[PATCH] drm/amdgpu:implement CONTEXT_CONTROL (v3)
Liu, Monk
Monk.Liu at amd.com
Mon Sep 5 04:09:44 UTC 2016
Okay, just put it simple
The approach my patch provided is absolutely correct, and I'm pretty sure of it, otherwise our close OGL UMD & windows d3d game already crashed for million times.
First of all: Move CONTEXT_CONTROL into RB is a must and the correct thing, and this is not a questionable patch.
Then give some definition:
UMD is aware of context switch within the process,
KMD is aware of context switch cross processes, because KMD give each context a unique ID globally, so this ID can detect process switch as well as context switch (within a process) easily.
Now back to your questions:
#1
So if I understand correctly, the new behavior is that the first submit containing a preamble always executes the loads in the preamble even if there is no context switch. The old behavior is that in that situation th preamble would be skipped in the new situation. Why do we want the new behavior?>If the application wants the loads to execute even without a context switch, it should not mark the IB as a preamble with AMDGPU_IB_FLAG_PREAMBLE.
[ML] there is no harm kmd to that, isn't it ? besides, no skip the first Preamble CEIB is the correct choice compared with skip it, no matter why UMD mark it as FLAG_PREAMBLE.
#2
If there is no IB with AMDGPU_IB_FLAG_PREAMBLE, then the CE_LOAD bit always gets disabled. Furthermore if there is a CE_LOAD bit, and no context switch the CE_LOAD bit also gets disabled for IB's without AMDGPU_IB_FLAG_PREAMBLE.
I think this is a bad move, as there are some uses for loading CE RAM that are not dependent on context switches, such as preloading things into L2 cache, or switching shader uniforms in CE RAM when a different shader gets bound. Therefore I think that the CE_LOAD bit should always be enabled for IB's without AMDGPU_IB_FLAG_PREAMBLE.
[ML] why my patch/scheme doesn’t show anything wrong when I run benchmark (Unigine heaven) ?
I admit I uses AMD close source OGL UMD, If KMD detects a context switch (including context switch within one process or process switch) then LOAD_CE_RAM is also kept.
For purpose of " such as preloading things into L2 cache, or switching shader uniforms in CE RAM when a different shader gets bound.".... that could be done by CE IB ( instead of Preamble CEIB) via commands like "write_const_ram, dump_const_ram"
#3
Furthermore, with this patch the preamble IB's always get executed and loads disabled with CONTEXT_CONTROL. As e.g. mesa uses its own CONTEXT_CONTROL (and we can't change that for old versions of mesa) this would override the kernel CONTEXT_CONTROL and always execute the loads.
[ML] I must say MESA use CONTEXT_CONTROL is really bad idea, MESA couldn't detect the context switch triggered by process switch.
No matter what reason, this wrong approach need be fixed.
#4
I also miss the CE_LOAD bit in the CONTEXT_CONTROL for gfx7. Does it not need it?
[ML] for GFX7, the CONTEXT_CONTROL doesn't support CE_LOAD bit. So CE_LOAD_RAM will always be kept from KMD perspective (which may sacrifice performance compared with GFX8).
#5
I would prefer keeping the old system for preamble IB's and just adding a generic CONTEXT_CONTROL that always enables the CE loads. I don't have an opinion the non-CE loads though, as I've never found a reason to use them.
[ML] No, that way our close UMD won't work correctly.
You can insist the wrong way although, and if you cannot accept the correct scheme of CONTEXT_CONTROL and change MESA's wrong behavior,
I'll consider upstream amdgpu KMD refuse to support SR-IOV/virtualization.
You need think it twice, you are insisting a wrong design/approach although it runs for years.
BR Monk
BR Monk
-----Original Message-----
From: Bas Nieuwenhuizen [mailto:bas at basnieuwenhuizen.nl]
Sent: Friday, September 02, 2016 12:09 AM
To: Liu, Monk <Monk.Liu at amd.com>
Cc: Christian König <deathsimple at vodafone.de>; amd-gfx at lists.freedesktop.org
Subject: Re: [PATCH] drm/amdgpu:implement CONTEXT_CONTROL (v3)
On Thu, Sep 1, 2016 at 12:55 PM, Liu, Monk <Monk.Liu at amd.com> wrote:
>> Why does that makes a difference if it is seen for the first time?
>>
>> [ml] if it is presented for the first time for belonging ctx, means even current CS do not involve context switch, we still need keep the actions in preamble IB.
