Nouveau - Jan 2014 - [Mesa-dev] [PATCH] nv50, nvc0: only clear out the buffers that we were asked to clear

On Thu, Jan 23, 2014 at 3:11 PM, Emil Velikov <emil.l.velikov at
gmail.com> wrote:
> On 17/01/14 02:23, Ilia Mirkin wrote:
>> Fixes fbo-drawbuffers-none glClearBuffer piglit test.
>>
>> Signed-off-by: Ilia Mirkin <imirkin at alum.mit.edu>
>> ---
>>
>> Only tested on nv50, but implementations seem similar enough.
>>
>>  src/gallium/drivers/nouveau/nv50/nv50_surface.c | 17 +++++++++--------
>>  src/gallium/drivers/nouveau/nvc0/nvc0_surface.c | 17 +++++++++--------
>>  2 files changed, 18 insertions(+), 16 deletions(-)
>>
>> diff --git a/src/gallium/drivers/nouveau/nv50/nv50_surface.c
b/src/gallium/drivers/nouveau/nv50/nv50_surface.c
>> index 358f57a..eb27429 100644
>> --- a/src/gallium/drivers/nouveau/nv50/nv50_surface.c
>> +++ b/src/gallium/drivers/nouveau/nv50/nv50_surface.c
>> @@ -408,9 +408,6 @@ nv50_clear(struct pipe_context *pipe, unsigned
buffers,
>>        PUSH_DATAf(push, color->f[1]);
>>        PUSH_DATAf(push, color->f[2]);
>>        PUSH_DATAf(push, color->f[3]);
>> -      mode >> -         NV50_3D_CLEAR_BUFFERS_R |
NV50_3D_CLEAR_BUFFERS_G |
>> -         NV50_3D_CLEAR_BUFFERS_B | NV50_3D_CLEAR_BUFFERS_A;
>>     }
>>
> I'm not sure why you've dropped the mode from above. I'm
guessing that
> the initial assumption was that if there is a color buffer it must be at
> cbuf[0].
Because I cleared the cbufs separately in a for loop below (the 0x3c
== the RGBA mask). The first RT may not have been there, and that
seemed simpler than the thing that I have now (as evidenced by the
code I have now which has more complex logic).
>
> The corrected version in your github branch looks alot better, it
> handles the above case and does not overwrite the clear buffer on rt0.
Christoph was really keen on doing the common-case color/depth clear
all in one clear buffers command, so yeah -- I redid it that way.
Enjoy the layered version of that in a later commit.
>
> That one is Reviewed-by: Emil Velikov <emil.l.velikov at gmail.com>
>>     if (buffers & PIPE_CLEAR_DEPTH) {
>> @@ -425,12 +422,16 @@ nv50_clear(struct pipe_context *pipe, unsigned
buffers,
>>        mode |= NV50_3D_CLEAR_BUFFERS_S;
>>     }
>>
>> -   BEGIN_NV04(push, NV50_3D(CLEAR_BUFFERS), 1);
>> -   PUSH_DATA (push, mode);
>> -
>> -   for (i = 1; i < fb->nr_cbufs; i++) {
>> +   if (mode) {
>>        BEGIN_NV04(push, NV50_3D(CLEAR_BUFFERS), 1);
>> -      PUSH_DATA (push, (i << 6) | 0x3c);
>> +      PUSH_DATA (push, mode);
>> +   }
>> +
>> +   for (i = 0; i < fb->nr_cbufs; i++) {
>> +      if (buffers & (PIPE_CLEAR_COLOR0 << i)) {
>> +         BEGIN_NV04(push, NV50_3D(CLEAR_BUFFERS), 1);
>> +         PUSH_DATA (push, (i << 6) | 0x3c);
>> +      }
>>     }
>>  }
>>
>> diff --git a/src/gallium/drivers/nouveau/nvc0/nvc0_surface.c
b/src/gallium/drivers/nouveau/nvc0/nvc0_surface.c
>> index 5375bd4..0c12bce 100644
>> --- a/src/gallium/drivers/nouveau/nvc0/nvc0_surface.c
>> +++ b/src/gallium/drivers/nouveau/nvc0/nvc0_surface.c
>> @@ -427,9 +427,6 @@ nvc0_clear(struct pipe_context *pipe, unsigned
buffers,
>>        PUSH_DATAf(push, color->f[1]);
>>        PUSH_DATAf(push, color->f[2]);
>>        PUSH_DATAf(push, color->f[3]);
>> -      mode >> -         NVC0_3D_CLEAR_BUFFERS_R |
NVC0_3D_CLEAR_BUFFERS_G |
>> -         NVC0_3D_CLEAR_BUFFERS_B | NVC0_3D_CLEAR_BUFFERS_A;
>>     }
>>
>>     if (buffers & PIPE_CLEAR_DEPTH) {
>> @@ -444,12 +441,16 @@ nvc0_clear(struct pipe_context *pipe, unsigned
buffers,
>>        mode |= NVC0_3D_CLEAR_BUFFERS_S;
>>     }
>>
>> -   BEGIN_NVC0(push, NVC0_3D(CLEAR_BUFFERS), 1);
>> -   PUSH_DATA (push, mode);
>> -
>> -   for (i = 1; i < fb->nr_cbufs; i++) {
>> +   if (mode) {
>>        BEGIN_NVC0(push, NVC0_3D(CLEAR_BUFFERS), 1);
>> -      PUSH_DATA (push, (i << 6) | 0x3c);
>> +      PUSH_DATA (push, mode);
>> +   }
>> +
>> +   for (i = 0; i < fb->nr_cbufs; i++) {
>> +      if (buffers & (PIPE_CLEAR_COLOR0 << i)) {
>> +         BEGIN_NVC0(push, NVC0_3D(CLEAR_BUFFERS), 1);
>> +         PUSH_DATA (push, (i << 6) | 0x3c);
>> +      }
>>     }
>>  }
>>
>>
>

On 23/01/14 20:15, Ilia Mirkin wrote:
> On Thu, Jan 23, 2014 at 3:11 PM, Emil Velikov <emil.l.velikov at
gmail.com> wrote:
>> On 17/01/14 02:23, Ilia Mirkin wrote:
>>> Fixes fbo-drawbuffers-none glClearBuffer piglit test.
