Module `Nir.Node`Source

Nodes descriptions

A node is composed of

a unique id of type int
a node description of type descr

descr describes several operations. When an operation shares the same name as an ONNX operation, it follows the standard defined in the ONNX IR v8 and ONNX Opset v13 standards, described here: https://onnx.ai/onnx/operators/index.html.

Nodes only require their inputs: it is assumed that a node only returns one value.

Sourcetype ty =

| BFloat16
| Float16
| Float
| UInt8
| Int8
| Int32
| Int64

Source

val pp_ty : 
  Ppx_deriving_runtime.Format.formatter ->
  ty ->
  Ppx_deriving_runtime.unit

Sourceval show_ty : ty -> Ppx_deriving_runtime.string

Sourcetype descr =

| Constant of {
1. data : Gentensor.t;
}
(*
A constant tensor, used to store non-varying parameters during inference.
*)
| Add of {
1. input1 : t;
2. input2 : t;
}
| Sub of {
1. input1 : t;
2. input2 : t;
}
| Mul of {
1. input1 : t;
2. input2 : t;
}
| Div of {
1. input1 : t;
2. input2 : t;
}
| Matmul of {
1. input1 : t;
2. input2 : t;
}
| QLinearMatMul of {
1. inputA : t;
2. inputA_scale : t;
3. inputA_zero_point : t;
4. inputB : t;
5. inputB_scale : t;
6. inputB_zero_point : t;
7. y_scale : t;
8. y_zero_point : t;
}
| Gemm of {
1. inputA : t;
2. inputB : t;
3. inputC : t Base.option;
4. alpha : Base.float;
5. beta : Base.float;
6. transA : Base.int;
7. transB : Base.int;
}
| QGemm of {
1. inputA : t;
2. inputA_scale : t;
3. inputA_zero_point : t;
4. inputB : t;
5. inputB_scale : t;
6. inputB_zero_point : t;
7. inputC : t Base.option;
8. y_scale : t Base.option;
9. y_zero_point : t Base.option;
10. alpha : Base.float;
11. transA : Base.int;
12. transB : Base.int;
}
(*
Not an ONNX operator of the default domain. Documentation at:
- https://xadupre.github.io/draft/inference/operators/onnx_commicrosoft_QGemm.html
- https://github.com/microsoft/onnxruntime/blob/main/docs/ContribOperators.md#commicrosoftqgemm
*)
| LogSoftmax
| ReLu of {
1. input : t;
}
| Transpose of {
1. input : t;
  (*
  Called "data" in ONNX documentation: https://onnx.ai/onnx/operators/onnx__Transpose.html .
  *)
2. perm : Base.int Base.list;
}
| Squeeze of {
1. data : t;
2. axes : t Base.option;
}
| MaxPool
| Conv
| Reshape of {
1. input : t;
2. shape : t;
}
| Flatten of {
1. input : t;
2. axis : Base.int;
}
| Identity of {
1. input : t;
}
| Input of {
1. shape : Shape.t;
}
| RW_Linearized_ReLu
| Concat of {
1. inputs : t Base.list;
2. axis : Base.int;
}
| Gather of {
1. input : t;
2. indices : t;
3. axis : Base.int;
}
| ReduceSum of {
1. input : t;
2. axes : t Base.option;
3. keepdims : Base.int;
4. noop_with_empty_axes : Base.int;
}
| GatherND of {
1. data : t;
2. indices : t;
3. batch_dims : Base.int;
}
| RandomNormal of {
1. dtype : Base.int;
2. mean : Base.float;
3. scale : Base.float;
4. seed : Base.float;
5. shape : Base.int Base.array;
}
| Abs of {
1. input : t;
}
| Log of {
1. input : t;
}
| Sign of {
1. input : t;
}
| ArgMax of {
1. input : t;
2. axis : Base.int;
3. keepdims : Base.bool;
}
| Pow of {
1. input1 : t;
2. input2 : t;
}
| QuantizeLinear of {
1. x : t;
2. y_scale : t;
3. y_zero_point : t Base.option;
4. axis : Base.int;
}
| DequantizeLinear of {
1. x : t;
2. x_scale : t;
3. x_zero_point : t Base.option;
4. axis : Base.int;
}

Sourceand t = private {

id : Base.int;
descr : descr;
shape : Shape.t;
ty : ty;
(*
Describes the shape of the result of the node computation.
*)

}

Source

val pp_descr : 
  Ppx_deriving_runtime.Format.formatter ->
  descr ->
  Ppx_deriving_runtime.unit

Sourceval show_descr : descr -> Ppx_deriving_runtime.string

Source

val pp : 
  Ppx_deriving_runtime.Format.formatter ->
  t ->
  Ppx_deriving_runtime.unit

Sourceval show : t -> Ppx_deriving_runtime.string

Sourceval equal : t -> t -> Base.bool

include Base.Hashtbl.Key.S with type t := t

Sourceval compare : t -> t -> int

Sourceval sexp_of_t : t -> Sexplib0.Sexp.t

Sourceval hash : t -> int

Two ts that compare equal must have equal hashes for the hashtable to behave properly.

include Base.Comparator.S with type t := t

Sourcetype comparator_witness

Sourceval comparator : (t, comparator_witness) Base.Comparator.comparator

Sourceval create : descr -> t

create descr returns a value of type node with proper indexing and the shape according to the ONNX semantic.

