Nouveau - Oct 2015 - [PATCH] nv50, nvc0: don't base decisions on available pushbuf space

On Sat, Oct 10, 2015 at 3:41 PM, Samuel Pitoiset
<samuel.pitoiset at gmail.com> wrote:
> This patch looks fine except that it should be a bit more normalized. I
> mean, sometimes you break when PUSH_SPACE fails, sometimes not. Same for
> PUSH_SPACE calls, sometimes you add it sometimes not.
Meh. We need to get our error checking situation straight, but this
isn't the patch to do it in.
>
> Did you run a full piglit test this time ? :)
Nope, but I ran a full piglit before this patch. Almost took down my
box. Probably won't be running it again for this patch.
>
> See my comment below.
>
>
> On 10/10/2015 11:09 AM, Ilia Mirkin wrote:
>>
>> We still have to push everything out, might as well kick earlier and
>> flip pushbufs when we know we'll need it. This resolves some issues
with
>> the new policy of making sure that we always leave a bit of room at the
>> end for fences.
>>
>> Signed-off-by: Ilia Mirkin <imirkin at alum.mit.edu>
>> Cc: mesa-stable at lists.freedesktop.org
>> ---
>>   src/gallium/drivers/nouveau/nv50/nv50_shader_state.c |  9 ++-------
>>   src/gallium/drivers/nouveau/nv50/nv50_transfer.c     | 16
>> +++-------------
>>   src/gallium/drivers/nouveau/nvc0/nvc0_transfer.c     | 20
>> +++++---------------
>>   3 files changed, 10 insertions(+), 35 deletions(-)
>>
>> diff --git a/src/gallium/drivers/nouveau/nv50/nv50_shader_state.c
>> b/src/gallium/drivers/nouveau/nv50/nv50_shader_state.c
>> index fdde11f..941555f 100644
>> --- a/src/gallium/drivers/nouveau/nv50/nv50_shader_state.c
>> +++ b/src/gallium/drivers/nouveau/nv50/nv50_shader_state.c
>> @@ -65,14 +65,9 @@ nv50_constbufs_validate(struct nv50_context *nv50)
>>                  PUSH_DATA (push, (b << 12) | (i << 8) | p
| 1);
>>               }
>>               while (words) {
>> -               unsigned nr;
>> -
>> -               if (!PUSH_SPACE(push, 16))
>> -                  break;
>> -               nr = PUSH_AVAIL(push);
>> -               assert(nr >= 16);
>> -               nr = MIN2(MIN2(nr - 3, words),
NV04_PFIFO_MAX_PACKET_LEN);
>> +               unsigned nr = MIN2(words, NV04_PFIFO_MAX_PACKET_LEN);
>>   +               PUSH_SPACE(push, nr + 3);
>
>
> This PUSH_SPACE call doesn't seem to be needed for me because
> NV50_PUSH_EXPLICIT_SPACE_CHECKING is not set and the following BEGIN_XXX
> calls will allocate space.
I want to ensure that both of the below commands are in the same
batch. Not sure if it's necessary, but... don't want to find out. They
were in the same batch before. And this batch stuff is what was
causing the M2MF errors I was seeing earlier.
