binius_math/
mle_adapters.rs

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
// Copyright 2023-2024 Irreducible Inc.

use super::{Error, MultilinearExtension, MultilinearPoly, MultilinearQueryRef};
use binius_field::{
	packed::{
		get_packed_slice, get_packed_slice_unchecked, set_packed_slice, set_packed_slice_unchecked,
	},
	ExtensionField, Field, PackedField, RepackedExtension,
};
use binius_utils::bail;
use rayon::prelude::*;
use std::{fmt::Debug, marker::PhantomData, ops::Deref, sync::Arc};

/// An adapter for [`MultilinearExtension`] that implements [`MultilinearPoly`] over a packed
/// extension field.
///
/// This struct implements `MultilinearPoly` for an extension field of the base field that the
/// multilinear extension is defined over.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct MLEEmbeddingAdapter<P, PE, Data = Vec<P>>(
	MultilinearExtension<P, Data>,
	PhantomData<PE>,
)
where
	P: PackedField,
	PE: PackedField,
	PE::Scalar: ExtensionField<P::Scalar>,
	Data: Deref<Target = [P]>;

impl<'a, P, PE, Data> MLEEmbeddingAdapter<P, PE, Data>
where
	P: PackedField,
	PE: PackedField + RepackedExtension<P>,
	PE::Scalar: ExtensionField<P::Scalar>,
	Data: Deref<Target = [P]> + Send + Sync + Debug + 'a,
{
	pub fn upcast_arc_dyn(self) -> Arc<dyn MultilinearPoly<PE> + Send + Sync + 'a> {
		Arc::new(self)
	}
}

impl<P, PE, Data> From<MultilinearExtension<P, Data>> for MLEEmbeddingAdapter<P, PE, Data>
where
	P: PackedField,
	PE: PackedField,
	PE::Scalar: ExtensionField<P::Scalar>,
	Data: Deref<Target = [P]>,
{
	fn from(inner: MultilinearExtension<P, Data>) -> Self {
		Self(inner, PhantomData)
	}
}

impl<P, PE, Data> AsRef<MultilinearExtension<P, Data>> for MLEEmbeddingAdapter<P, PE, Data>
where
	P: PackedField,
	PE: PackedField,
	PE::Scalar: ExtensionField<P::Scalar>,
	Data: Deref<Target = [P]>,
{
	fn as_ref(&self) -> &MultilinearExtension<P, Data> {
		&self.0
	}
}

impl<P, PE, Data> MultilinearPoly<PE> for MLEEmbeddingAdapter<P, PE, Data>
where
	P: PackedField + Debug,
	PE: PackedField + RepackedExtension<P>,
	PE::Scalar: ExtensionField<P::Scalar>,
	Data: Deref<Target = [P]> + Send + Sync + Debug,
{
	fn n_vars(&self) -> usize {
		self.0.n_vars()
	}

	fn log_extension_degree(&self) -> usize {
		PE::Scalar::LOG_DEGREE
	}

	fn evaluate_on_hypercube(&self, index: usize) -> Result<PE::Scalar, Error> {
		let eval = self.0.evaluate_on_hypercube(index)?;
		Ok(eval.into())
	}

	fn evaluate_on_hypercube_and_scale(
		&self,
		index: usize,
		scalar: PE::Scalar,
	) -> Result<PE::Scalar, Error> {
		let eval = self.0.evaluate_on_hypercube(index)?;
		Ok(scalar * eval)
	}

	fn evaluate(&self, query: MultilinearQueryRef<PE>) -> Result<PE::Scalar, Error> {
		self.0.evaluate(query)
	}

	fn evaluate_partial_low(
		&self,
		query: MultilinearQueryRef<PE>,
	) -> Result<MultilinearExtension<PE>, Error> {
		self.0.evaluate_partial_low(query)
	}

	fn evaluate_partial_high(
		&self,
		query: MultilinearQueryRef<PE>,
	) -> Result<MultilinearExtension<PE>, Error> {
		self.0.evaluate_partial_high(query)
	}

	fn subcube_inner_products(
		&self,
		query: MultilinearQueryRef<PE>,
		subcube_vars: usize,
		subcube_index: usize,
		inner_products: &mut [PE],
	) -> Result<(), Error> {
		let query_n_vars = query.n_vars();
		if query_n_vars + subcube_vars > self.n_vars() {
			bail!(Error::ArgumentRangeError {
				arg: "query.n_vars() + subcube_vars".into(),
				range: 0..self.n_vars(),
			});
		}

		let max_index = 1 << (self.n_vars() - query_n_vars - subcube_vars);
		if subcube_index >= max_index {
			bail!(Error::ArgumentRangeError {
				arg: "subcube_index".into(),
				range: 0..max_index,
			});
		}

		let correct_len = 1 << subcube_vars.saturating_sub(PE::LOG_WIDTH);
		if inner_products.len() != correct_len {
			bail!(Error::ArgumentRangeError {
				arg: "evals.len()".to_string(),
				range: correct_len..correct_len + 1,
			});
		}

