use std::collections::BTreeMap;
use computegraph::graph::{Graph, GraphBuilder};
use computegraph::types::ValueRef;
use super::{
all_primitive_ad_support, identity_matrix, one_like, primitive_ad_support, promote_dtype,
AdRuleSupport,
};
use crate::dim_expr::DimExpr;
use crate::input_key::TensorInputKey;
use crate::std_tensor_op::StdTensorOp;
use tenferro_core_ops::{all_primitive_descriptors, PrimitiveOpKind};
use tenferro_tensor::DType;
fn op_kind_name(op: &StdTensorOp) -> &'static str {
match op {
StdTensorOp::Constant { .. } => "Constant",
StdTensorOp::BroadcastInDim { .. } => "BroadcastInDim",
StdTensorOp::EmbedDiag { .. } => "EmbedDiag",
StdTensorOp::Exp => "Exp",
StdTensorOp::Log => "Log",
StdTensorOp::Sin => "Sin",
StdTensorOp::Cos => "Cos",
StdTensorOp::Tanh => "Tanh",
StdTensorOp::Sqrt => "Sqrt",
StdTensorOp::Rsqrt => "Rsqrt",
StdTensorOp::Expm1 => "Expm1",
StdTensorOp::Log1p => "Log1p",
_ => "Other",
}
}
fn op_kind_histogram(graph: &Graph<StdTensorOp>) -> BTreeMap<&'static str, usize> {
let mut counts = BTreeMap::new();
for op in graph.operations() {
*counts.entry(op_kind_name(&op.operation)).or_insert(0) += 1;
}
counts
}
fn assert_no_analytic_constant_shortcuts(graph: &Graph<StdTensorOp>) {
for op in graph.operations() {
assert!(
!matches!(
op.operation,
StdTensorOp::Exp
| StdTensorOp::Log
| StdTensorOp::Sin
| StdTensorOp::Cos
| StdTensorOp::Tanh
| StdTensorOp::Sqrt
| StdTensorOp::Rsqrt
| StdTensorOp::Expm1
| StdTensorOp::Log1p
),
"AD semantic constant helpers must not use analytic shortcuts; saw {:?}",
op.operation
);
}
}
#[test]
fn primitive_ad_support_manifest_covers_core_catalog_order() {
let manifest = all_primitive_ad_support();
assert_eq!(manifest.len(), PrimitiveOpKind::COUNT);
for descriptor in all_primitive_descriptors() {
let entry = primitive_ad_support(descriptor.kind);
assert_eq!(entry.kind, descriptor.kind);
assert_eq!(manifest[descriptor.kind.as_index()], *entry);
}
}
#[test]
fn primitive_ad_support_manifest_has_no_unsupported_entries() {
for entry in all_primitive_ad_support() {
assert_ne!(entry.linearize, AdRuleSupport::Unsupported);
assert_ne!(entry.transpose, AdRuleSupport::Unsupported);
}
}
#[test]
fn primitive_ad_support_manifest_marks_known_non_differentiable_ops() {
for kind in [
PrimitiveOpKind::Compare,
PrimitiveOpKind::ShapeOf,
PrimitiveOpKind::Constant,
] {
let entry = primitive_ad_support(kind);
assert_eq!(entry.linearize, AdRuleSupport::NonDifferentiable);
assert_eq!(entry.transpose, AdRuleSupport::NonDifferentiable);
}
}
#[test]
fn promote_dtype_covers_supported_pairs_without_runtime_unreachable() {
let source = include_str!("../support.rs");
assert!(
!source.contains("promote_dtype: unhandled pair"),
"dtype promotion should use exhaustive DType matching, not a runtime unreachable"
);
assert_eq!(promote_dtype(DType::Bool, DType::F64), DType::F64);
assert_eq!(promote_dtype(DType::I32, DType::I64), DType::I64);
assert_eq!(promote_dtype(DType::I64, DType::F32), DType::F64);
assert_eq!(promote_dtype(DType::F32, DType::C32), DType::C32);
assert_eq!(promote_dtype(DType::F64, DType::C32), DType::C64);
assert_eq!(promote_dtype(DType::C32, DType::C64), DType::C64);
}
#[test]
fn promote_dtype_delegates_to_canonical_tensor_validation_rules() {
