Skip to contentSkip to conversation

drm/amdgpu: don't invalidate caches in RELEASE_MEM, only do the writeback

PATCH In review Marek Olšák 2 participants 3 messages Jul 2, 2019 – Jul 8, 2019 2 messages not in the archive

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 drm/amdgpu: don't invalidate caches in RELEASE_MEM, only do the writeback

From: Marek Olšák <[email protected]>

This RELEASE_MEM use has the Release semantic, which means we should write back but not invalidate. Invalidations only make sense with the Acquire semantic (ACQUIRE_MEM), or when RELEASE_MEM is used to do the combined Acquire-Release semantic, which is a barrier, not a fence.

The undesirable side effect of doing invalidations for the Release semantic is that it invalidates caches while shaders are running, because the Release can execute in the middle of the next IB.

UMDs should use ACQUIRE_MEM at the beginning of IBs. Doing cache invalidations for a fence (like in this case) doesn't do anything for correctness.

---
 drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c | 6 +-----
 1 file changed, 1 insertion(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c +1 -5
index 210d24511dc6..a30f5d4913b9 100644
@@ -4296,25 +4296,21 @@ static void gfx_v10_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr,
 	bool int_sel = flags & AMDGPU_FENCE_FLAG_INT;
 
 	/* Interrupt not work fine on GFX10.1 model yet. Use fallback instead */
 	if (adev->pdev->device == 0x50)
 		int_sel = false;
 
 	/* RELEASE_MEM - flush caches, send int */
 	amdgpu_ring_write(ring, PACKET3(PACKET3_RELEASE_MEM, 6));
 	amdgpu_ring_write(ring, (PACKET3_RELEASE_MEM_GCR_SEQ |
 				 PACKET3_RELEASE_MEM_GCR_GL2_WB |
-				 PACKET3_RELEASE_MEM_GCR_GL2_INV |
-				 PACKET3_RELEASE_MEM_GCR_GL2_US |
-				 PACKET3_RELEASE_MEM_GCR_GL1_INV |
-				 PACKET3_RELEASE_MEM_GCR_GLV_INV |
-				 PACKET3_RELEASE_MEM_GCR_GLM_INV |
+				 PACKET3_RELEASE_MEM_GCR_GLM_INV | /* must be set with GLM_WB */
 				 PACKET3_RELEASE_MEM_GCR_GLM_WB |
 				 PACKET3_RELEASE_MEM_CACHE_POLICY(3) |
 				 PACKET3_RELEASE_MEM_EVENT_TYPE(CACHE_FLUSH_AND_INV_TS_EVENT) |
 				 PACKET3_RELEASE_MEM_EVENT_INDEX(5)));
 	amdgpu_ring_write(ring, (PACKET3_RELEASE_MEM_DATA_SEL(write64bit ? 2 : 1) |
 				 PACKET3_RELEASE_MEM_INT_SEL(int_sel ? 2 : 0)));
 
 	/*
 	 * the address should be Qword aligned if 64bit write, Dword
 	 * aligned if only send 32bit data low (discard data high)
Signature · 6 lines
2.17.1

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 20190702182901.22491-1-maraeo-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org

#2 lore raw

ping

Quoted 56 lines · Marek Olšák

From: Marek Olšák <[email protected]>

This RELEASE_MEM use has the Release semantic, which means we should write back but not invalidate. Invalidations only make sense with the Acquire semantic (ACQUIRE_MEM), or when RELEASE_MEM is used to do the combined Acquire-Release semantic, which is a barrier, not a fence.

The undesirable side effect of doing invalidations for the Release semantic is that it invalidates caches while shaders are running, because the Release
can execute in the middle of the next IB.

UMDs should use ACQUIRE_MEM at the beginning of IBs. Doing cache invalidations for a fence (like in this case) doesn't do anything for correctness.

---
 drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c | 6 +-----
 1 file changed, 1 insertion(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c +0 -0
index 210d24511dc6..a30f5d4913b9 100644
@@ -4296,25 +4296,21 @@ static void gfx_v10_0_ring_emit_fence(struct
 amdgpu_ring *ring, u64 addr,
        bool int_sel = flags & AMDGPU_FENCE_FLAG_INT;
 
        /* Interrupt not work fine on GFX10.1 model yet. Use fallback

instead */

        if (adev->pdev->device == 0x50)
                int_sel = false;