>
>
>>                  BEGIN_NV04(push, NV50_3D(CB_ADDR), 1);
>>                  PUSH_DATA (push, (start << 8) | b);
>>                  BEGIN_NI04(push, NV50_3D(CB_DATA(0)), nr);
>> diff --git a/src/gallium/drivers/nouveau/nv50/nv50_transfer.c
>> b/src/gallium/drivers/nouveau/nv50/nv50_transfer.c
>> index be51407..9a3fd1e 100644
>> --- a/src/gallium/drivers/nouveau/nv50/nv50_transfer.c
>> +++ b/src/gallium/drivers/nouveau/nv50/nv50_transfer.c
>> @@ -187,14 +187,7 @@ nv50_sifc_linear_u8(struct nouveau_context *nv,
>>      PUSH_DATA (push, 0);
>>        while (count) {
>> -      unsigned nr;
>> -
>> -      if (!PUSH_SPACE(push, 16))
>> -         break;
>> -      nr = PUSH_AVAIL(push);
>> -      assert(nr >= 16);
>> -      nr = MIN2(count, nr - 1);
>> -      nr = MIN2(nr, NV04_PFIFO_MAX_PACKET_LEN);
>> +      unsigned nr = MIN2(count, NV04_PFIFO_MAX_PACKET_LEN);
>>           BEGIN_NI04(push, NV50_2D(SIFC_DATA), nr);
>>         PUSH_DATAp(push, src, nr);
>> @@ -395,12 +388,9 @@ nv50_cb_push(struct nouveau_context *nv,
>>      nouveau_pushbuf_validate(push);
>>        while (words) {
>> -      unsigned nr;
>> -
>> -      nr = PUSH_AVAIL(push);
>> -      nr = MIN2(nr - 7, words);
>> -      nr = MIN2(nr, NV04_PFIFO_MAX_PACKET_LEN - 1);
>> +      unsigned nr = MIN2(words, NV04_PFIFO_MAX_PACKET_LEN);
>>   +      PUSH_SPACE(push, nr + 7);
>>         BEGIN_NV04(push, NV50_3D(CB_DEF_ADDRESS_HIGH), 3);
>>         PUSH_DATAh(push, bo->offset + base);
>>         PUSH_DATA (push, bo->offset + base);
>> diff --git a/src/gallium/drivers/nouveau/nvc0/nvc0_transfer.c
>> b/src/gallium/drivers/nouveau/nvc0/nvc0_transfer.c
>> index aaec60a..d459dd6 100644
>> --- a/src/gallium/drivers/nouveau/nvc0/nvc0_transfer.c
>> +++ b/src/gallium/drivers/nouveau/nvc0/nvc0_transfer.c
>> @@ -188,14 +188,10 @@ nvc0_m2mf_push_linear(struct nouveau_context *nv,
>>      nouveau_pushbuf_validate(push);
>>        while (count) {
>> -      unsigned nr;
>> +      unsigned nr = MIN2(count, NV04_PFIFO_MAX_PACKET_LEN);
>>   -      if (!PUSH_SPACE(push, 16))
>> +      if (!PUSH_SPACE(push, nr + 9))
>>            break;
>> -      nr = PUSH_AVAIL(push);
>> -      assert(nr >= 16);
>> -      nr = MIN2(count, nr - 9);
>> -      nr = MIN2(nr, NV04_PFIFO_MAX_PACKET_LEN);
>>           BEGIN_NVC0(push, NVC0_M2MF(OFFSET_OUT_HIGH), 2);
>>         PUSH_DATAh(push, dst->offset + offset);
>> @@ -234,14 +230,10 @@ nve4_p2mf_push_linear(struct nouveau_context *nv,
>>      nouveau_pushbuf_validate(push);
>>        while (count) {
>> -      unsigned nr;
>> +      unsigned nr = MIN2(count, (NV04_PFIFO_MAX_PACKET_LEN - 1));
>>   -      if (!PUSH_SPACE(push, 16))
>> +      if (!PUSH_SPACE(push, nr + 10))
>>            break;
>> -      nr = PUSH_AVAIL(push);
>> -      assert(nr >= 16);
>> -      nr = MIN2(count, nr - 8);
>> -      nr = MIN2(nr, (NV04_PFIFO_MAX_PACKET_LEN - 1));
>>           BEGIN_NVC0(push, NVE4_P2MF(UPLOAD_DST_ADDRESS_HIGH), 2);
>>         PUSH_DATAh(push, dst->offset + offset);
>> @@ -571,9 +563,7 @@ nvc0_cb_bo_push(struct nouveau_context *nv,
>>      PUSH_DATA (push, bo->offset + base);
>>        while (words) {
>> -      unsigned nr = PUSH_AVAIL(push);
>> -      nr = MIN2(nr, words);
>> -      nr = MIN2(nr, NV04_PFIFO_MAX_PACKET_LEN - 1);
>> +      unsigned nr = MIN2(words, NV04_PFIFO_MAX_PACKET_LEN - 1);
>>           PUSH_SPACE(push, nr + 2);
>>         PUSH_REFN (push, bo, NOUVEAU_BO_WR | domain);
>
>

On 10/10/2015 09:42 PM, Ilia Mirkin wrote:
> On Sat, Oct 10, 2015 at 3:41 PM, Samuel Pitoiset
> <samuel.pitoiset at gmail.com> wrote:
>> This patch looks fine except that it should be a bit more normalized. I
>> mean, sometimes you break when PUSH_SPACE fails, sometimes not. Same
for
>> PUSH_SPACE calls, sometimes you add it sometimes not.
> Meh. We need to get our error checking situation straight, but this
> isn't the patch to do it in.
Yeah, but this needs to be clarified.