let source = include_str!("../support.rs");
assert!(
source.contains("tenferro_tensor::validate::promote_dtype"),
"AD support must delegate to the canonical tensor dtype promotion lattice"
);
let dtypes = [
DType::Bool,
DType::I32,
DType::I64,
DType::F32,
DType::F64,
DType::C32,
DType::C64,
];
for lhs in dtypes {
for rhs in dtypes {
assert_eq!(
promote_dtype(lhs, rhs),
tenferro_tensor::validate::promote_dtype(lhs, rhs),
"promotion mismatch for {lhs:?}, {rhs:?}"
);
}
}
}
#[test]
fn one_like_helper_emits_semantic_constant_not_analytic_shortcut() {
let mut builder = GraphBuilder::<StdTensorOp>::new();
let anchor = builder.add_input(TensorInputKey::User { id: 901 });
let one = one_like(&mut builder, DType::F64, ValueRef::Local(anchor), 2);
let graph = builder.build();
assert!(graph.values()[one].producer.is_some());
assert_no_analytic_constant_shortcuts(&graph);
assert_eq!(
op_kind_histogram(&graph),
BTreeMap::from([("BroadcastInDim", 1), ("Constant", 1)])
);
}
#[test]
fn identity_matrix_helper_emits_semantic_constant_and_remaps_shape_source() {
let mut builder = GraphBuilder::<StdTensorOp>::new();
let anchor = builder.add_input(TensorInputKey::User { id: 900 });
let identity = identity_matrix(
&mut builder,
DType::F64,
3,
&[DimExpr::InputDim {
input_idx: 0,
axis: 2,
}],
ValueRef::Local(anchor),
0,
);
let graph = builder.build();
assert_no_analytic_constant_shortcuts(&graph);
assert_eq!(
op_kind_histogram(&graph),
BTreeMap::from([("BroadcastInDim", 1), ("Constant", 1), ("EmbedDiag", 1)])
);
let (embed_diag_id, _) = graph.values()[identity]
.producer
.expect("identity output must be produced by embed diag");
assert!(matches!(
graph.operations()[embed_diag_id].operation,
StdTensorOp::EmbedDiag { .. }
));
let broadcast_id = match graph.operations()[embed_diag_id].inputs[0] {
ValueRef::Local(id) => id,
ref other => panic!("embed diag should consume broadcast output, got {other:?}"),
};
let (broadcast_op_id, _) = graph.values()[broadcast_id]
.producer
.expect("identity diagonal must be produced by broadcast");
let broadcast = &graph.operations()[broadcast_op_id];
assert_eq!(
broadcast.operation,
StdTensorOp::BroadcastInDim {
shape: vec![
DimExpr::Const(3),
DimExpr::InputDim {
input_idx: 1,
axis: 2,
},
],
dims: vec![],
}
);
assert_eq!(broadcast.inputs[1], ValueRef::Local(anchor));
}
#[test]
#[should_panic(
expected = "identity_matrix shape expressions must reference only the provided shape_source_idx"
)]
fn identity_matrix_helper_rejects_unbound_shape_sources() {
let mut builder = GraphBuilder::<StdTensorOp>::new();
let anchor = builder.add_input(TensorInputKey::User { id: 902 });
let _ = identity_matrix(
&mut builder,
DType::F64,
3,
&[DimExpr::InputDim {
input_idx: 1,
axis: 0,
}],
ValueRef::Local(anchor),
0,
);
}
fn preset_scalar_cases() -> Vec<(DType, Vec<u8>, Vec<u8>)> {
fn c32(re: f32, im: f32) -> Vec<u8> {
let mut bytes = re.to_le_bytes().to_vec();
bytes.extend_from_slice(&im.to_le_bytes());
bytes
}
fn c64(re: f64, im: f64) -> Vec<u8> {
let mut bytes = re.to_le_bytes().to_vec();
bytes.extend_from_slice(&im.to_le_bytes());
bytes
}
vec![
(
DType::F32,
0.0_f32.to_le_bytes().to_vec(),
1.0_f32.to_le_bytes().to_vec(),
),
(
DType::F64,
0.0_f64.to_le_bytes().to_vec(),