        /* RELEASE_MEM - flush caches, send int */
        amdgpu_ring_write(ring, PACKET3(PACKET3_RELEASE_MEM, 6));
        amdgpu_ring_write(ring, (PACKET3_RELEASE_MEM_GCR_SEQ |
                                 PACKET3_RELEASE_MEM_GCR_GL2_WB |
drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c +1 -5
-                                PACKET3_RELEASE_MEM_GCR_GL2_INV |
-                                PACKET3_RELEASE_MEM_GCR_GL2_US |
-                                PACKET3_RELEASE_MEM_GCR_GL1_INV |
-                                PACKET3_RELEASE_MEM_GCR_GLV_INV |
-                                PACKET3_RELEASE_MEM_GCR_GLM_INV |
+                                PACKET3_RELEASE_MEM_GCR_GLM_INV | /* must

be set with GLM_WB */
PACKET3_RELEASE_MEM_GCR_GLM_WB | PACKET3_RELEASE_MEM_CACHE_POLICY(3) |

PACKET3_RELEASE_MEM_EVENT_TYPE(CACHE_FLUSH_AND_INV_TS_EVENT) |
PACKET3_RELEASE_MEM_EVENT_INDEX(5)));
amdgpu_ring_write(ring, (PACKET3_RELEASE_MEM_DATA_SEL(write64bit ?
2 : 1) |
PACKET3_RELEASE_MEM_INT_SEL(int_sel ? 2 :
0)));

Quoted 7 lines

/*
* the address should be Qword aligned if 64bit write, Dword
* aligned if only send 32bit data low (discard data high)
--
2.17.1

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): CAAxE2A5FUU-TukfWC0-Od9u_deQKV3JZjPXCbwin=ccx-dOx-g-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org

#4 lore raw

________________________________
From: amd-gfx <amd-gfx-bounces-PD4FTy7X32lNgt0PjOBp9y5qC8QIuHrW@public.gmane.org> on behalf of Marek Olšák <[email protected]>
Sent: Monday, July 8, 2019 1:31 PM
To: amd-gfx mailing list
Subject: Re: [PATCH] drm/amdgpu: don't invalidate caches in RELEASE_MEM, only do the writeback

ping

On Tue, Jul 2, 2019 at 2:29 PM Marek Olšák <[email protected]<mailto:[email protected]>> wrote:
From: Marek Olšák <[email protected]<mailto:[email protected]>>

This RELEASE_MEM use has the Release semantic, which means we should write back but not invalidate. Invalidations only make sense with the Acquire semantic (ACQUIRE_MEM), or when RELEASE_MEM is used to do the combined Acquire-Release semantic, which is a barrier, not a fence.

The undesirable side effect of doing invalidations for the Release semantic is that it invalidates caches while shaders are running, because the Release can execute in the middle of the next IB.

UMDs should use ACQUIRE_MEM at the beginning of IBs. Doing cache invalidations for a fence (like in this case) doesn't do anything for correctness.

---
 drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c | 6 +-----
 1 file changed, 1 insertion(+), 5 deletions(-)
drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c +1 -5
index 210d24511dc6..a30f5d4913b9 100644
@@ -4296,25 +4296,21 @@ static void gfx_v10_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr,
        bool int_sel = flags & AMDGPU_FENCE_FLAG_INT;
 
        /* Interrupt not work fine on GFX10.1 model yet. Use fallback instead */
        if (adev->pdev->device == 0x50)
                int_sel = false;
 
        /* RELEASE_MEM - flush caches, send int */
        amdgpu_ring_write(ring, PACKET3(PACKET3_RELEASE_MEM, 6));
        amdgpu_ring_write(ring, (PACKET3_RELEASE_MEM_GCR_SEQ |
                                 PACKET3_RELEASE_MEM_GCR_GL2_WB |
-                                PACKET3_RELEASE_MEM_GCR_GL2_INV |
-                                PACKET3_RELEASE_MEM_GCR_GL2_US |
-                                PACKET3_RELEASE_MEM_GCR_GL1_INV |
-                                PACKET3_RELEASE_MEM_GCR_GLV_INV |
-                                PACKET3_RELEASE_MEM_GCR_GLM_INV |
+                                PACKET3_RELEASE_MEM_GCR_GLM_INV | /* must be set with GLM_WB */
                                 PACKET3_RELEASE_MEM_GCR_GLM_WB |
                                 PACKET3_RELEASE_MEM_CACHE_POLICY(3) |
                                 PACKET3_RELEASE_MEM_EVENT_TYPE(CACHE_FLUSH_AND_INV_TS_EVENT) |
                                 PACKET3_RELEASE_MEM_EVENT_INDEX(5)));
        amdgpu_ring_write(ring, (PACKET3_RELEASE_MEM_DATA_SEL(write64bit ? 2 : 1) |
                                 PACKET3_RELEASE_MEM_INT_SEL(int_sel ? 2 : 0)));
 
        /*
         * the address should be Qword aligned if 64bit write, Dword
         * aligned if only send 32bit data low (discard data high)

--
2.17.1

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx