r600g/compute: Defer the creation of the temporary resource

For the first use of a buffer, we will only need the temporary
resource in the case that a user wants to write/map to this buffer.

But in the cases where the user creates a buffer to act as an
output of a kernel, then we were creating an unneeded resource,
because it will contain garbage, and would be copied to the pool,
and destroyed when promoting.

This patch avoids the creation and copies of resources in
this case.

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

diff --git a/src/gallium/drivers/r600/compute_memory_pool.c b/src/gallium/drivers/r600/compute_memory_pool.c
index 9cb16f8c29e42f08c10bd2b9bbcb55a500430c9e..fe19d9ef81c578b696809955e67200cb150a27e6 100644 (file)
--- a/src/gallium/drivers/r600/compute_memory_pool.c
+++ b/src/gallium/drivers/r600/compute_memory_pool.c
@@ -353,19 +353,21 @@ int compute_memory_promote_item(struct compute_memory_pool *pool,
        list_add(&item->link, pos);
        item->start_in_dw = start_in_dw;
 
-       u_box_1d(0, item->size_in_dw * 4, &box);
-
-       rctx->b.b.resource_copy_region(pipe,
-                       dst, 0, item->start_in_dw * 4, 0 ,0,
-                       src, 0, &box);
-
-       /* We check if the item is mapped for reading.
-        * In this case, we need to keep the temporary buffer 'alive'
-        * because it is possible to keep a map active for reading
-        * while a kernel (that reads from it) executes */
-       if (!(item->status & ITEM_MAPPED_FOR_READING)) {
-               pool->screen->b.b.resource_destroy(screen, src);
-               item->real_buffer = NULL;
+       if (src != NULL) {
+               u_box_1d(0, item->size_in_dw * 4, &box);
+
+               rctx->b.b.resource_copy_region(pipe,
+                               dst, 0, item->start_in_dw * 4, 0 ,0,
+                               src, 0, &box);
+
+               /* We check if the item is mapped for reading.
+                * In this case, we need to keep the temporary buffer 'alive'
+                * because it is possible to keep a map active for reading
+                * while a kernel (that reads from it) executes */
+               if (!(item->status & ITEM_MAPPED_FOR_READING)) {
+                       pool->screen->b.b.resource_destroy(screen, src);
+                       item->real_buffer = NULL;
+               }
        }
 
        return 0;
@@ -475,8 +477,7 @@ 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);
+       new_item->real_buffer = NULL;
 
        list_addtail(&new_item->link, pool->unallocated_list);
 

diff --git a/src/gallium/drivers/r600/evergreen_compute.c b/src/gallium/drivers/r600/evergreen_compute.c
index 5c115dcb282b8d926e2c53750af7e08c5c2c2ff8..12e9c85586d21dd1ff5cff0dafe72aebb11c3c91 100644 (file)
--- a/src/gallium/drivers/r600/evergreen_compute.c
+++ b/src/gallium/drivers/r600/evergreen_compute.c
@@ -970,14 +970,21 @@ void *r600_compute_global_transfer_map(
        struct r600_resource_global* buffer =
                (struct r600_resource_global*)resource;
 
-       struct pipe_resource *dst;
+       struct compute_memory_item *item = buffer->chunk;
+       struct pipe_resource *dst = NULL;
        unsigned offset = box->x;
 
-       if (is_item_in_pool(buffer->chunk)) {
-               compute_memory_demote_item(pool, buffer->chunk, ctx_);
+       if (is_item_in_pool(item)) {
+               compute_memory_demote_item(pool, item, ctx_);
+       }
+       else {
+               if (item->real_buffer == NULL) {
+                       item->real_buffer = (struct r600_resource*)
+                                       r600_compute_buffer_alloc_vram(pool->screen, item->size_in_dw * 4);
+               }
        }
 
-       dst = (struct pipe_resource*)buffer->chunk->real_buffer;
+       dst = (struct pipe_resource*)item->real_buffer;
 
        if (usage & PIPE_TRANSFER_READ)
                buffer->chunk->status |= ITEM_MAPPED_FOR_READING;
@@ -988,7 +995,7 @@ void *r600_compute_global_transfer_map(
                        box->x, box->y, box->z, box->width, box->height,
                        box->depth);
        COMPUTE_DBG(rctx->screen, "Buffer id = %u offset = "
-               "%u (box.x)\n", buffer->chunk->id, box->x);
+               "%u (box.x)\n", item->id, box->x);
 
 
        assert(resource->target == PIPE_BUFFER);