[Mesa-dev] [PATCH 2/2] nir: add lowering for gl_HelperInvocation
Rob Clark
robdclark at gmail.com
Tue Jun 12 17:49:33 UTC 2018
On Mon, Jun 11, 2018 at 10:59 PM, Roland Scheidegger <sroland at vmware.com> wrote:
> Am 12.06.2018 um 01:17 schrieb Rob Clark:
>> On Mon, Jun 11, 2018 at 6:59 PM, Roland Scheidegger <sroland at vmware.com> wrote:
>>> Am 12.06.2018 um 00:32 schrieb Jason Ekstrand:
>>>> On Wed, Jun 6, 2018 at 7:43 AM, Rob Clark <robdclark at gmail.com
>>>> <mailto:robdclark at gmail.com>> wrote:
>>>>
>>>> Signed-off-by: Rob Clark <robdclark at gmail.com
>>>> <mailto:robdclark at gmail.com>>
>>>> ---
>>>> I can't say for sure that this will work on all drivers, but it is
>>>> what the blob driver does, and it seems to make deqp happy. I could
>>>> move this to it's own pass inside ir3, but that seemed like overkill
>>>>
>>>> src/compiler/nir/nir.h | 10 ++++++++++
>>>> src/compiler/nir/nir_lower_system_values.c | 17 +++++++++++++++++
>>>> src/gallium/drivers/freedreno/ir3/ir3_nir.c | 1 +
>>>> 3 files changed, 28 insertions(+)
>>>>
>>>> diff --git a/src/compiler/nir/nir.h b/src/compiler/nir/nir.h
>>>> index 073ab4e82ea..de3d55d83af 100644
>>>> --- a/src/compiler/nir/nir.h
>>>> +++ b/src/compiler/nir/nir.h
>>>> @@ -1963,6 +1963,16 @@ typedef struct nir_shader_compiler_options {
>>>> */
>>>> bool lower_base_vertex;
>>>>
>>>> + /**
>>>> + * If enabled, gl_HelperInvocation will be lowered as:
>>>> + *
>>>> + * !((1 << gl_SampleID) & gl_SampleMaskIN[0]))
>>>>
>>>>
>>>> This only works for multi-sampling. What about the single-sampled case?
>>> It doesn't make sense to me for msaa neither.
>>> gl_SampleID forces per-sample execution, which clearly isn't what you
>>> want.
>>
>> open to suggestions about how to try and force blob to do something
>> different.. I tried a few things and the blob popped out shader code
>> that worked the same way in all cases, but ofc I could have missed
>> something.
>>
>>> Plus, gl_SampleMaskIN is specified to only contain bits for the
>>> current shader invocation, so for msaa with forced per-sample execution
>>> that would only contain the single bit corresponding to gl_SampleID
>>> anyway (that is my interpretation at least - I know hw sample mask
>>> inputs will probably always contains all the bits from rasterization,
>>> regardless of gl_SampleID), so the "1 << gl_SampleID &" part will do
>>> nothing at all. So I think that part is more about lowering the hw
>>> rasterization sample mask to gl_SampleMaskIN rather than lowering to
>>> gl_HelperInvocation.
>>> But yes, !gl_SampleMaskIN should give gl_HelperInvocation - I think all
>>> hw can give you raster sample mask even without msaa but I'm not
>>> entirely sure if it's guaranteed to work in GL with single sampling
>>> (similar for sample id, which should just be stuck at 0). But using the
>>> gl names here looks very, very fishy to me here.
>>
>> I *guess* this is relying on behaviour that gl does not guarantee, but
>> also doesn't prohibit, but the hw happens work that way and it works
>> out..
>>
>> Maybe the answer is to simply better document the assumptions that
>> this lowering depends on?
> Well I think if you have a nir shader which uses sample id but is
> supposed to not be run per-sample that is quite confusing (I'd just
> assume that adrenos will use sample id 0 just like in single sampled
> case). Hopefully noone will deduce shader frequency from the nir shader...
> I'm definitely no NIR expert but there's quite a lot of hacky
> assumptions in there:
> - sample id in nir doesn't cause per-sample execution (and needs to be 0
> in case of msaa but no per-sample execution)
> - sampleMaskIn is available in single-sampled
> - sampleMaskIn doesn't actually have gl semantics but hw ones (so simply
> the mask from rasterization, not taking into account the sample id
> already if there's per-sample execution (for legitimate reasons),
> although as mentioned I think there's still some debate about this and
> not all drivers may actually agree what the semantics of it are in case
> of per-sample execution)
>
> As such I'm not sure it really makes sense to do that as a generic
> lowering pass? Albeit you're quite right that it might work for a lot of
> hw... But some comments would at the very least imho definitely be required.
