AlexIndustrial/mesa
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

radv: Bypass L2 for gang semaphore BO with SDMA/ACE

Browse Source 
When the "gang leader" is SDMA, we need to ensure that the
gang semaphores BO is coherent between SDMA and CP.
To achieve this, we need bypass the L2 cache when either SDMA
or CP are connected to L2.

Suggested-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Konstantin Seurer <konstantin.seurer@gmail.com>
Reviewed-by: Samuel Pitoiset <samuel.pitoiset@gmail.com>
Part-of: <https://gitlab.freedesktop.org/mesa/mesa/-/merge_requests/39057>
This commit is contained in:
Timur Kristóf
2025-12-21 09:29:29 -06:00
committed by Marge Bot
parent 7700ba28f5
commit f55771a17d
1 changed files with 25 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
src/amd/vulkan/radv_queue.c
View File

@@ -1341,13 +1341,36 @@ radv_create_gang_wait_preambles_postambles(struct radv_queue *queue)
VkResult r = VK_SUCCESS; VkResult r = VK_SUCCESS;
struct radeon_winsys *ws = device->ws; struct radeon_winsys *ws = device->ws;
struct radeon_winsys_bo *gang_sem_bo = NULL; struct radeon_winsys_bo *gang_sem_bo = NULL;
enum radeon_bo_flag gang_sem_bo_flags = RADEON_FLAG_NO_INTERPROCESS_SHARING | RADEON_FLAG_ZERO_VRAM;
/* When the "gang leader" is SDMA, we need to ensure that the gang semaphores BO
* is coherent between SDMA and CP. To achieve this, we need bypass the L2 cache
* when either SDMA or CP are connected to L2.
*
* GFX6-8:
* neither CP nor SDMA are connected to L2
* GFX9:
* CP is connected to L2, SDMA isn't connected to L2
* GFX10:
* CP is connected to L2
* SDMA is connected to L2 but there are possible hw bugs with
* cache_policy=BYPASS which is currently the default configuration set by SDMA0_UTCL1_PAGE
* GFX10.3-11.5:
* CP is connected to L2
* SDMA is connected to L2, and an alternative solution would be to
* configure the cache policy in SDMA packets (ie. with CVP=1) directly to overwrite the
* default behavior, but it doesn't seem worth it.
* GFX12:
* neither CP nor SDMA are connected to L2
*/
if (ip == AMD_IP_SDMA && pdev->info.gfx_level >= GFX9 && pdev->info.gfx_level < GFX12)
gang_sem_bo_flags |= RADEON_FLAG_GL2_BYPASS;
/* Gang semaphores BO. /* Gang semaphores BO.
* DWORD 0: used in preambles, gang leader writes, gang members wait. * DWORD 0: used in preambles, gang leader writes, gang members wait.
* DWORD 1: used in postambles, gang leader waits, gang members write. * DWORD 1: used in postambles, gang leader waits, gang members write.
*/ */
r = radv_bo_create(device, NULL, 8, 4, RADEON_DOMAIN_VRAM, r = radv_bo_create(device, NULL, 8, 4, RADEON_DOMAIN_VRAM, gang_sem_bo_flags, RADV_BO_PRIORITY_SCRATCH, 0, true,
RADEON_FLAG_NO_INTERPROCESS_SHARING | RADEON_FLAG_ZERO_VRAM, RADV_BO_PRIORITY_SCRATCH, 0, true,
&gang_sem_bo); &gang_sem_bo);
if (r != VK_SUCCESS) if (r != VK_SUCCESS)
return r; return r;