>>>
>>> Signed-off-by: Ilia Mirkin <imirkin at alum.mit.edu>
>>> ---
>>>
>>> Only tested on nv50, but implementations seem similar enough.
>>>
>>>  src/gallium/drivers/nouveau/nv50/nv50_surface.c | 17
+++++++++--------
>>>  src/gallium/drivers/nouveau/nvc0/nvc0_surface.c | 17
+++++++++--------
>>>  2 files changed, 18 insertions(+), 16 deletions(-)
>>>
>>> diff --git a/src/gallium/drivers/nouveau/nv50/nv50_surface.c
b/src/gallium/drivers/nouveau/nv50/nv50_surface.c
>>> index 358f57a..eb27429 100644
>>> --- a/src/gallium/drivers/nouveau/nv50/nv50_surface.c
>>> +++ b/src/gallium/drivers/nouveau/nv50/nv50_surface.c
>>> @@ -408,9 +408,6 @@ nv50_clear(struct pipe_context *pipe, unsigned
buffers,
>>>        PUSH_DATAf(push, color->f[1]);
>>>        PUSH_DATAf(push, color->f[2]);
>>>        PUSH_DATAf(push, color->f[3]);
>>> -      mode >>> -         NV50_3D_CLEAR_BUFFERS_R |
NV50_3D_CLEAR_BUFFERS_G |
>>> -         NV50_3D_CLEAR_BUFFERS_B | NV50_3D_CLEAR_BUFFERS_A;
>>>     }
>>>
>> I'm not sure why you've dropped the mode from above. I'm
guessing that
>> the initial assumption was that if there is a color buffer it must be
at
>> cbuf[0].
> 
> Because I cleared the cbufs separately in a for loop below (the 0x3c
> == the RGBA mask). The first RT may not have been there, and that
> seemed simpler than the thing that I have now (as evidenced by the
> code I have now which has more complex logic).
> 
With the original patch, you explicitly set clear_depth and
clear_stencil for the first buffer, regardless of what is being passed
by gallium. I'm not sure that was strictly correct.

-Emil
>>
>> The corrected version in your github branch looks alot better, it
>> handles the above case and does not overwrite the clear buffer on rt0.
> 
> Christoph was really keen on doing the common-case color/depth clear
> all in one clear buffers command, so yeah -- I redid it that way.
> Enjoy the layered version of that in a later commit.