>
>> Did you run a full piglit test this time ? :)
> Nope, but I ran a full piglit before this patch. Almost took down my
> box. Probably won't be running it again for this patch.
Ok, I'll run a full piglit this night then.
>
>> See my comment below.
>>
>>
>> On 10/10/2015 11:09 AM, Ilia Mirkin wrote:
>>> We still have to push everything out, might as well kick earlier
and
>>> flip pushbufs when we know we'll need it. This resolves some
issues with
>>> the new policy of making sure that we always leave a bit of room at
the
>>> end for fences.
>>>
>>> Signed-off-by: Ilia Mirkin <imirkin at alum.mit.edu>
>>> Cc: mesa-stable at lists.freedesktop.org
>>> ---
>>>    src/gallium/drivers/nouveau/nv50/nv50_shader_state.c |  9
++-------
>>>    src/gallium/drivers/nouveau/nv50/nv50_transfer.c     | 16
>>> +++-------------
>>>    src/gallium/drivers/nouveau/nvc0/nvc0_transfer.c     | 20
>>> +++++---------------
>>>    3 files changed, 10 insertions(+), 35 deletions(-)
>>>
>>> diff --git a/src/gallium/drivers/nouveau/nv50/nv50_shader_state.c
>>> b/src/gallium/drivers/nouveau/nv50/nv50_shader_state.c
>>> index fdde11f..941555f 100644
>>> --- a/src/gallium/drivers/nouveau/nv50/nv50_shader_state.c
>>> +++ b/src/gallium/drivers/nouveau/nv50/nv50_shader_state.c
>>> @@ -65,14 +65,9 @@ nv50_constbufs_validate(struct nv50_context
*nv50)
>>>                   PUSH_DATA (push, (b << 12) | (i << 8)
| p | 1);
>>>                }
>>>                while (words) {
>>> -               unsigned nr;
>>> -
>>> -               if (!PUSH_SPACE(push, 16))
>>> -                  break;
>>> -               nr = PUSH_AVAIL(push);
>>> -               assert(nr >= 16);
>>> -               nr = MIN2(MIN2(nr - 3, words),
NV04_PFIFO_MAX_PACKET_LEN);
>>> +               unsigned nr = MIN2(words,
NV04_PFIFO_MAX_PACKET_LEN);
>>>    +               PUSH_SPACE(push, nr + 3);
>>
>> This PUSH_SPACE call doesn't seem to be needed for me because
>> NV50_PUSH_EXPLICIT_SPACE_CHECKING is not set and the following
BEGIN_XXX
>> calls will allocate space.
> I want to ensure that both of the below commands are in the same
> batch. Not sure if it's necessary, but... don't want to find out.
They
> were in the same batch before. And this batch stuff is what was
> causing the M2MF errors I was seeing earlier.
>
>>
>>>                   BEGIN_NV04(push, NV50_3D(CB_ADDR), 1);
>>>                   PUSH_DATA (push, (start << 8) | b);
>>>                   BEGIN_NI04(push, NV50_3D(CB_DATA(0)), nr);
>>> diff --git a/src/gallium/drivers/nouveau/nv50/nv50_transfer.c
>>> b/src/gallium/drivers/nouveau/nv50/nv50_transfer.c
>>> index be51407..9a3fd1e 100644
>>> --- a/src/gallium/drivers/nouveau/nv50/nv50_transfer.c
>>> +++ b/src/gallium/drivers/nouveau/nv50/nv50_transfer.c
>>> @@ -187,14 +187,7 @@ nv50_sifc_linear_u8(struct nouveau_context
*nv,
>>>       PUSH_DATA (push, 0);
>>>         while (count) {
>>> -      unsigned nr;
>>> -
>>> -      if (!PUSH_SPACE(push, 16))
>>> -         break;
>>> -      nr = PUSH_AVAIL(push);
>>> -      assert(nr >= 16);
>>> -      nr = MIN2(count, nr - 1);
>>> -      nr = MIN2(nr, NV04_PFIFO_MAX_PACKET_LEN);
>>> +      unsigned nr = MIN2(count, NV04_PFIFO_MAX_PACKET_LEN);