Sourceval gather_int : ?encode:Base.bool -> t -> Base.int -> t

gather_int n i gathers the ith element of the node n.

Sourceval gather_ints : ?encode:Base.bool -> t -> Base.int Base.list -> t

gather_ints n l gathers the ith element of the node n for each element i of l.

Sourceval map : (t -> t) -> t -> t

map f n replace the direct inputs i of n by f i

Sourceval map_rec : (t -> t) -> t -> t

map_rec f n replace top-bottom the nodes i accessible from n by f i

Sourceval replace_input : (Base.unit -> t) -> t -> t

replace_input f n replace the input in n by f ()

Sourceval preds : t -> t Base.list

Direct predecessors of a t.

Sourceval iter_rec : (t -> Base.unit) -> t -> Base.unit

Iterate on the predecessors of a t and itself. Repect topological order.

Sourceval compute_shape : t -> Shape.t

Sourceval (+) : t -> t -> t

Sourceval (*) : t -> t -> t

Sourceval (-) : t -> t -> t

Sourceval mul_float : t -> Base.float -> t

Sourceval div_float : ?encode:Base.bool -> t -> Base.float -> t

Sourceval concat_0 : t Base.list -> t

Sourceval reshape : Shape.t -> t -> t

Sourceval sum_list : ?shp:Shape.t -> t Base.list -> t

sum_list shp ns is a node corresponding to the sum of the nodes in ns. If ns is empty, this returns a tensor of shape shp filled with 0s. By default, shp is a single float.

Source

val partial_dot_product : 
  ?shp:Shape.t ->
  t Base.array ->
  t Base.array ->
  Base.int ->
  Base.int ->
  t

partial_dot_product shp arr1 arr2 first last where arr1 = [\|n11, n12, ..., n1k1\|] and arr2 = [\|n21, n22, ..., n2k2\|] is a node corresponding to (n1first * n2first) + (n1first + 1 * n2first + 1) + ... + (n1last - 1 * n2last - 1) if this exists. It is assumed that arr1 and arr2 contain tensors with same shape. Edge cases include:

if last > length n1 or last > length n2, then fails
if last >= first, then returns a tensor where all values are initialized to 0. The shape of this tensor is determined using the following order:

if length arr1 <> 0 then use the shape of arr1.(0)
if length arr2 <> 0 then use the shape of arr2.(0)
if shp <> None, then use shp
otherwise, fails

Source

val transpose : 
  Base.int Base.list ->
  Base.int Base.array ->
  Base.int Base.array

transpose perm id is the position of the component at position id when the permutation perm is applied as the result of a tensor transposition. For instance, if id = [\|10; 20; 30\|] and \|perm = [2;0;1]\|, then transpose perm id will equal [\|30; 10; 20\|]. The empty permutation is represented by [] and is interpreted as [r-1; r-2; ...; 1; 0] (following <https://onnx.ai/onnx/operators/onnx__Transpose.html#transpose-23>); otherwise it is assumed that perm is a permutation of [0; 1; ...; r-1].

Source

val untranspose : 
  Base.int Base.list ->
  Base.int Base.array ->
  Base.int Base.array

untranspose is the reverse of transpose so that untranspose perm @@ transpose perm id equals id.

Sourceval flatten : Shape.t -> Base.int -> Base.int Base.array -> Base.int Base.array

flatten shp axis id computes the position of the component at position id when the flattening on axis is performed over shape shp. Following the definition of Flatten <https://onnx.ai/onnx/operators/onnx__Flatten.html#flatten-23>, axis can be negative in which case it computes from the end. For instance, if the shape is 10 * 10 * 10 * 10 * 10 (rank 5), axis = 3, and id = [\|1; 2; 3; 4; 5\|], then the result will be [\|123; 45\|] (computed as [\| 3 + (10 * ( 2 + 10 * 1)); 5 + 10 * 4) \|]).

Source

val unflatten : 
  Shape.t ->
  Base.int ->
  Base.int Base.array ->
  Base.int Base.array

unflatten is the reverse of flatten so that unflatten sh axis @@ flatten sh axis id equals id (for correct inputs id).

Sourceval encode_qgemm : descr -> descr

Encode QGemm operator in terms of DequantizeLinear, QuantizeLinear and Gemm operators.

Install

Dune Dependency

Authors

Maintainers

Sources

doc/caisar.nir/Nir/Node/index.html

Module `Nir.Node`Source

Nodes descriptions

package caisar

Install

Dune Dependency

Authors

Maintainers

Sources

doc/caisar.nir/Nir/Node/index.html

Module Nir.NodeSource

Nodes descriptions

Module `Nir.Node`Source