Mesa (main): radeonsi: make the DRI_PRIME dGPU -> iGPU copy async
GitLab Mirror
gitlab-mirror at kemper.freedesktop.org
Thu Oct 7 09:44:50 UTC 2021
Module: Mesa
Branch: main
Commit: a1dea665d023df0f0d240a32a5c7df259c3db92c
URL: http://cgit.freedesktop.org/mesa/mesa/commit/?id=a1dea665d023df0f0d240a32a5c7df259c3db92c
Author: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer at amd.com>
Date: Tue Sep 7 20:12:30 2021 +0200
radeonsi: make the DRI_PRIME dGPU -> iGPU copy async
Doing this copy using SDMA frees up the dGPU to do more
interesting things while the copy is happening; for instance
the rendering of the next frame.
hw queue activity before:
------------------------
dGPU:
gfx: [renderframe 1][copy->iGPU][renderframe 2][copy->iGPU]...
iGPU:
gfx: [Xorg] [Xorg]
hw queue activity before after:
------------------------------
dGPU:
gfx: [renderframe 1][renderframe 2][renderframe 3]....
sdma: [copy->iGPU] [copy->iGPU] [copy->iGPU]
iGPU:
gfx: [Xorg] [Xorg] ...
If SDMA isn't available or can't do the copy, use an async compute
context instead.
Reviewed-by: Marek Olšák <marek.olsak at amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/12763>
---
src/gallium/drivers/radeonsi/si_blit.c | 37 ++++++++++++++++++++++++++++++++++
1 file changed, 37 insertions(+)
diff --git a/src/gallium/drivers/radeonsi/si_blit.c b/src/gallium/drivers/radeonsi/si_blit.c
index 6de37ab2bd8..bf5ccb1cc9a 100644
--- a/src/gallium/drivers/radeonsi/si_blit.c
+++ b/src/gallium/drivers/radeonsi/si_blit.c
@@ -1205,11 +1205,48 @@ resolve_to_temp:
static void si_blit(struct pipe_context *ctx, const struct pipe_blit_info *info)
{
struct si_context *sctx = (struct si_context *)ctx;
+ struct si_texture *sdst = (struct si_texture *)info->dst.resource;
if (do_hardware_msaa_resolve(ctx, info)) {
return;
}
+ if (info->is_dri_blit_image && sdst->surface.is_linear &&
+ sctx->chip_class >= GFX7 && sdst->surface.flags & RADEON_SURF_IMPORTED) {
+ struct si_texture *ssrc = (struct si_texture *)info->src.resource;
+ /* Use SDMA or async compute when copying to a DRI_PRIME imported linear surface. */
+ bool async_copy = info->dst.box.x == 0 && info->dst.box.y == 0 && info->dst.box.z == 0 &&
+ info->src.box.x == 0 && info->src.box.y == 0 && info->src.box.z == 0 &&
+ info->dst.level == 0 && info->src.level == 0 &&
+ info->src.box.width == info->dst.resource->width0 &&
+ info->src.box.height == info->dst.resource->height0 &&
+ info->src.box.depth == 1 && util_can_blit_via_copy_region(info, true);
+ /* Try SDMA first... */
+ /* TODO: figure out why SDMA copies are slow on GFX10_3 */
+ if (async_copy && sctx->chip_class < GFX10_3 && si_sdma_copy_image(sctx, sdst, ssrc))
+ return;
+
+ /* ... and use async compute as the fallback. */
+ if (async_copy) {
+ struct si_screen *sscreen = sctx->screen;
+
+ simple_mtx_lock(&sscreen->async_compute_context_lock);
+ if (!sscreen->async_compute_context)
+ si_init_aux_async_compute_ctx(sscreen);
+
+ if (sscreen->async_compute_context) {
+ si_compute_copy_image((struct si_context*)sctx->screen->async_compute_context,
+ info->dst.resource, 0, info->src.resource, 0, 0, 0, 0,
+ &info->src.box, false, 0);
+ si_flush_gfx_cs((struct si_context*)sctx->screen->async_compute_context, 0, NULL);
+ simple_mtx_unlock(&sscreen->async_compute_context_lock);
+ return;
+ }
+
+ simple_mtx_unlock(&sscreen->async_compute_context_lock);
+ }
+ }
+
if (unlikely(sctx->thread_trace_enabled))
sctx->sqtt_next_event = EventCmdCopyImage;
More information about the mesa-commit
mailing list