* [PATCH] drm/amdgpu: Add graphics cache rinse packet for sdma 5.0
@ 2021-04-20 19:27 Alex Deucher
2021-04-29 3:40 ` Alex Deucher
0 siblings, 1 reply; 5+ messages in thread
From: Alex Deucher @ 2021-04-20 19:27 UTC (permalink / raw)
To: amd-gfx; +Cc: Alex Deucher
Add emit mem sync callback for sdma_v5_0
In amdgpu sync object test, three threads created jobs
to send GFX IB and SDMA IB in sequence. After the first
GFX thread joined, sometimes the third thread will reuse
the same physical page to store the SDMA IB. There will
be a risk that SDMA will read GFX IB in the previous physical
page. So it's better to flush the cache before commit sdma IB.
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
---
drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c | 28 ++++++++++++++++++++++++++
1 file changed, 28 insertions(+)
diff --git a/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c b/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c
index 920fc6d4a127..d294ef6a625a 100644
--- a/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c
+++ b/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c
@@ -437,6 +437,33 @@ static void sdma_v5_0_ring_emit_ib(struct amdgpu_ring *ring,
amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr));
}
+/**
+ * sdma_v5_0_ring_emit_mem_sync - flush the IB by graphics cache rinse
+ *
+ * @ring: amdgpu ring pointer
+ * @job: job to retrieve vmid from
+ * @ib: IB object to schedule
+ *
+ * flush the IB by graphics cache rinse.
+ */
+static void sdma_v5_0_ring_emit_mem_sync(struct amdgpu_ring *ring)
+{
+ uint32_t gcr_cntl =
+ SDMA_GCR_GL2_INV | SDMA_GCR_GL2_WB | SDMA_GCR_GLM_INV |
+ SDMA_GCR_GL1_INV | SDMA_GCR_GLV_INV | SDMA_GCR_GLK_INV |
+ SDMA_GCR_GLI_INV(1);
+
+ /* flush entire cache L0/L1/L2, this can be optimized by performance requirement */
+ amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_GCR_REQ));
+ amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD1_BASE_VA_31_7(0));
+ amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD2_GCR_CONTROL_15_0(gcr_cntl) |
+ SDMA_PKT_GCR_REQ_PAYLOAD2_BASE_VA_47_32(0));
+ amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD3_LIMIT_VA_31_7(0) |
+ SDMA_PKT_GCR_REQ_PAYLOAD3_GCR_CONTROL_18_16(gcr_cntl >> 16));
+ amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD4_LIMIT_VA_47_32(0) |
+ SDMA_PKT_GCR_REQ_PAYLOAD4_VMID(0));
+}
+
/**
* sdma_v5_0_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
*
@@ -1643,6 +1670,7 @@ static const struct amdgpu_ring_funcs sdma_v5_0_ring_funcs = {
10 + 10 + 10, /* sdma_v5_0_ring_emit_fence x3 for user fence, vm fence */
.emit_ib_size = 5 + 7 + 6, /* sdma_v5_0_ring_emit_ib */
.emit_ib = sdma_v5_0_ring_emit_ib,
+ .emit_mem_sync = sdma_v5_0_ring_emit_mem_sync,
.emit_fence = sdma_v5_0_ring_emit_fence,
.emit_pipeline_sync = sdma_v5_0_ring_emit_pipeline_sync,
.emit_vm_flush = sdma_v5_0_ring_emit_vm_flush,
--
2.30.2
_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
^ permalink raw reply related [flat|nested] 5+ messages in thread
* Re: [PATCH] drm/amdgpu: Add graphics cache rinse packet for sdma 5.0
2021-04-20 19:27 [PATCH] drm/amdgpu: Add graphics cache rinse packet for sdma 5.0 Alex Deucher
@ 2021-04-29 3:40 ` Alex Deucher
2021-04-29 3:43 ` Zhang, Hawking
2021-04-29 8:11 ` Chen, Jiansong (Simon)
0 siblings, 2 replies; 5+ messages in thread
From: Alex Deucher @ 2021-04-29 3:40 UTC (permalink / raw)
To: Alex Deucher; +Cc: amd-gfx list
Ping?
On Tue, Apr 20, 2021 at 3:28 PM Alex Deucher <alexander.deucher@amd.com> wrote:
>
> Add emit mem sync callback for sdma_v5_0
>
> In amdgpu sync object test, three threads created jobs
> to send GFX IB and SDMA IB in sequence. After the first
> GFX thread joined, sometimes the third thread will reuse
> the same physical page to store the SDMA IB. There will
> be a risk that SDMA will read GFX IB in the previous physical
> page. So it's better to flush the cache before commit sdma IB.