>> Usually if current CS is from the same cntx of previous CS, that means no ctx switch occurs, so we can skip the actions in preamble IB. but above case is the exception.
>
> Can't userspace just not set the preamble flag for the first submit with a preamble? I think that would result in the same behavior, unless having two non-preamble CE IB's in a single submit is an issue.
>
> - Bas
>
>
> [ML] I'm confused, what's your point?
So if I understand correctly, the new behavior is that the first submit containing a preamble always executes the loads in the preamble even if there is no context switch. The old behavior is that in that situation the preamble would be skipped in the new situation. Why do we want the new behavior? If the application wants the loads to execute even without a context switch, it should not mark the IB as a preamble with AMDGPU_IB_FLAG_PREAMBLE.
On inspecting the patch more closely I think there are more issues with this patch.
If there is no IB with AMDGPU_IB_FLAG_PREAMBLE, then the CE_LOAD bit always gets disabled. Furthermore if there is a CE_LOAD bit, and no context switch the CE_LOAD bit also gets disabled for IB's without AMDGPU_IB_FLAG_PREAMBLE.
I think this is a bad move, as there are some uses for loading CE RAM that are not dependent on context switches, such as preloading things into L2 cache, or switching shader uniforms in CE RAM when a different shader gets bound. Therefore I think that the CE_LOAD bit should always be enabled for IB's without AMDGPU_IB_FLAG_PREAMBLE.
Furthermore, with this patch the preamble IB's always get executed and loads disabled with CONTEXT_CONTROL. As e.g. mesa uses its own CONTEXT_CONTROL (and we can't change that for old versions of mesa) this would override the kernel CONTEXT_CONTROL and always execute the loads.
I also miss the CE_LOAD bit in the CONTEXT_CONTROL for gfx7. Does it not need it?
I would prefer keeping the old system for preamble IB's and just adding a generic CONTEXT_CONTROL that always enables the CE loads. I don't have an opinion the non-CE loads though, as I've never found a reason to use them.
- Bas
> With this patch, preamble_flag is not needed at all.
> Without this patch, many original assumption and logic is not correct.
> Besides, CONTEXT_CONTROL not only deals CE but also deal DE.
>
> BR Monk
>
>
> -----Original Message-----
> From: Bas Nieuwenhuizen [mailto:bas at basnieuwenhuizen.nl]
> Sent: Thursday, September 01, 2016 4:19 PM
> To: Liu, Monk <Monk.Liu at amd.com>
> Cc: Christian König <deathsimple at vodafone.de>;
> amd-gfx at lists.freedesktop.org
> Subject: Re: [PATCH] drm/amdgpu:implement CONTEXT_CONTROL (v3)
>
> On Thu, Sep 1, 2016 at 9:37 AM, Liu, Monk <Monk.Liu at amd.com> wrote:
>>
>>
>> -----Original Message-----
>> From: amd-gfx [mailto:amd-gfx-bounces at lists.freedesktop.org] On
>> Behalf Of Christian K?nig
>> Sent: Wednesday, August 31, 2016 7:53 PM
>> To: Liu, Monk <Monk.Liu at amd.com>; amd-gfx at lists.freedesktop.org
>> Subject: Re: [PATCH] drm/amdgpu:implement CONTEXT_CONTROL (v3)
>>
>> Looks good to me in general, a few nit picks and sugegstions below.
>>
>> Am 31.08.2016 um 05:49 schrieb Monk Liu:
>>> v1:
>>> for gfx8, use CONTEXT_CONTROL package to dynamically skip preamble
>>> CEIB and other load_xxx command in sequence.
>>>
>>> v2:
>>> support GFX7 as well, and bump up version.
>>> remove cntxcntl in compute ring funcs because CPC doesn't support
>>> this packet.
>>>
>>> v3: fix reduntant judgement in cntxcntl.
>>>
>>> Change-Id: I4b87ca84ea8c11ba4f7fb4c0e8a5be537ccde851
>>> Signed-off-by: Monk Liu <Monk.Liu at amd.com>
>>>
>>> Change-Id: I5d24c1bb5c14190ce4adeb6a331ee3d92b3d5c83
>>> Signed-off-by: Monk Liu <Monk.Liu at amd.com>
>>
>> Only one signed of by line is enough and remove the change-ids.
