r600g/compute: Add an intermediate resource for OpenCL buffers

This patch changes completely the way buffers are added to the
compute_memory_pool. Before this, whenever we were going to
map a buffer or write to or read from it, it would get placed
into the pool. Now, every unallocated buffer has its own
r600_resource until it is allocated in the pool.

NOTE: This patch also increase the GPU memory usage at the moment
of putting every buffer in it's place. More or less, the memory
usage is ~2x(sum of every buffer size)

v2: Cleanup

v3: Use temporary variables to avoid so many castings in functions,
    as suggested by Tom Stellard

Reviewed-by: Tom Stellard <thomas.stellard@amd.com>

3 files changed, 41 insertions, 6 deletions

diff --git a/src/gallium/drivers/r600/compute_memory_pool.c b/src/gallium/drivers/r600/compute_memory_pool.c
index ec8c470fc65..0b126a80191 100644
--- a/src/gallium/drivers/r600/compute_memory_pool.c
+++ b/src/gallium/drivers/r600/compute_memory_pool.c
@@ -71,7 +71,6 @@ static void compute_memory_pool_init(struct compute_memory_pool * pool,
 	if (pool->shadow == NULL)
 		return;
 
-	pool->next_id = 1;
 	pool->size_in_dw = initial_size_in_dw;
 	pool->bo = (struct r600_resource*)r600_compute_buffer_alloc_vram(pool->screen,
 							pool->size_in_dw * 4);
@@ -314,6 +313,14 @@ int compute_memory_finalize_pending(struct compute_memory_pool* pool,
 	for (item = pending_list; item; item = next) {
 		next = item->next;
 
+		struct pipe_screen *screen = (struct pipe_screen *)pool->screen;
+		struct r600_context *rctx = (struct r600_context *)pipe;
+		struct pipe_resource *dst = (struct pipe_resource *)pool->bo;
+		struct pipe_resource *src = (struct pipe_resource *)item->real_buffer;
+		struct pipe_box box;
+
+		u_box_1d(0, item->size_in_dw * 4, &box);
+
 		/* Search for free space in the pool for this item. */
 		while ((start_in_dw=compute_memory_prealloc_chunk(pool,
 						item->size_in_dw)) == -1) {
@@ -365,6 +372,14 @@ int compute_memory_finalize_pending(struct compute_memory_pool* pool,
 			pool->item_list = item;
 		}
 
+		rctx->b.b.resource_copy_region(pipe,
+				dst, 0, item->start_in_dw * 4, 0 ,0,
+				src, 0, &box);
+
+		pool->screen->b.b.resource_destroy(
+			screen, src);
+		item->real_buffer = NULL;
+
 		allocated += item->size_in_dw;
 	}
 
@@ -375,6 +390,8 @@ int compute_memory_finalize_pending(struct compute_memory_pool* pool,
 void compute_memory_free(struct compute_memory_pool* pool, int64_t id)
 {
 	struct compute_memory_item *item, *next;
+	struct pipe_screen *screen = (struct pipe_screen *)pool->screen;
+	struct pipe_resource *res;
 
 	COMPUTE_DBG(pool->screen, "* compute_memory_free() id + %ld \n", id);
 
@@ -393,6 +410,12 @@ void compute_memory_free(struct compute_memory_pool* pool, int64_t id)
 				item->next->prev = item->prev;
 			}
 
+			if (item->real_buffer) {
+				res = (struct pipe_resource *)item->real_buffer;
+				pool->screen->b.b.resource_destroy(
+						screen, res);
+			}
+
 			free(item);
 
 			return;
@@ -426,6 +449,8 @@ struct compute_memory_item* compute_memory_alloc(
 	new_item->start_in_dw = -1; /* mark pending */
 	new_item->id = pool->next_id++;
 	new_item->pool = pool;
+	new_item->real_buffer = (struct r600_resource*)r600_compute_buffer_alloc_vram(
+							pool->screen, size_in_dw * 4);
 
 	if (pool->item_list) {
 		for (last_item = pool->item_list; last_item->next;
diff --git a/src/gallium/drivers/r600/compute_memory_pool.h b/src/gallium/drivers/r600/compute_memory_pool.h
index c711c59bcfa..e94159cedac 100644
--- a/src/gallium/drivers/r600/compute_memory_pool.h
+++ b/src/gallium/drivers/r600/compute_memory_pool.h
@@ -38,6 +38,8 @@ struct compute_memory_item
 	int64_t start_in_dw; ///Start pointer in dwords relative in the pool bo
 	int64_t size_in_dw; ///Size of the chunk in dwords
 
+	struct r600_resource *real_buffer;
+
 	struct compute_memory_pool* pool;
 
 	struct compute_memory_item* prev;
diff --git a/src/gallium/drivers/r600/evergreen_compute.c b/src/gallium/drivers/r600/evergreen_compute.c
index a2abf1546dd..c152e54146a 100644
--- a/src/gallium/drivers/r600/evergreen_compute.c
+++ b/src/gallium/drivers/r600/evergreen_compute.c
@@ -958,6 +958,17 @@ void *r600_compute_global_transfer_map(
 	struct r600_resource_global* buffer =
 		(struct r600_resource_global*)resource;
 
+	struct pipe_resource *dst;
+	unsigned offset = box->x;
+
+	if (buffer->chunk->real_buffer) {
+		dst = (struct pipe_resource*)buffer->chunk->real_buffer;
+	}
+	else {
+		dst = (struct pipe_resource*)buffer->chunk->pool->bo;
+		offset += (buffer->chunk->start_in_dw * 4);
+	}
+
 	COMPUTE_DBG(rctx->screen, "* r600_compute_global_transfer_map()\n"
 			"level = %u, usage = %u, box(x = %u, y = %u, z = %u "
 			"width = %u, height = %u, depth = %u)\n", level, usage,
@@ -967,8 +978,6 @@ void *r600_compute_global_transfer_map(
 		"%u (box.x)\n", buffer->chunk->id, box->x);
 
 
-	compute_memory_finalize_pending(pool, ctx_);
-
 	assert(resource->target == PIPE_BUFFER);
 	assert(resource->bind & PIPE_BIND_GLOBAL);
 	assert(box->x >= 0);
@@ -976,9 +985,8 @@ void *r600_compute_global_transfer_map(
 	assert(box->z == 0);
 
 	///TODO: do it better, mapping is not possible if the pool is too big
-	return pipe_buffer_map_range(ctx_, (struct pipe_resource*)buffer->chunk->pool->bo,
-			box->x + (buffer->chunk->start_in_dw * 4),
-			box->width, usage, ptransfer);
+	return pipe_buffer_map_range(ctx_, dst,
+			offset, box->width, usage, ptransfer);
 }
 
 void r600_compute_global_transfer_unmap(