"expecting function with parameter domain" with "isl_aff_unbind_params_insert_domain"

[Pei Mu]

Hi,

I got some problems when coding to get the schedule tree about the following code:

```

  for (int iter0 = 0; iter0 < 64; iter0++) {

    for (int iter1 = 0; iter1 < 8; iter1++) {

S:    Var2[iter0*8+iter1] = Var0[iter0*8+iter1] + Var1[iter0*8+iter1];

    }   

  }

```

And I get the schedule tree by ppcg, like:

```

domain: "{ S[iter0, iter1] : 0 <= iter0 <= 63 and 0 <= iter1 <= 7 }"

child:

  schedule: "L_0[{ S[iter0, iter1] -> [(iter0)] }]"

  permutable: 1

  coincident: [ 1 ]

  child:

    schedule: "L_1[{ S[iter0, iter1] -> [(iter1)] }]"

    permutable: 1

    coincident: [ 1 ]

```

It's cool.

Now I want to implement my own schedule tree builder. but when I try to get the memory access relation, I met the error report: "expecting function with parameter domain" with "isl_aff_unbind_params_insert_domain"

Actually, I have no idea what it means.

I wish I could describe my code clearly:

1. Get the domain schedule: 

```domain: "{ S[iter0, iter1] : 0 <= iter0 <= 63 and 0 <= iter1 <= 7 }"```

2. Get the memory access by adding "Var2" to the domain space by the "isl_space_add_named_tuple_id_ui", and named it "tmp_space":

```

[iter0, iter1] -> { Var2[i0, i1] } // I guess this might need to be correct to [iter0, iter1] -> { Var2[iter0, iter1] }

```

And the isl_multi_id from such tmp_space, by "isl_multi_id_from_id_list", and named it "coordinate":

```

[iter0, iter1] -> { Var2[iter0, iter1] }

```

3. Get the first aff param on such tmp_space, by "isl_aff_param_on_domain_space_id", and named it "domain_aff_bounds":

```

[iter0, iter1] -> { Var2[i0, i1] -> [(iter0)] }

```

Then I try to write:

```

isl_aff_unbind_params_insert_domain(isl_aff_copy(domain_aff_bounds), coordinate)

```

And the error "expecting function with parameter domain" occurred.

Thank you very much for your patient, I appreciate that a lot.

The isl version is 0.24, and the cpp source code is attached. I would appreciate it a lot if you could reply!

Pei

[Sven Verdoolaege]

On Mon, Aug 16, 2021 at 06:00:06AM -0700, Pei Mu wrote:
> Hi,
> I got some problems when coding to get the schedule tree about the
> following code:
> ```
> for (int iter0 = 0; iter0 < 64; iter0++) {
> for (int iter1 = 0; iter1 < 8; iter1++) {
> S: Var2[iter0*8+iter1] = Var0[iter0*8+iter1] + Var1[iter0*8+iter1];
> }
> }
> ```
>

> Now I want to implement my own schedule tree builder. but when I try to get
> the memory access relation, I met the error report: "expecting function
> with parameter domain" with "isl_aff_unbind_params_insert_domain"
> Actually, I have no idea what it means.
>
> I wish I could describe my code clearly:
> 1. Get the domain schedule:
> ```domain: "{ S[iter0, iter1] : 0 <= iter0 <= 63 and 0 <= iter1 <= 7 }"```

I don't understand why you bring up a schedule here.
Apparently you don't have a schedule yet and you are trying
to construct an access relation.
I'll assume this is meant to be the domain of the access relation.

> 2. Get the memory access by adding "Var2" to the domain space by the
> "isl_space_add_named_tuple_id_ui", and named it "tmp_space":
> ```
> [iter0, iter1] -> { Var2[i0, i1] } // I guess this might need to be correct
> to [iter0, iter1] -> { Var2[iter0, iter1] }
> ```

Are you trying to construct `[iter0, iter1] -> { Var2[iter0*8+iter1] }`?
Then clearly, this is the wrong space.

> And the isl_multi_id from such tmp_space, by "isl_multi_id_from_id_list",
> and named it "coordinate":
> ```
> [iter0, iter1] -> { Var2[iter0, iter1] }
> ```
> 3. Get the first aff param on such tmp_space, by
> "isl_aff_param_on_domain_space_id", and named it "domain_aff_bounds":
> ```
> [iter0, iter1] -> { Var2[i0, i1] -> [(iter0)] }
> ```

What is this meant to represent? That is, why did you construct this?

> Then I try to write:
> ```
> isl_aff_unbind_params_insert_domain(isl_aff_copy(domain_aff_bounds),
> coordinate)
> ```

The function you constructed above already has a domain,
so you can't insert one.
isl_aff_unbind_params_insert_domain can only be used on a function
that does not have a domain tuple.
What are you trying to do here?

Btw, you may want to use the C++ interface.
In particular, if you use the templated C++ interface,
then the equivalent of the code above wouldn't even compile.

skimo

[Pei Mu]

Hi skimo,

Thanks a lot for your reply! 

> I don't understand why you bring up a schedule here. 
> Apparently you don't have a schedule yet and you are trying 
> to construct an access relation. 
> I'll assume this is meant to be the domain of the access relation. 

Yes, you are right. It's the domain of the access relation.

> > for (int iter0 = 0; iter0 < 64; iter0++) { 
> > for (int iter1 = 0; iter1 < 8; iter1++) { 
> > S: Var2[iter0*8+iter1] = Var0[iter0*8+iter1] + Var1[iter0*8+iter1]; 
> > } 
> > } 

> Are you trying to construct `[iter0, iter1] -> { Var2[iter0*8+iter1] }`? 
> Then clearly, this is the wrong space. 

Sorry for the error input C code, which should be:

```S: Var2[iter0][iter1] = Var0[iter0][iter1] + Var1[iter0][iter1]; ```

> > 3. Get the first aff param on such tmp_space, by 
> > "isl_aff_param_on_domain_space_id", and named it "domain_aff_bounds": 
> > ``` 
> > [iter0, iter1] -> { Var2[i0, i1] -> [(iter0)] } 
> > ``` 
>
> What is this meant to represent? That is, why did you construct this? 

I just learned from the AKG(Auto Kernel Gerenation) code and I not sure about what it actually mean.

I guess this code want to get the loop iterator info - "iter0". I will go to the AKG community(if exist) to confirm.

> The function you constructed above already has a domain, 

> so you can't insert one. 
> isl_aff_unbind_params_insert_domain can only be used on a function 
> that does not have a domain tuple. 

I've now understod, thank you.

> Btw, you may want to use the C++ interface. 

> In particular, if you use the templated C++ interface, 
> then the equivalent of the code above wouldn't even compile. 

Yes, I'm writting with the c++ interface.

Actually, I just want to make a toy demo with isl, which includes:

1. extracting schedule tree from the AST(I've finished my lexcer and parser, and get an AST);

2. instance access relation analysis and dependency analysis;

3. schedule transformation.

I chos the AKG at the first glance. PPCG and TC seem a bit complex. But it looks difficult either to learn from the AKG.

May be I should start with iscc or  barvinok?

Thanks again,

Pei

[Sven Verdoolaege]

On Tue, Aug 17, 2021 at 08:19:33PM -0700, Pei Mu wrote:
> Actually, I just want to make a toy demo with isl, which includes:
> 1. extracting schedule tree from the AST(I've finished my lexcer and
> parser, and get an AST);
> 2. instance access relation analysis and dependency analysis;
> 3. schedule transformation.
>
> I chos the AKG at the first glance. PPCG and TC seem a bit complex. But it
> looks difficult either to learn from the AKG.

The schedule tree and access relation construction in TC
are a lot simpler than in pet, partly because the input
is simpler language is simpler to analyze.
I'm not familiar with AKG.

> May be I should start with iscc or barvinok?

I guess that depends on what you want to do with them,
but iscc doesn't expose any low-level construction funtions
(since it's an interactive environment where set/relations
are assumed to be written be hand) or schedule tree functionality.
The pyhton interface may be more useful for experimenting
with these things.

skimo

[Pei Mu]

Awesome! Thanks for your help! I will check out TC and isl python interface.

Pei