>>
>>> ---
>>> drivers/gpu/drm/amd/amdgpu/amdgpu.h | 9 +++++++++
>>> drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 12 ++++++++++++
>>> drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c | 3 ++-
>>> drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c | 16 +++++++++-------
>>> drivers/gpu/drm/amd/amdgpu/gfx_v7_0.c | 20 ++++++++++++++++++++
>>> drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c | 30 ++++++++++++++++++++++++++++++
>>> 6 files changed, 82 insertions(+), 8 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
>>> index 1254410..0de5f08 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
>>> @@ -321,6 +321,7 @@ struct amdgpu_ring_funcs {
>>> void (*begin_use)(struct amdgpu_ring *ring);
>>> void (*end_use)(struct amdgpu_ring *ring);
>>> void (*emit_switch_buffer) (struct amdgpu_ring *ring);
>>> + void (*emit_cntxcntl) (struct amdgpu_ring *ring, uint32_t
>>> + flags);
>>> };
>>>
>>> /*
>>> @@ -965,6 +966,7 @@ struct amdgpu_ctx {
>>> spinlock_t ring_lock;
>>> struct fence **fences;
>>> struct amdgpu_ctx_ring rings[AMDGPU_MAX_RINGS];
>>> + bool preamble_presented;
>>> };
>>>
>>> struct amdgpu_ctx_mgr {
>>> @@ -1227,8 +1229,13 @@ struct amdgpu_cs_parser {
>>>
>>> /* user fence */
>>> struct amdgpu_bo_list_entry uf_entry;
>>> + bool preamble_present; /* True means this command submit
>>> +involves a preamble IB */
>>
>> We only need this in amdgpu_cs_ib_fill() don't we? See below as well.
>>
>> [ML] seems good advice
>>
>>> };
>>>
>>> +#define PREAMBLE_IB_PRESENT (1 << 0) /* bit set means command submit involves a preamble IB */
>>> +#define PREAMBLE_IB_PRESENT_FIRST (1 << 1) /* bit set means preamble IB is first presented in belonging context */
>>
>> Why does that makes a difference if it is seen for the first time?
>>
>> [ml] if it is presented for the first time for belonging ctx, means even current CS do not involve context switch, we still need keep the actions in preamble IB.
>> Usually if current CS is from the same cntx of previous CS, that means no ctx switch occurs, so we can skip the actions in preamble IB. but above case is the exception.
>
> Can't userspace just not set the preamble flag for the first submit with a preamble? I think that would result in the same behavior, unless having two non-preamble CE IB's in a single submit is an issue.
>
> - Bas
>
>>
>>> +#define HAVE_CTX_SWITCH (1 << 2) /* bit set means context switch occured */
>>> +
>>> struct amdgpu_job {
>>> struct amd_sched_job base;
>>> struct amdgpu_device *adev;
>>> @@ -1237,6 +1244,7 @@ struct amdgpu_job {
>>> struct amdgpu_sync sync;
>>> struct amdgpu_ib *ibs;
>>> struct fence *fence; /* the hw fence */
>>> + uint32_t preamble_status;
>>> uint32_t num_ibs;
>>> void *owner;
>>> uint64_t fence_ctx; /* the fence_context this job uses */
>>> @@ -2264,6 +2272,7 @@ amdgpu_get_sdma_instance(struct amdgpu_ring *ring)
>>> #define amdgpu_ring_emit_hdp_flush(r) (r)->funcs->emit_hdp_flush((r))
>>> #define amdgpu_ring_emit_hdp_invalidate(r) (r)->funcs->emit_hdp_invalidate((r))
>>> #define amdgpu_ring_emit_switch_buffer(r)
>>> (r)->funcs->emit_switch_buffer((r))
>>> +#define amdgpu_ring_emit_cntxcntl(r, d)
>>> +(r)->funcs->emit_cntxcntl((r), (d))
>>> #define amdgpu_ring_pad_ib(r, ib) ((r)->funcs->pad_ib((r), (ib)))
>>> #define amdgpu_ring_init_cond_exec(r) (r)->funcs->init_cond_exec((r))
>>> #define amdgpu_ring_patch_cond_exec(r,o)
>>> (r)->funcs->patch_cond_exec((r),(o))
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
>>> index 2d4e005..6d8c050 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c
>>> @@ -792,6 +792,9 @@ static int amdgpu_cs_ib_fill(struct amdgpu_device *adev,
>>> if (r)
>>> return r;
>>>
>>> + if (ib->flags & AMDGPU_IB_FLAG_PREAMBLE)
>>> + parser->preamble_present = true;
>>> +
>>> if (parser->job->ring && parser->job->ring != ring)
>>> return -EINVAL;
>>>
>>> @@ -930,6 +933,12 @@ static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
>>> return r;
>>> }
>>>
>>> + if (p->preamble_present) {
>>> + job->preamble_status |= PREAMBLE_IB_PRESENT;
>>> + if (!p->ctx->preamble_presented)
>>> + job->preamble_status |= PREAMBLE_IB_PRESENT_FIRST;
>>> + }
>>> +
>>
>> Better move this to the end of amdgpu_cs_ib_fill() where we allocate the IBs as well.