1.0_f64.to_le_bytes().to_vec(),
),
(
DType::I32,
0_i32.to_le_bytes().to_vec(),
1_i32.to_le_bytes().to_vec(),
),
(
DType::I64,
0_i64.to_le_bytes().to_vec(),
1_i64.to_le_bytes().to_vec(),
),
(DType::Bool, vec![0], vec![1]),
(DType::C32, c32(0.0, 0.0), c32(1.0, 0.0)),
(DType::C64, c64(0.0, 0.0), c64(1.0, 0.0)),
]
}
fn sole_constant_bytes(
graph: &Graph<StdTensorOp>,
value: computegraph::types::LocalValueId,
) -> (DType, Vec<u8>) {
let (op_id, _) = graph.values()[value]
.producer
.expect("scalar seed must be produced by a constant");
match &graph.operations()[op_id].operation {
StdTensorOp::Constant { dtype, bytes } => (*dtype, bytes.clone()),
other => panic!("expected a Constant, saw {other:?}"),
}
}
#[test]
fn zero_like_emits_exact_zero_bytes_for_every_preset_dtype() {
for (idx, (dtype, zero, _)) in preset_scalar_cases().into_iter().enumerate() {
let mut builder = GraphBuilder::<StdTensorOp>::new();
let anchor = builder.add_input(TensorInputKey::User {
id: 920 + idx as u64,
});
let value = super::zero_like(&mut builder, dtype, ValueRef::Local(anchor), 0);
let graph = builder.build();
assert_eq!(
op_kind_histogram(&graph),
BTreeMap::from([("Constant", 1)]),
"a rank-0 seed for {dtype:?} must be one constant and nothing else"
);
assert_eq!(sole_constant_bytes(&graph, value), (dtype, zero));
}
}
#[test]
fn one_like_emits_exact_one_bytes_for_every_preset_dtype() {
for (idx, (dtype, _, one)) in preset_scalar_cases().into_iter().enumerate() {
let mut builder = GraphBuilder::<StdTensorOp>::new();
let anchor = builder.add_input(TensorInputKey::User {
id: 930 + idx as u64,
});
let value = super::one_like(&mut builder, dtype, ValueRef::Local(anchor), 0);
let graph = builder.build();
assert_eq!(
op_kind_histogram(&graph),
BTreeMap::from([("Constant", 1)]),
"a rank-0 seed for {dtype:?} must be one constant and nothing else"
);
assert_eq!(sole_constant_bytes(&graph, value), (dtype, one));
}
}
#[test]
fn scalar_seeds_broadcast_once_per_anchor_axis() {
for (idx, (dtype, zero, one)) in preset_scalar_cases().into_iter().enumerate() {
for (label, expected_bytes) in [("zero", zero), ("one", one)] {
let mut builder = GraphBuilder::<StdTensorOp>::new();
let anchor = builder.add_input(TensorInputKey::User {
id: 940 + idx as u64,
});
let value = if label == "zero" {
super::zero_like(&mut builder, dtype, ValueRef::Local(anchor), 3)
} else {
super::one_like(&mut builder, dtype, ValueRef::Local(anchor), 3)
};
let graph = builder.build();
assert_eq!(
op_kind_histogram(&graph),
BTreeMap::from([("BroadcastInDim", 1), ("Constant", 1)]),
"{label} seed on a rank-3 anchor must lift one scalar once"
);
let (broadcast_id, _) = graph.values()[value]
.producer
.expect("non-scalar seed must be produced by a broadcast");
let broadcast = &graph.operations()[broadcast_id];
assert_eq!(
broadcast.operation,
StdTensorOp::BroadcastInDim {
shape: (0..3)
.map(|axis| DimExpr::InputDim { input_idx: 1, axis })
.collect(),
dims: vec![],
}
);
assert_eq!(broadcast.inputs[1], ValueRef::Local(anchor));
let scalar = match broadcast.inputs[0] {
ValueRef::Local(id) => id,
ref other => panic!("broadcast must consume a local scalar, got {other:?}"),
};
assert_eq!(sole_constant_bytes(&graph, scalar), (dtype, expected_bytes));
}
}
}