>
> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
> ---
> drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c | 28 ++++++++++++++++++++++++++
> 1 file changed, 28 insertions(+)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c b/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c
> index 920fc6d4a127..d294ef6a625a 100644
> --- a/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c
> +++ b/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c
> @@ -437,6 +437,33 @@ static void sdma_v5_0_ring_emit_ib(struct amdgpu_ring *ring,
> amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr));
> }
>
> +/**
> + * sdma_v5_0_ring_emit_mem_sync - flush the IB by graphics cache rinse
> + *
> + * @ring: amdgpu ring pointer
> + * @job: job to retrieve vmid from
> + * @ib: IB object to schedule
> + *
> + * flush the IB by graphics cache rinse.
> + */
> +static void sdma_v5_0_ring_emit_mem_sync(struct amdgpu_ring *ring)
> +{
> + uint32_t gcr_cntl =
> + SDMA_GCR_GL2_INV | SDMA_GCR_GL2_WB | SDMA_GCR_GLM_INV |
> + SDMA_GCR_GL1_INV | SDMA_GCR_GLV_INV | SDMA_GCR_GLK_INV |
> + SDMA_GCR_GLI_INV(1);
> +
> + /* flush entire cache L0/L1/L2, this can be optimized by performance requirement */
> + amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_GCR_REQ));
> + amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD1_BASE_VA_31_7(0));
> + amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD2_GCR_CONTROL_15_0(gcr_cntl) |
> + SDMA_PKT_GCR_REQ_PAYLOAD2_BASE_VA_47_32(0));
> + amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD3_LIMIT_VA_31_7(0) |
> + SDMA_PKT_GCR_REQ_PAYLOAD3_GCR_CONTROL_18_16(gcr_cntl >> 16));
> + amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD4_LIMIT_VA_47_32(0) |
> + SDMA_PKT_GCR_REQ_PAYLOAD4_VMID(0));
> +}
> +
> /**
> * sdma_v5_0_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
> *
> @@ -1643,6 +1670,7 @@ static const struct amdgpu_ring_funcs sdma_v5_0_ring_funcs = {
> 10 + 10 + 10, /* sdma_v5_0_ring_emit_fence x3 for user fence, vm fence */
> .emit_ib_size = 5 + 7 + 6, /* sdma_v5_0_ring_emit_ib */
> .emit_ib = sdma_v5_0_ring_emit_ib,
> + .emit_mem_sync = sdma_v5_0_ring_emit_mem_sync,
> .emit_fence = sdma_v5_0_ring_emit_fence,
> .emit_pipeline_sync = sdma_v5_0_ring_emit_pipeline_sync,
> .emit_vm_flush = sdma_v5_0_ring_emit_vm_flush,
> --
> 2.30.2
>
> _______________________________________________
> amd-gfx mailing list
> amd-gfx@lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/amd-gfx
_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
^ permalink raw reply [flat|nested] 5+ messages in thread
* RE: [PATCH] drm/amdgpu: Add graphics cache rinse packet for sdma 5.0
2021-04-29 3:40 ` Alex Deucher
@ 2021-04-29 3:43 ` Zhang, Hawking
2021-04-29 8:11 ` Chen, Jiansong (Simon)
1 sibling, 0 replies; 5+ messages in thread
From: Zhang, Hawking @ 2021-04-29 3:43 UTC (permalink / raw)
To: Alex Deucher, Deucher, Alexander; +Cc: amd-gfx list
[AMD Public Use]
Reviewed-by: Hawking Zhang <Hawking.Zhang@amd.com>
Regards,
Hawking
-----Original Message-----
From: amd-gfx <amd-gfx-bounces@lists.freedesktop.org> On Behalf Of Alex Deucher
Sent: Thursday, April 29, 2021 11:41
To: Deucher, Alexander <Alexander.Deucher@amd.com>
Cc: amd-gfx list <amd-gfx@lists.freedesktop.org>
Subject: Re: [PATCH] drm/amdgpu: Add graphics cache rinse packet for sdma 5.0
Ping?
On Tue, Apr 20, 2021 at 3:28 PM Alex Deucher <alexander.deucher@amd.com> wrote:
>
> Add emit mem sync callback for sdma_v5_0
>
> In amdgpu sync object test, three threads created jobs to send GFX IB
> and SDMA IB in sequence. After the first GFX thread joined, sometimes
> the third thread will reuse the same physical page to store the SDMA
> IB. There will be a risk that SDMA will read GFX IB in the previous
> physical page. So it's better to flush the cache before commit sdma
> IB.