>> [ML] okay, good change.
>>
>>
>>
>>> job->owner = p->filp;
>>> job->fence_ctx = entity->fence_context;
>>> p->fence = fence_get(&job->base.s_fence->finished);
>>> @@ -940,6 +949,9 @@ static int amdgpu_cs_submit(struct amdgpu_cs_parser *p,
>>> trace_amdgpu_cs_ioctl(job);
>>> amd_sched_entity_push_job(&job->base);
>>>
>>> + if (p->preamble_present)
>>> + p->ctx->preamble_presented = true;
>>> +
>>> return 0;
>>> }
>>>
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
>>> index 56c85e6..44db0ab 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_drv.c
>>> @@ -55,9 +55,10 @@
>>> * - 3.3.0 - Add VM support for UVD on supported hardware.
>>> * - 3.4.0 - Add AMDGPU_INFO_NUM_EVICTIONS.
>>> * - 3.5.0 - Add support for new UVD_NO_OP register.
>>> + * - 3.6.0 - UMD doesn't/shouldn't need to use CONTEXT_CONTROL in
>>> + IB, KMD should do it
>>> */
>>> #define KMS_DRIVER_MAJOR 3
>>> -#define KMS_DRIVER_MINOR 5
>>> +#define KMS_DRIVER_MINOR 6
>>> #define KMS_DRIVER_PATCHLEVEL 0
>>>
>>> int amdgpu_vram_limit = 0;
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
>>> index 04263f0..b12b5ba 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
>>> @@ -121,10 +121,11 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
>>> {
>>> struct amdgpu_device *adev = ring->adev;
>>> struct amdgpu_ib *ib = &ibs[0];
>>> - bool skip_preamble, need_ctx_switch;
>>> + bool need_ctx_switch;
>>> unsigned patch_offset = ~0;
>>> struct amdgpu_vm *vm;
>>> uint64_t fence_ctx;
>>> + uint32_t status = 0;
>>>
>>> unsigned i;
>>> int r = 0;
>>> @@ -174,15 +175,16 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
>>> /* always set cond_exec_polling to CONTINUE */
>>> *ring->cond_exe_cpu_addr = 1;
>>>
>>> - skip_preamble = ring->current_ctx == fence_ctx;
>>> need_ctx_switch = ring->current_ctx != fence_ctx;
>>> + if (job && ring->funcs->emit_cntxcntl) {
>>> + if (need_ctx_switch)
>>> + status |= HAVE_CTX_SWITCH;
>>> + status |= job->preamble_status;
>>> + amdgpu_ring_emit_cntxcntl(ring, status);
>>> + }
>>> +
>>> for (i = 0; i < num_ibs; ++i) {
>>> ib = &ibs[i];
>>> -
>>> - /* drop preamble IBs if we don't have a context switch */
>>> - if ((ib->flags & AMDGPU_IB_FLAG_PREAMBLE) && skip_preamble)
>>> - continue;
>>> -
>>
>> Would be nice to keep this functionality for cases where we don't support emit_cntxcntl (e.g. SI?).
>> [ML] SI support CONTEXT_CONTROL as well, and the package structure is exactly the same as CI.
>>
>>> amdgpu_ring_emit_ib(ring, ib, job ? job->vm_id : 0,
>>> need_ctx_switch);
>>> need_ctx_switch = false; diff --git
>>> a/drivers/gpu/drm/amd/amdgpu/gfx_v7_0.c
>>> b/drivers/gpu/drm/amd/amdgpu/gfx_v7_0.c
>>> index f055d49..0d5addb 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/gfx_v7_0.c
>>> +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v7_0.c
>>> @@ -2096,6 +2096,25 @@ static void gfx_v7_0_ring_emit_ib_compute(struct amdgpu_ring *ring,
>>> amdgpu_ring_write(ring, control);
>>> }
>>>
>>> +static void gfx_v7_ring_emit_cntxcntl(struct amdgpu_ring *ring,
>>> +uint32_t flags) {
>>> + uint32_t dw2 = 0;
>>> +
>>> + dw2 |= 0x80000000; /* set load_enable otherwise this package is just NOPs */
>>> + if (flags & HAVE_CTX_SWITCH) {
>>> + /* set load_global_config & load_global_uconfig */
>>> + dw2 |= 0x8001;
>>> + /* set load_cs_sh_regs */
>>> + dw2 |= 0x01000000;
>>> + /* set load_per_context_state & load_gfx_sh_regs */
>>> + dw2 |= 0x10002;
>>
>> Better define some constants for those.
