freedreno_batch.c
+++ b/src/gallium/drivers/freedreno/freedreno_batch.c
@@ -36,6 +36,7 @@ fd_batch_create(struct fd_context *ctx)
 {
 	struct fd_batch *batch = CALLOC_STRUCT(fd_batch);
 	static unsigned seqno = 0;
+	unsigned size = 0;
 
 	if (!batch)
 		return NULL;
@@ -44,13 +45,18 @@ fd_batch_create(struct fd_context *ctx)
 	batch->seqno = ++seqno;
 	batch->ctx = ctx;
 
-	/* TODO how to pick a good size?  Or maybe we should introduce
-	 * fd_ringlist?  Also, make sure size is aligned with bo-cache
-	 * bucket size, since otherwise that will round up size..
+	/* if kernel is too old to support unlimited # of cmd buffers, we
+	 * have no option but to allocate large worst-case sizes so that
+	 * we don't need to grow the ringbuffer.  Performance is likely to
+	 * suffer, but there is no good alternative.
 	 */
-	batch->draw    = fd_ringbuffer_new(ctx->screen->pipe, 0x10000);
-	batch->binning = fd_ringbuffer_new(ctx->screen->pipe, 0x10000);
-	batch->gmem    = fd_ringbuffer_new(ctx->screen->pipe, 0x10000);
+	if (fd_device_version(ctx->screen->dev) < FD_VERSION_UNLIMITED_CMDS) {
+		size = 0x100000;
+	}
+
+	batch->draw    = fd_ringbuffer_new(ctx->screen->pipe, size);
+	batch->binning = fd_ringbuffer_new(ctx->screen->pipe, size);
+	batch->gmem    = fd_ringbuffer_new(ctx->screen->pipe, size);
 
 	fd_ringbuffer_set_parent(batch->gmem, NULL);
 	fd_ringbuffer_set_parent(batch->draw, batch->gmem);
@@ -117,10 +123,9 @@ fd_batch_resource_used(struct fd_batch *batch, struct fd_resource *rsc,
 void
 fd_batch_check_size(struct fd_batch *batch)
 {
-	/* TODO eventually support having a list of draw/binning rb's
-	 * and if we are too close to the end, add another to the
-	 * list.  For now we just flush.
-	 */
+	if (fd_device_version(batch->ctx->screen->dev) >= FD_VERSION_UNLIMITED_CMDS)
+		return;
+
 	struct fd_ringbuffer *ring = batch->draw;
 	if (((ring->cur - ring->start) > (ring->size/4 - 0x1000)) ||
 			(fd_mesa_debug & FD_DBG_FLUSH))
diff --git a/src/gallium/drivers/freedreno/freedreno_util.h b/src/gallium/drivers/freedreno/freedreno_util.h
index b6b91f9..8f125d9 100644
--- a/src/gallium/drivers/freedreno/freedreno_util.h
+++ b/src/gallium/drivers/freedreno/freedreno_util.h
@@ -238,13 +238,8 @@ OUT_RELOCW(struct fd_ringbuffer *ring, struct fd_bo *bo,
 
 static inline void BEGIN_RING(struct fd_ringbuffer *ring, uint32_t ndwords)
 {
-	if ((ring->cur + ndwords) >= ring->end) {
-		/* this probably won't really work if we have multiple tiles..
-		 * but it is ok for 2d..  we might need different behavior
-		 * depending on 2d or 3d pipe.
-		 */
-		DBG("uh oh..");
-	}
+	if (ring->cur + ndwords >= ring->end)
+		fd_ringbuffer_grow(ring, ndwords);
 }
 
 static inline void
@@ -255,6 +250,13 @@ OUT_PKT0(struct fd_ringbuffer *ring, uint16_t regindx, uint16_t cnt)
 }
 
 static inline void
+OUT_PKT2(struct fd_ringbuffer *ring)
+{
+	BEGIN_RING(ring, 1);
+	OUT_RING(ring, CP_TYPE2_PKT);
+}
+
+static inline void
 OUT_PKT3(struct fd_ringbuffer *ring, uint8_t opcode, uint16_t cnt)
 {
 	BEGIN_RING(ring, cnt+1);
@@ -271,9 +273,7 @@ OUT_WFI(struct fd_ringbuffer *ring)
 static inline void
 __OUT_IB(struct fd_ringbuffer *ring, bool prefetch, struct fd_ringbuffer *target)
 {
-	uint32_t dwords = target->cur - target->start;
-
-	assert(dwords > 0);
+	unsigned count = fd_ringbuffer_cmd_count(target);
 
 	/* for debug after a lock up, write a unique counter value
 	 * to scratch6 for each IB, to make it easier to match up
@@ -283,9 +283,14 @@ __OUT_IB(struct fd_ringbuffer *ring, bool prefetch, struct fd_ringbuffer *target
 	 */
 	emit_marker(ring, 6);
 
-	OUT_PKT3(ring, prefetch ? CP_INDIRECT_BUFFER_PFE : CP_INDIRECT_BUFFER_PFD, 2);
-	fd_ringbuffer_emit_reloc_ring_full(ring, target, 0);
-	OUT_RING(ring, dwords);
+	for (unsigned i = 0; i < count; i++) {
+		uint32_t dwords;
+		OUT_PKT3(ring, prefetch ? CP_INDIRECT_BUFFER_PFE : CP_INDIRECT_BUFFER_PFD, 2);
+		dwords = fd_ringbuffer_emit_reloc_ring_full(ring, target, i) / 4;
+		assert(dwords > 0);
+		OUT_RING(ring, dwords);
+		OUT_PKT2(ring);
+	}
 
 	emit_marker(ring, 6);
 }
-- 
2.7.4