>
> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
> ---
> drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c | 28
> ++++++++++++++++++++++++++
> 1 file changed, 28 insertions(+)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c
> b/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c
> index 920fc6d4a127..d294ef6a625a 100644
> --- a/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c
> +++ b/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c
> @@ -437,6 +437,33 @@ static void sdma_v5_0_ring_emit_ib(struct amdgpu_ring *ring,
> amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr)); }
>
> +/**
> + * sdma_v5_0_ring_emit_mem_sync - flush the IB by graphics cache
> +rinse
> + *
> + * @ring: amdgpu ring pointer
> + * @job: job to retrieve vmid from
> + * @ib: IB object to schedule
> + *
> + * flush the IB by graphics cache rinse.
> + */
> +static void sdma_v5_0_ring_emit_mem_sync(struct amdgpu_ring *ring) {
> + uint32_t gcr_cntl =
> + SDMA_GCR_GL2_INV | SDMA_GCR_GL2_WB | SDMA_GCR_GLM_INV |
> + SDMA_GCR_GL1_INV | SDMA_GCR_GLV_INV | SDMA_GCR_GLK_INV |
> + SDMA_GCR_GLI_INV(1);
> +
> + /* flush entire cache L0/L1/L2, this can be optimized by performance requirement */
> + amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_GCR_REQ));
> + amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD1_BASE_VA_31_7(0));
> + amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD2_GCR_CONTROL_15_0(gcr_cntl) |
> + SDMA_PKT_GCR_REQ_PAYLOAD2_BASE_VA_47_32(0));
> + amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD3_LIMIT_VA_31_7(0) |
> + SDMA_PKT_GCR_REQ_PAYLOAD3_GCR_CONTROL_18_16(gcr_cntl >> 16));
> + amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD4_LIMIT_VA_47_32(0) |
> + SDMA_PKT_GCR_REQ_PAYLOAD4_VMID(0));
> +}
> +
> /**
> * sdma_v5_0_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
> *
> @@ -1643,6 +1670,7 @@ static const struct amdgpu_ring_funcs sdma_v5_0_ring_funcs = {
> 10 + 10 + 10, /* sdma_v5_0_ring_emit_fence x3 for user fence, vm fence */
> .emit_ib_size = 5 + 7 + 6, /* sdma_v5_0_ring_emit_ib */
> .emit_ib = sdma_v5_0_ring_emit_ib,
> + .emit_mem_sync = sdma_v5_0_ring_emit_mem_sync,
> .emit_fence = sdma_v5_0_ring_emit_fence,
> .emit_pipeline_sync = sdma_v5_0_ring_emit_pipeline_sync,
> .emit_vm_flush = sdma_v5_0_ring_emit_vm_flush,
> --
> 2.30.2
>
> _______________________________________________
> amd-gfx mailing list
> amd-gfx@lists.freedesktop.org
> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flist
> s.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&data=04%7C01%7Cha
> wking.zhang%40amd.com%7C27b92616b764440e6f6908d90ac0a302%7C3dd8961fe48
> 84e608e11a82d994e183d%7C0%7C0%7C637552644756636394%7CUnknown%7CTWFpbGZ
> sb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3
> D%7C1000&sdata=W%2F3m8Hla7J1prF%2BlApw60YpW2rSF8qq9XIjIqgV8DUQ%3D&
> amp;reserved=0
_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&data=04%7C01%7Chawking.zhang%40amd.com%7C27b92616b764440e6f6908d90ac0a302%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637552644756636394%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=W%2F3m8Hla7J1prF%2BlApw60YpW2rSF8qq9XIjIqgV8DUQ%3D&reserved=0
_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
^ permalink raw reply [flat|nested] 5+ messages in thread
* RE: [PATCH] drm/amdgpu: Add graphics cache rinse packet for sdma 5.0
2021-04-29 3:40 ` Alex Deucher
2021-04-29 3:43 ` Zhang, Hawking
@ 2021-04-29 8:11 ` Chen, Jiansong (Simon)
2021-04-29 16:03 ` Alex Deucher
1 sibling, 1 reply; 5+ messages in thread
From: Chen, Jiansong (Simon) @ 2021-04-29 8:11 UTC (permalink / raw)
To: Alex Deucher, Deucher, Alexander; +Cc: amd-gfx list
[AMD Official Use Only - Internal Distribution Only]
Hi Alex,
I notice there is already similar logic in sdma_v5_0_ring_emit_ib, do we need remove it?