> 
>>
>> That one is Reviewed-by: Emil Velikov <emil.l.velikov at
gmail.com>
>>>     if (buffers & PIPE_CLEAR_DEPTH) {
>>> @@ -425,12 +422,16 @@ nv50_clear(struct pipe_context *pipe,
unsigned buffers,
>>>        mode |= NV50_3D_CLEAR_BUFFERS_S;
>>>     }
>>>
>>> -   BEGIN_NV04(push, NV50_3D(CLEAR_BUFFERS), 1);
>>> -   PUSH_DATA (push, mode);
>>> -
>>> -   for (i = 1; i < fb->nr_cbufs; i++) {
>>> +   if (mode) {
>>>        BEGIN_NV04(push, NV50_3D(CLEAR_BUFFERS), 1);
>>> -      PUSH_DATA (push, (i << 6) | 0x3c);
>>> +      PUSH_DATA (push, mode);
>>> +   }
>>> +
>>> +   for (i = 0; i < fb->nr_cbufs; i++) {
>>> +      if (buffers & (PIPE_CLEAR_COLOR0 << i)) {
>>> +         BEGIN_NV04(push, NV50_3D(CLEAR_BUFFERS), 1);
>>> +         PUSH_DATA (push, (i << 6) | 0x3c);
>>> +      }
>>>     }
>>>  }
>>>
>>> diff --git a/src/gallium/drivers/nouveau/nvc0/nvc0_surface.c
b/src/gallium/drivers/nouveau/nvc0/nvc0_surface.c
>>> index 5375bd4..0c12bce 100644
>>> --- a/src/gallium/drivers/nouveau/nvc0/nvc0_surface.c
>>> +++ b/src/gallium/drivers/nouveau/nvc0/nvc0_surface.c
>>> @@ -427,9 +427,6 @@ nvc0_clear(struct pipe_context *pipe, unsigned
buffers,
>>>        PUSH_DATAf(push, color->f[1]);
>>>        PUSH_DATAf(push, color->f[2]);
>>>        PUSH_DATAf(push, color->f[3]);
>>> -      mode >>> -         NVC0_3D_CLEAR_BUFFERS_R |
NVC0_3D_CLEAR_BUFFERS_G |
>>> -         NVC0_3D_CLEAR_BUFFERS_B | NVC0_3D_CLEAR_BUFFERS_A;
>>>     }
>>>
>>>     if (buffers & PIPE_CLEAR_DEPTH) {
>>> @@ -444,12 +441,16 @@ nvc0_clear(struct pipe_context *pipe,
unsigned buffers,
>>>        mode |= NVC0_3D_CLEAR_BUFFERS_S;
>>>     }
>>>
>>> -   BEGIN_NVC0(push, NVC0_3D(CLEAR_BUFFERS), 1);
>>> -   PUSH_DATA (push, mode);
>>> -
>>> -   for (i = 1; i < fb->nr_cbufs; i++) {
>>> +   if (mode) {
>>>        BEGIN_NVC0(push, NVC0_3D(CLEAR_BUFFERS), 1);
>>> -      PUSH_DATA (push, (i << 6) | 0x3c);
>>> +      PUSH_DATA (push, mode);
>>> +   }
>>> +
>>> +   for (i = 0; i < fb->nr_cbufs; i++) {
>>> +      if (buffers & (PIPE_CLEAR_COLOR0 << i)) {
>>> +         BEGIN_NVC0(push, NVC0_3D(CLEAR_BUFFERS), 1);
>>> +         PUSH_DATA (push, (i << 6) | 0x3c);
>>> +      }
>>>     }
>>>  }
>>>
>>>
>>

On Thu, Jan 23, 2014 at 3:31 PM, Emil Velikov <emil.l.velikov at
gmail.com> wrote:
> On 23/01/14 20:15, Ilia Mirkin wrote:
>> On Thu, Jan 23, 2014 at 3:11 PM, Emil Velikov <emil.l.velikov at
gmail.com> wrote:
>>> On 17/01/14 02:23, Ilia Mirkin wrote:
>>>> Fixes fbo-drawbuffers-none glClearBuffer piglit test.
>>>>
>>>> Signed-off-by: Ilia Mirkin <imirkin at alum.mit.edu>
>>>> ---
>>>>
>>>> Only tested on nv50, but implementations seem similar enough.
>>>>
>>>>  src/gallium/drivers/nouveau/nv50/nv50_surface.c | 17
+++++++++--------
>>>>  src/gallium/drivers/nouveau/nvc0/nvc0_surface.c | 17
+++++++++--------
>>>>  2 files changed, 18 insertions(+), 16 deletions(-)
>>>>
>>>> diff --git a/src/gallium/drivers/nouveau/nv50/nv50_surface.c
b/src/gallium/drivers/nouveau/nv50/nv50_surface.c
>>>> index 358f57a..eb27429 100644
>>>> --- a/src/gallium/drivers/nouveau/nv50/nv50_surface.c
>>>> +++ b/src/gallium/drivers/nouveau/nv50/nv50_surface.c
>>>> @@ -408,9 +408,6 @@ nv50_clear(struct pipe_context *pipe,
unsigned buffers,
>>>>        PUSH_DATAf(push, color->f[1]);
>>>>        PUSH_DATAf(push, color->f[2]);
>>>>        PUSH_DATAf(push, color->f[3]);
>>>> -      mode >>>> -         NV50_3D_CLEAR_BUFFERS_R
| NV50_3D_CLEAR_BUFFERS_G |
>>>> -         NV50_3D_CLEAR_BUFFERS_B | NV50_3D_CLEAR_BUFFERS_A;
>>>>     }
>>>>
>>> I'm not sure why you've dropped the mode from above.
I'm guessing that
>>> the initial assumption was that if there is a color buffer it must
be at
>>> cbuf[0].
>>
>> Because I cleared the cbufs separately in a for loop below (the 0x3c
>> == the RGBA mask). The first RT may not have been there, and that
>> seemed simpler than the thing that I have now (as evidenced by the
>> code I have now which has more complex logic).
>>
> With the original patch, you explicitly set clear_depth and
> clear_stencil for the first buffer, regardless of what is being passed
> by gallium. I'm not sure that was strictly correct.
Only if buffers & PIPE_CLEAR_DEPTH/STENCIL were set... of which there
can only be one, unlike the color buffers, of which there can be many.
Well, whatever. I think everyone's happy with the new version, right?