>>
>> [ML] I'll leave it to other guys when doing cleanups, a little hurry for other jobs now ...
>>
>> Regards,
>> Christian.
>>
>>> + }
>>> +
>>> + amdgpu_ring_write(ring, PACKET3(PACKET3_CONTEXT_CONTROL, 1));
>>> + amdgpu_ring_write(ring, dw2);
>>> + amdgpu_ring_write(ring, 0);
>>> +}
>>> +
>>> /**
>>> * gfx_v7_0_ring_test_ib - basic ring IB test
>>> *
>>> @@ -4929,6 +4948,7 @@ static const struct amdgpu_ring_funcs gfx_v7_0_ring_funcs_gfx = {
>>> .test_ib = gfx_v7_0_ring_test_ib,
>>> .insert_nop = amdgpu_ring_insert_nop,
>>> .pad_ib = amdgpu_ring_generic_pad_ib,
>>> + .emit_cntxcntl = gfx_v7_ring_emit_cntxcntl,
>>> };
>>>
>>> static const struct amdgpu_ring_funcs gfx_v7_0_ring_funcs_compute
>>> = { diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c
>>> b/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c
>>> index 8ba8e42..73f6ffa 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c
>>> +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c
>>> @@ -6085,6 +6085,35 @@ static void gfx_v8_ring_emit_sb(struct amdgpu_ring *ring)
>>> amdgpu_ring_write(ring, 0);
>>> }
>>>
>>> +static void gfx_v8_ring_emit_cntxcntl(struct amdgpu_ring *ring,
>>> +uint32_t flags) {
>>> + uint32_t dw2 = 0;
>>> +
>>> + dw2 |= 0x80000000; /* set load_enable otherwise this package is just NOPs */
>>> + if (flags & HAVE_CTX_SWITCH) {
>>> + /* set load_global_config & load_global_uconfig */
>>> + dw2 |= 0x8001;
>>> + /* set load_cs_sh_regs */
>>> + dw2 |= 0x01000000;
>>> + /* set load_per_context_state & load_gfx_sh_regs for GFX */
>>> + dw2 |= 0x10002;
>>> +
>>> + /* set load_ce_ram if preamble presented */
>>> + if (PREAMBLE_IB_PRESENT & flags)
>>> + dw2 |= 0x10000000;
>>> + } else {
>>> + /* still load_ce_ram if this is the first time preamble presented
>>> + * although there is no context switch happens.
>>> + */
>>> + if (PREAMBLE_IB_PRESENT_FIRST & flags)
>>> + dw2 |= 0x10000000;
>>> + }
>>> +
>>> + amdgpu_ring_write(ring, PACKET3(PACKET3_CONTEXT_CONTROL, 1));
>>> + amdgpu_ring_write(ring, dw2);
>>> + amdgpu_ring_write(ring, 0);
>>> +}
>>> +
>>> static void gfx_v8_0_set_gfx_eop_interrupt_state(struct amdgpu_device *adev,
>>> enum amdgpu_interrupt_state state)
>>> {
>>> @@ -6267,6 +6296,7 @@ static const struct amdgpu_ring_funcs gfx_v8_0_ring_funcs_gfx = {
>>> .insert_nop = amdgpu_ring_insert_nop,
>>> .pad_ib = amdgpu_ring_generic_pad_ib,
>>> .emit_switch_buffer = gfx_v8_ring_emit_sb,
>>> + .emit_cntxcntl = gfx_v8_ring_emit_cntxcntl,
>>> };
>>>
>>> static const struct amdgpu_ring_funcs gfx_v8_0_ring_funcs_compute
>>> = {
>>
>>
>> _______________________________________________
>> amd-gfx mailing list
>> amd-gfx at lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/amd-gfx
>> _______________________________________________
>> amd-gfx mailing list
>> amd-gfx at lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/amd-gfx
More information about the amd-gfx
mailing list