Regards,
Jiansong
@@ -410,6 +410,18 @@ static void sdma_v5_0_ring_emit_ib(struct amdgpu_ring *ring,
unsigned vmid = AMDGPU_JOB_GET_VMID(job);
uint64_t csa_mc_addr = amdgpu_sdma_get_csa_mc_addr(ring, vmid);
+ /* Invalidate L2, because if we don't do it, we might get stale cache
+ * lines from previous IBs.
+ */
+ amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_GCR_REQ));
+ amdgpu_ring_write(ring, 0);
+ amdgpu_ring_write(ring, (SDMA_GCR_GL2_INV |
+ SDMA_GCR_GL2_WB |
+ SDMA_GCR_GLM_INV |
+ SDMA_GCR_GLM_WB) << 16);
+ amdgpu_ring_write(ring, 0xffffff80);
+ amdgpu_ring_write(ring, 0xffff);
+
-----Original Message-----
From: amd-gfx <amd-gfx-bounces@lists.freedesktop.org> On Behalf Of Alex Deucher
Sent: Thursday, April 29, 2021 11:41 AM
To: Deucher, Alexander <Alexander.Deucher@amd.com>
Cc: amd-gfx list <amd-gfx@lists.freedesktop.org>
Subject: Re: [PATCH] drm/amdgpu: Add graphics cache rinse packet for sdma 5.0
Ping?
On Tue, Apr 20, 2021 at 3:28 PM Alex Deucher <alexander.deucher@amd.com> wrote:
>
> Add emit mem sync callback for sdma_v5_0
>
> In amdgpu sync object test, three threads created jobs to send GFX IB
> and SDMA IB in sequence. After the first GFX thread joined, sometimes
> the third thread will reuse the same physical page to store the SDMA
> IB. There will be a risk that SDMA will read GFX IB in the previous
> physical page. So it's better to flush the cache before commit sdma
> IB.
>
> Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
> ---
> drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c | 28
> ++++++++++++++++++++++++++
> 1 file changed, 28 insertions(+)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c
> b/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c
> index 920fc6d4a127..d294ef6a625a 100644
> --- a/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c
> +++ b/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c
> @@ -437,6 +437,33 @@ static void sdma_v5_0_ring_emit_ib(struct amdgpu_ring *ring,
> amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr)); }
>
> +/**
> + * sdma_v5_0_ring_emit_mem_sync - flush the IB by graphics cache
> +rinse
> + *
> + * @ring: amdgpu ring pointer
> + * @job: job to retrieve vmid from
> + * @ib: IB object to schedule
> + *
> + * flush the IB by graphics cache rinse.
> + */
> +static void sdma_v5_0_ring_emit_mem_sync(struct amdgpu_ring *ring) {
> + uint32_t gcr_cntl =
> + SDMA_GCR_GL2_INV | SDMA_GCR_GL2_WB | SDMA_GCR_GLM_INV |
> + SDMA_GCR_GL1_INV | SDMA_GCR_GLV_INV | SDMA_GCR_GLK_INV |
> + SDMA_GCR_GLI_INV(1);
> +
> + /* flush entire cache L0/L1/L2, this can be optimized by performance requirement */
> + amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_GCR_REQ));
> + amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD1_BASE_VA_31_7(0));
> + amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD2_GCR_CONTROL_15_0(gcr_cntl) |
> + SDMA_PKT_GCR_REQ_PAYLOAD2_BASE_VA_47_32(0));
> + amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD3_LIMIT_VA_31_7(0) |
> + SDMA_PKT_GCR_REQ_PAYLOAD3_GCR_CONTROL_18_16(gcr_cntl >> 16));
> + amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD4_LIMIT_VA_47_32(0) |
> + SDMA_PKT_GCR_REQ_PAYLOAD4_VMID(0));
> +}
> +
> /**
> * sdma_v5_0_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
> *
> @@ -1643,6 +1670,7 @@ static const struct amdgpu_ring_funcs sdma_v5_0_ring_funcs = {
> 10 + 10 + 10, /* sdma_v5_0_ring_emit_fence x3 for user fence, vm fence */
