Hi Alex,
We might want to move the mes_v11_0's mqd to VRAM, too. According to my test result, moving mes queue's mqd to vram will lead to better performance than only moving gfx&compute queues' mqd.
Regards & Thanks,
Yubiao
-----Original Message-----
From: amd-gfx <amd-gfx-bounces@xxxxxxxxxxxxxxxxxxxxx> On Behalf Of Alex Deucher
Sent: Thursday, April 20, 2023 11:59 PM
To: amd-gfx@xxxxxxxxxxxxxxxxxxxxx
Cc: Deucher, Alexander <Alexander.Deucher@xxxxxxx>
Subject: [PATCH] drm/amdgpu: put MQDs in VRAM
Reduces preemption latency.
Signed-off-by: Alex Deucher <alexander.deucher@xxxxxxx>
---
drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c | 2 ++
1 file changed, 2 insertions(+)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c
index 60bb4bba1994..f52feee1d64d 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.c
@@ -413,6 +413,7 @@ int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
ring = &adev->gfx.gfx_ring[i];
if (!ring->mqd_obj) {
r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
+ AMDGPU_GEM_DOMAIN_VRAM |
AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
&ring->mqd_gpu_addr, &ring->mqd_ptr);
if (r) {
@@ -434,6 +435,7 @@ int amdgpu_gfx_mqd_sw_init(struct amdgpu_device *adev,
ring = &adev->gfx.compute_ring[i + xcc_id * adev->gfx.num_compute_rings];
if (!ring->mqd_obj) {
r = amdgpu_bo_create_kernel(adev, mqd_size, PAGE_SIZE,
+ AMDGPU_GEM_DOMAIN_VRAM |
AMDGPU_GEM_DOMAIN_GTT, &ring->mqd_obj,
&ring->mqd_gpu_addr, &ring->mqd_ptr);
if (r) {
--
2.40.0
