Mesa (master): radeonsi: don't wait for idle at the end of gfx IBs

[GitLab Mirror]

Module: Mesa
Branch: master
Commit: 266fec1307b26a544007423582afd8618791893c
URL:    http://cgit.freedesktop.org/mesa/mesa/commit/?id=266fec1307b26a544007423582afd8618791893c

Author: Marek Olšák <marek.olsak at amd.com>
Date:   Mon May  4 17:41:41 2020 -0400

radeonsi: don't wait for idle at the end of gfx IBs

Reviewed-by: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer at amd.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/4894>

---

 src/gallium/drivers/radeonsi/si_gfx_cs.c | 13 +++++++++++++
 1 file changed, 13 insertions(+)

diff --git a/src/gallium/drivers/radeonsi/si_gfx_cs.c b/src/gallium/drivers/radeonsi/si_gfx_cs.c
index 30ba6b02f87..087df67d61d 100644
--- a/src/gallium/drivers/radeonsi/si_gfx_cs.c
+++ b/src/gallium/drivers/radeonsi/si_gfx_cs.c
@@ -78,6 +78,19 @@ void si_flush_gfx_cs(struct si_context *ctx, unsigned flags, struct pipe_fence_h
    if (ctx->gfx_flush_in_progress)
       return;
 
+   /* The amdgpu kernel driver always synchronizes execution for shared DMABUFs
+    * between processes, so we don't have to wait at the end of IBs to make sure
+    * everything is idle.
+    *
+    * The amdgpu winsys synchronizes execution for buffers shared by different
+    * contexts within the same process.
+    *
+    * Interop with AMDVLK, RADV, or OpenCL within the same process requires
+    * explicit fences or glFinish.
+    */
+   if (ctx->screen->info.is_amdgpu)
+      flags |= RADEON_FLUSH_START_NEXT_GFX_IB_NOW;
+
    if (!ctx->screen->info.kernel_flushes_tc_l2_after_ib) {
       wait_flags |= wait_ps_cs | SI_CONTEXT_INV_L2;
    } else if (ctx->chip_class == GFX6) {