> .emit_ib_size = 5 + 7 + 6, /* sdma_v5_0_ring_emit_ib */
> .emit_ib = sdma_v5_0_ring_emit_ib,
> + .emit_mem_sync = sdma_v5_0_ring_emit_mem_sync,
> .emit_fence = sdma_v5_0_ring_emit_fence,
> .emit_pipeline_sync = sdma_v5_0_ring_emit_pipeline_sync,
> .emit_vm_flush = sdma_v5_0_ring_emit_vm_flush,
> --
> 2.30.2
>
> _______________________________________________
> amd-gfx mailing list
> amd-gfx@lists.freedesktop.org
> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flist
> s.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&data=04%7C01%7CJi
> ansong.Chen%40amd.com%7C1b777b80b17145712a7b08d90ac0a1f4%7C3dd8961fe48
> 84e608e11a82d994e183d%7C0%7C0%7C637552644738458840%7CUnknown%7CTWFpbGZ
> sb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3
> D%7C1000&sdata=T9jZUIJIQHS2gY8bU%2F7uM1ealAP3qxelkg2Slj3JASA%3D&am
> p;reserved=0
_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&data=04%7C01%7CJiansong.Chen%40amd.com%7C1b777b80b17145712a7b08d90ac0a1f4%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637552644738458840%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=T9jZUIJIQHS2gY8bU%2F7uM1ealAP3qxelkg2Slj3JASA%3D&reserved=0
_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
^ permalink raw reply [flat|nested] 5+ messages in thread
* Re: [PATCH] drm/amdgpu: Add graphics cache rinse packet for sdma 5.0
2021-04-29 8:11 ` Chen, Jiansong (Simon)
@ 2021-04-29 16:03 ` Alex Deucher
0 siblings, 0 replies; 5+ messages in thread
From: Alex Deucher @ 2021-04-29 16:03 UTC (permalink / raw)
To: Chen, Jiansong (Simon); +Cc: Deucher, Alexander, amd-gfx list
On Thu, Apr 29, 2021 at 4:11 AM Chen, Jiansong (Simon)
<Jiansong.Chen@amd.com> wrote:
>
> [AMD Official Use Only - Internal Distribution Only]
>
> Hi Alex,
> I notice there is already similar logic in sdma_v5_0_ring_emit_ib, do we need remove it?
>
I think we can also remove this from sdma_v5_0_ring_emit_ib in a
follow up patch.
Alex
> Regards,
> Jiansong
>
> @@ -410,6 +410,18 @@ static void sdma_v5_0_ring_emit_ib(struct amdgpu_ring *ring,
> unsigned vmid = AMDGPU_JOB_GET_VMID(job);
> uint64_t csa_mc_addr = amdgpu_sdma_get_csa_mc_addr(ring, vmid);
>
> + /* Invalidate L2, because if we don't do it, we might get stale cache
> + * lines from previous IBs.
> + */
> + amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_GCR_REQ));
> + amdgpu_ring_write(ring, 0);
> + amdgpu_ring_write(ring, (SDMA_GCR_GL2_INV |
> + SDMA_GCR_GL2_WB |
> + SDMA_GCR_GLM_INV |
> + SDMA_GCR_GLM_WB) << 16);
> + amdgpu_ring_write(ring, 0xffffff80);
> + amdgpu_ring_write(ring, 0xffff);
> +
>
> -----Original Message-----
> From: amd-gfx <amd-gfx-bounces@lists.freedesktop.org> On Behalf Of Alex Deucher
> Sent: Thursday, April 29, 2021 11:41 AM
> To: Deucher, Alexander <Alexander.Deucher@amd.com>
> Cc: amd-gfx list <amd-gfx@lists.freedesktop.org>
> Subject: Re: [PATCH] drm/amdgpu: Add graphics cache rinse packet for sdma 5.0
>
> Ping?
>
> On Tue, Apr 20, 2021 at 3:28 PM Alex Deucher <alexander.deucher@amd.com> wrote:
> >
> > Add emit mem sync callback for sdma_v5_0
> >
> > In amdgpu sync object test, three threads created jobs to send GFX IB
> > and SDMA IB in sequence. After the first GFX thread joined, sometimes
> > the third thread will reuse the same physical page to store the SDMA
> > IB. There will be a risk that SDMA will read GFX IB in the previous
> > physical page. So it's better to flush the cache before commit sdma
> > IB.