>
so I reworded the comment about the enable flag a bit:
/**
* If enabled, gl_HelperInvocation will be lowered as:
*
* !((1 << sample_id) & sample_mask_in))
*
* This depends on some possibly hw implementation details, which may
* not be true for all hw. In particular that the FS is only executed
* for covered samples or for helper invocations. So, do not blindly
* enable this option.
*/
bool lower_helper_invocation;
But I guess I probably do need to introduce a new intrinsic to access
sample_id without necessarily triggering per-sample mode.. hmm..
BR,
-R
> Roland
>
>
>
>>
>> BR,
>> -R
>>
>>>
>>> Roland
>>>
>>>
>>>
>>>
>>>>
>>>> --Jason
>>>>
>>>>
>>>>
>>>> + *
>>>> + * TODO any hw w/ more than 32 samples? For them (if they
>>>> + * used this option), a bit more math would be involved.
>>>> + */
>>>> + bool lower_helper_invocation;
>>>> +
>>>> bool lower_cs_local_index_from_id;
>>>>
>>>> bool lower_device_index_to_zero;
>>>> diff --git a/src/compiler/nir/nir_lower_system_values.c
>>>> b/src/compiler/nir/nir_lower_system_values.c
>>>> index 487da042620..6668cbb5dcd 100644
>>>> --- a/src/compiler/nir/nir_lower_system_values.c
>>>> +++ b/src/compiler/nir/nir_lower_system_values.c
>>>> @@ -136,6 +136,23 @@ convert_block(nir_block *block, nir_builder *b)
>>>> nir_load_first_vertex(b));
>>>> break;
>>>>
>>>> + case SYSTEM_VALUE_HELPER_INVOCATION:
>>>> + if (b->shader->options->lower_helper_invocation) {
>>>> + nir_ssa_def *tmp;
>>>> +
>>>> + tmp = nir_ushr(b,
>>>> + nir_imm_int(b, 1),
>>>> + nir_load_sample_id(b));
>>>> +
>>>> + tmp = nir_iand(b,
>>>> + nir_load_sample_mask_in(b),
>>>> + tmp);
>>>> +
>>>> + sysval = nir_inot(b, nir_i2b(b, tmp));
>>>> + }
>>>> +
>>>> + break;
>>>> +
>>>> case SYSTEM_VALUE_INSTANCE_INDEX:
>>>> sysval = nir_iadd(b,
>>>> nir_load_instance_id(b),
>>>> diff --git a/src/gallium/drivers/freedreno/ir3/ir3_nir.c
>>>> b/src/gallium/drivers/freedreno/ir3/ir3_nir.c
>>>> index cd1f9c526f2..341d990b269 100644
>>>> --- a/src/gallium/drivers/freedreno/ir3/ir3_nir.c
>>>> +++ b/src/gallium/drivers/freedreno/ir3/ir3_nir.c
>>>> @@ -51,6 +51,7 @@ static const nir_shader_compiler_options options = {
>>>> .lower_extract_byte = true,
>>>> .lower_extract_word = true,
>>>> .lower_all_io_to_temps = true,
>>>> + .lower_helper_invocation = true,
>>>> };
>>>>
>>>> struct nir_shader *
>>>> --
>>>> 2.17.0
>>>>
>>>> _______________________________________________
>>>> mesa-dev mailing list
>>>> mesa-dev at lists.freedesktop.org <mailto:mesa-dev at lists.freedesktop.org>
>>>> https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Fmesa-dev&data=02%7C01%7Csroland%40vmware.com%7C3de626b5c92844260bfb08d5cff17c34%7Cb39138ca3cee4b4aa4d6cd83d9dd62f0%7C1%7C0%7C636643558433673773&sdata=rA8s3IuEmEOR6NKdxwa5eTIwAOOhbAQWNAxW6FmiUOU%3D&reserved=0
>>>> <https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Fmesa-dev&data=02%7C01%7Csroland%40vmware.com%7Cc50ea7f234a34e635ac308d5cfeb432a%7Cb39138ca3cee4b4aa4d6cd83d9dd62f0%7C1%7C0%7C636643531719133589&sdata=k8EjcVGoBZyrmPfu5s5yUFZ7OerpIrUi3W3bK7qVz5o%3D&reserved=0>
>>>>
>>>>
>>>>
>>>>
>>>> _______________________________________________
>>>> mesa-dev mailing list
>>>> mesa-dev at lists.freedesktop.org
>>>> https://na01.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Fmesa-dev&data=02%7C01%7Csroland%40vmware.com%7Cc50ea7f234a34e635ac308d5cfeb432a%7Cb39138ca3cee4b4aa4d6cd83d9dd62f0%7C1%7C0%7C636643531719133589&sdata=k8EjcVGoBZyrmPfu5s5yUFZ7OerpIrUi3W3bK7qVz5o%3D&reserved=0
>>>>
>>>
>
More information about the mesa-dev
mailing list