>>>            BEGIN_NI04(push, NV50_2D(SIFC_DATA), nr);
>>>          PUSH_DATAp(push, src, nr);
>>> @@ -395,12 +388,9 @@ nv50_cb_push(struct nouveau_context *nv,
>>>       nouveau_pushbuf_validate(push);
>>>         while (words) {
>>> -      unsigned nr;
>>> -
>>> -      nr = PUSH_AVAIL(push);
>>> -      nr = MIN2(nr - 7, words);
>>> -      nr = MIN2(nr, NV04_PFIFO_MAX_PACKET_LEN - 1);
>>> +      unsigned nr = MIN2(words, NV04_PFIFO_MAX_PACKET_LEN);
>>>    +      PUSH_SPACE(push, nr + 7);
>>>          BEGIN_NV04(push, NV50_3D(CB_DEF_ADDRESS_HIGH), 3);
>>>          PUSH_DATAh(push, bo->offset + base);
>>>          PUSH_DATA (push, bo->offset + base);
>>> diff --git a/src/gallium/drivers/nouveau/nvc0/nvc0_transfer.c
>>> b/src/gallium/drivers/nouveau/nvc0/nvc0_transfer.c
>>> index aaec60a..d459dd6 100644
>>> --- a/src/gallium/drivers/nouveau/nvc0/nvc0_transfer.c
>>> +++ b/src/gallium/drivers/nouveau/nvc0/nvc0_transfer.c
>>> @@ -188,14 +188,10 @@ nvc0_m2mf_push_linear(struct nouveau_context
*nv,
>>>       nouveau_pushbuf_validate(push);
>>>         while (count) {
>>> -      unsigned nr;
>>> +      unsigned nr = MIN2(count, NV04_PFIFO_MAX_PACKET_LEN);
>>>    -      if (!PUSH_SPACE(push, 16))
>>> +      if (!PUSH_SPACE(push, nr + 9))
>>>             break;
>>> -      nr = PUSH_AVAIL(push);
>>> -      assert(nr >= 16);
>>> -      nr = MIN2(count, nr - 9);
>>> -      nr = MIN2(nr, NV04_PFIFO_MAX_PACKET_LEN);
>>>            BEGIN_NVC0(push, NVC0_M2MF(OFFSET_OUT_HIGH), 2);
>>>          PUSH_DATAh(push, dst->offset + offset);
>>> @@ -234,14 +230,10 @@ nve4_p2mf_push_linear(struct nouveau_context
*nv,
>>>       nouveau_pushbuf_validate(push);
>>>         while (count) {
>>> -      unsigned nr;
>>> +      unsigned nr = MIN2(count, (NV04_PFIFO_MAX_PACKET_LEN - 1));
>>>    -      if (!PUSH_SPACE(push, 16))
>>> +      if (!PUSH_SPACE(push, nr + 10))
>>>             break;
>>> -      nr = PUSH_AVAIL(push);
>>> -      assert(nr >= 16);
>>> -      nr = MIN2(count, nr - 8);
>>> -      nr = MIN2(nr, (NV04_PFIFO_MAX_PACKET_LEN - 1));
>>>            BEGIN_NVC0(push, NVE4_P2MF(UPLOAD_DST_ADDRESS_HIGH), 2);
>>>          PUSH_DATAh(push, dst->offset + offset);
>>> @@ -571,9 +563,7 @@ nvc0_cb_bo_push(struct nouveau_context *nv,
>>>       PUSH_DATA (push, bo->offset + base);
>>>         while (words) {
>>> -      unsigned nr = PUSH_AVAIL(push);
>>> -      nr = MIN2(nr, words);
>>> -      nr = MIN2(nr, NV04_PFIFO_MAX_PACKET_LEN - 1);
>>> +      unsigned nr = MIN2(words, NV04_PFIFO_MAX_PACKET_LEN - 1);
>>>            PUSH_SPACE(push, nr + 2);
>>>          PUSH_REFN (push, bo, NOUVEAU_BO_WR | domain);
>>

On Sat, Oct 10, 2015 at 3:55 PM, Samuel Pitoiset
<samuel.pitoiset at gmail.com> wrote:
>
>
> On 10/10/2015 09:42 PM, Ilia Mirkin wrote:
>>
>> On Sat, Oct 10, 2015 at 3:41 PM, Samuel Pitoiset
>> <samuel.pitoiset at gmail.com> wrote:
>>>
>>> This patch looks fine except that it should be a bit more
normalized. I
>>> mean, sometimes you break when PUSH_SPACE fails, sometimes not.
Same for
>>> PUSH_SPACE calls, sometimes you add it sometimes not.
>>
>> Meh. We need to get our error checking situation straight, but this
>> isn't the patch to do it in.
>
>
> Yeah, but this needs to be clarified.
What does?