> >
> > Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
> > ---
> > drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c | 28
> > ++++++++++++++++++++++++++
> > 1 file changed, 28 insertions(+)
> >
> > diff --git a/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c
> > b/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c
> > index 920fc6d4a127..d294ef6a625a 100644
> > --- a/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c
> > +++ b/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c
> > @@ -437,6 +437,33 @@ static void sdma_v5_0_ring_emit_ib(struct amdgpu_ring *ring,
> > amdgpu_ring_write(ring, upper_32_bits(csa_mc_addr)); }
> >
> > +/**
> > + * sdma_v5_0_ring_emit_mem_sync - flush the IB by graphics cache
> > +rinse
> > + *
> > + * @ring: amdgpu ring pointer
> > + * @job: job to retrieve vmid from
> > + * @ib: IB object to schedule
> > + *
> > + * flush the IB by graphics cache rinse.
> > + */
> > +static void sdma_v5_0_ring_emit_mem_sync(struct amdgpu_ring *ring) {
> > + uint32_t gcr_cntl =
> > + SDMA_GCR_GL2_INV | SDMA_GCR_GL2_WB | SDMA_GCR_GLM_INV |
> > + SDMA_GCR_GL1_INV | SDMA_GCR_GLV_INV | SDMA_GCR_GLK_INV |
> > + SDMA_GCR_GLI_INV(1);
> > +
> > + /* flush entire cache L0/L1/L2, this can be optimized by performance requirement */
> > + amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_GCR_REQ));
> > + amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD1_BASE_VA_31_7(0));
> > + amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD2_GCR_CONTROL_15_0(gcr_cntl) |
> > + SDMA_PKT_GCR_REQ_PAYLOAD2_BASE_VA_47_32(0));
> > + amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD3_LIMIT_VA_31_7(0) |
> > + SDMA_PKT_GCR_REQ_PAYLOAD3_GCR_CONTROL_18_16(gcr_cntl >> 16));
> > + amdgpu_ring_write(ring, SDMA_PKT_GCR_REQ_PAYLOAD4_LIMIT_VA_47_32(0) |
> > + SDMA_PKT_GCR_REQ_PAYLOAD4_VMID(0));
> > +}
> > +
> > /**
> > * sdma_v5_0_ring_emit_hdp_flush - emit an hdp flush on the DMA ring
> > *
> > @@ -1643,6 +1670,7 @@ static const struct amdgpu_ring_funcs sdma_v5_0_ring_funcs = {
> > 10 + 10 + 10, /* sdma_v5_0_ring_emit_fence x3 for user fence, vm fence */
> > .emit_ib_size = 5 + 7 + 6, /* sdma_v5_0_ring_emit_ib */
> > .emit_ib = sdma_v5_0_ring_emit_ib,
> > + .emit_mem_sync = sdma_v5_0_ring_emit_mem_sync,
> > .emit_fence = sdma_v5_0_ring_emit_fence,
> > .emit_pipeline_sync = sdma_v5_0_ring_emit_pipeline_sync,
> > .emit_vm_flush = sdma_v5_0_ring_emit_vm_flush,
> > --
> > 2.30.2
> >
> > _______________________________________________
> > amd-gfx mailing list
> > amd-gfx@lists.freedesktop.org
> > https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flist
> > s.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&data=04%7C01%7CJi
> > ansong.Chen%40amd.com%7C1b777b80b17145712a7b08d90ac0a1f4%7C3dd8961fe48
> > 84e608e11a82d994e183d%7C0%7C0%7C637552644738458840%7CUnknown%7CTWFpbGZ
> > sb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3
> > D%7C1000&sdata=T9jZUIJIQHS2gY8bU%2F7uM1ealAP3qxelkg2Slj3JASA%3D&am
> > p;reserved=0
> _______________________________________________
> amd-gfx mailing list
> amd-gfx@lists.freedesktop.org
> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&data=04%7C01%7CJiansong.Chen%40amd.com%7C1b777b80b17145712a7b08d90ac0a1f4%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637552644738458840%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=T9jZUIJIQHS2gY8bU%2F7uM1ealAP3qxelkg2Slj3JASA%3D&reserved=0
_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
^ permalink raw reply [flat|nested] 5+ messages in thread
end of thread, other threads:[~2021-04-29 16:03 UTC | newest]
Thread overview: 5+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2021-04-20 19:27 [PATCH] drm/amdgpu: Add graphics cache rinse packet for sdma 5.0 Alex Deucher
2021-04-29 3:40 ` Alex Deucher
2021-04-29 3:43 ` Zhang, Hawking
2021-04-29 8:11 ` Chen, Jiansong (Simon)
2021-04-29 16:03 ` Alex Deucher
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).