Reviewed-by: Nicolai Hähnle <nicolai.haeh...@amd.com>
On 31.05.2016 23:42, Marek Olšák wrote:
From: Marek Olšák <marek.ol...@amd.com>
This improves throughput by keeping TTM overhead down.
Some piglit tests such as texelFetch and streaming-texture-leak will
use less memory now.
v2: use gart_size / 4 as the threshold
---
src/gallium/drivers/radeon/r600_pipe_common.h | 1 +
src/gallium/drivers/radeon/r600_texture.c | 32 +++++++++++++++++++++++----
2 files changed, 29 insertions(+), 4 deletions(-)
diff --git a/src/gallium/drivers/radeon/r600_pipe_common.h
b/src/gallium/drivers/radeon/r600_pipe_common.h
index cf3756b..d693004 100644
--- a/src/gallium/drivers/radeon/r600_pipe_common.h
+++ b/src/gallium/drivers/radeon/r600_pipe_common.h
@@ -501,6 +501,7 @@ struct r600_common_context {
/* Misc stats. */
unsigned num_draw_calls;
unsigned num_dma_calls;
+ uint64_t num_alloc_tex_transfer_bytes;
/* Render condition. */
struct r600_atom render_cond_atom;
diff --git a/src/gallium/drivers/radeon/r600_texture.c
b/src/gallium/drivers/radeon/r600_texture.c
index dbabbd0..bf6cfe8 100644
--- a/src/gallium/drivers/radeon/r600_texture.c
+++ b/src/gallium/drivers/radeon/r600_texture.c
@@ -1302,9 +1302,11 @@ static bool r600_can_invalidate_texture(struct
r600_common_screen *rscreen,
box->depth);
}
-static void r600_texture_invalidate_storage(struct r600_common_screen *rscreen,
+static void r600_texture_invalidate_storage(struct r600_common_context *rctx,
struct r600_texture *rtex)
{
+ struct r600_common_screen *rscreen = rctx->screen;
+
/* There is no point in discarding depth and tiled buffers. */
assert(!rtex->is_depth);
assert(rtex->surface.level[0].mode == RADEON_SURF_MODE_LINEAR_ALIGNED);
@@ -1319,6 +1321,8 @@ static void r600_texture_invalidate_storage(struct
r600_common_screen *rscreen,
r600_dirty_all_framebuffer_states(rscreen);
p_atomic_inc(&rscreen->dirty_tex_descriptor_counter);
+
+ rctx->num_alloc_tex_transfer_bytes += rtex->size;
}
static void *r600_texture_transfer_map(struct pipe_context *ctx,
@@ -1378,8 +1382,7 @@ static void *r600_texture_transfer_map(struct
pipe_context *ctx,
/* It's busy. */
if (r600_can_invalidate_texture(rctx->screen, rtex,
usage, box))
- r600_texture_invalidate_storage(rctx->screen,
- rtex);
+ r600_texture_invalidate_storage(rctx, rtex);
else
use_staging_texture = true;
}
@@ -1499,6 +1502,7 @@ static void *r600_texture_transfer_map(struct
pipe_context *ctx,
static void r600_texture_transfer_unmap(struct pipe_context *ctx,
struct pipe_transfer* transfer)
{
+ struct r600_common_context *rctx = (struct r600_common_context*)ctx;
struct r600_transfer *rtransfer = (struct r600_transfer*)transfer;
struct pipe_resource *texture = transfer->resource;
struct r600_texture *rtex = (struct r600_texture*)texture;
@@ -1514,8 +1518,28 @@ static void r600_texture_transfer_unmap(struct
pipe_context *ctx,
}
}
- if (rtransfer->staging)
+ if (rtransfer->staging) {
+ rctx->num_alloc_tex_transfer_bytes +=
rtransfer->staging->buf->size;
pipe_resource_reference((struct
pipe_resource**)&rtransfer->staging, NULL);
+ }
+
+ /* Heuristic for {upload, draw, upload, draw, ..}:
+ *
+ * Flush the gfx IB if we've allocated too much texture storage.
+ *
+ * The idea is that we don't want to build IBs that use too much
+ * memory and put pressure on the kernel memory manager and we also
+ * want to make temporary and invalidated buffers go idle ASAP to
+ * decrease the total memory usage or make them reusable. The memory
+ * usage will be slightly higher than given here because of the buffer
+ * cache in the winsys.
+ *
+ * The result is that the kernel memory manager is never a bottleneck.
+ */
+ if (rctx->num_alloc_tex_transfer_bytes > rctx->screen->info.gart_size /
4) {
+ rctx->gfx.flush(rctx, RADEON_FLUSH_ASYNC, NULL);
+ rctx->num_alloc_tex_transfer_bytes = 0;
+ }
FREE(transfer);
}
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