		// REVIEW: not spending effort to optimize this as the future of switchover
		//         is somewhat unclear in light of univariate skip
		let subcube_start = subcube_index << (query_n_vars + subcube_vars);
		for scalar_index in 0..1 << subcube_vars {
			let evals_start = subcube_start + (scalar_index << query_n_vars);
			let mut inner_product = PE::Scalar::ZERO;
			for i in 0..1 << query_n_vars {
				inner_product += get_packed_slice(query.expansion(), i)
					* get_packed_slice(self.0.evals(), evals_start + i);
			}

			set_packed_slice(inner_products, scalar_index, inner_product);
		}

		Ok(())
	}

	fn subcube_evals(
		&self,
		subcube_vars: usize,
		subcube_index: usize,
		log_embedding_degree: usize,
		evals: &mut [PE],
	) -> Result<(), Error> {
		let log_extension_degree = PE::Scalar::LOG_DEGREE;

		if subcube_vars > self.n_vars() {
			bail!(Error::ArgumentRangeError {
				arg: "subcube_vars".to_string(),
				range: 0..self.n_vars() + 1,
			});
		}

		// Check that chosen embedding subfield is large enough.
		// We also use a stack allocated array of bases, which imposes
		// a maximum tower height restriction.
		const MAX_TOWER_HEIGHT: usize = 7;
		if log_embedding_degree > log_extension_degree.min(MAX_TOWER_HEIGHT) {
			bail!(Error::LogEmbeddingDegreeTooLarge {
				log_embedding_degree
			});
		}

		let correct_len = 1 << subcube_vars.saturating_sub(log_embedding_degree + PE::LOG_WIDTH);
		if evals.len() != correct_len {
			bail!(Error::ArgumentRangeError {
				arg: "evals.len()".to_string(),
				range: correct_len..correct_len + 1,
			});
		}

		let max_index = 1 << (self.n_vars() - subcube_vars);
		if subcube_index >= max_index {
			bail!(Error::ArgumentRangeError {
				arg: "subcube_index".to_string(),
				range: 0..max_index,
			});
		}

		let subcube_start = subcube_index << subcube_vars;

		if log_embedding_degree == 0 {
			// One-to-one embedding can bypass the extension field construction overhead.
			for i in 0..1 << subcube_vars {
				// Safety: subcube_index < max_index check
				let scalar =
					unsafe { get_packed_slice_unchecked(self.0.evals(), subcube_start + i) };

				let extension_scalar = scalar.into();

				// Safety: i < 1 << min(0, subcube_vars) <= correct_len * PE::WIDTH
				unsafe {
					set_packed_slice_unchecked(evals, i, extension_scalar);
				}
			}
		} else {
			// For many-to-one embedding, use ExtensionField::from_bases_sparse
			let mut bases = [P::Scalar::default(); 1 << MAX_TOWER_HEIGHT];
			let bases = &mut bases[0..1 << log_embedding_degree];

			let bases_count = 1 << log_embedding_degree.min(subcube_vars);
			for i in 0..1 << subcube_vars.saturating_sub(log_embedding_degree) {
				for (j, base) in bases[..bases_count].iter_mut().enumerate() {
					// Safety: i > 0 iff log_embedding_degree < subcube_vars and subcube_index < max_index check
					*base = unsafe {
						get_packed_slice_unchecked(
							self.0.evals(),
							subcube_start + (i << log_embedding_degree) + j,
						)
					};
				}

				let extension_scalar = PE::Scalar::from_bases_sparse(
					bases,
					log_extension_degree - log_embedding_degree,
				)?;

				// Safety: i < 1 << min(0, subcube_vars - log_embedding_degree) <= correct_len * PE::WIDTH
				unsafe {
					set_packed_slice_unchecked(evals, i, extension_scalar);
				}
			}
		}

		Ok(())
	}

	fn packed_evals(&self) -> Option<&[PE]> {
		Some(PE::cast_exts(self.0.evals()))
	}
}

impl<P, Data> MultilinearExtension<P, Data>
where
	P: PackedField,
	Data: Deref<Target = [P]>,
{
	pub fn specialize<PE>(self) -> MLEEmbeddingAdapter<P, PE, Data>
	where
		PE: PackedField,
		PE::Scalar: ExtensionField<P::Scalar>,
	{
		MLEEmbeddingAdapter::from(self)
	}
}

impl<'a, P, Data> MultilinearExtension<P, Data>
where
	P: PackedField,
	Data: Deref<Target = [P]> + Send + Sync + Debug + 'a,
{
	pub fn specialize_arc_dyn<PE>(self) -> Arc<dyn MultilinearPoly<PE> + Send + Sync + 'a>
	where
		PE: PackedField + RepackedExtension<P>,
		PE::Scalar: ExtensionField<P::Scalar>,
	{
		self.specialize().upcast_arc_dyn()
	}
}

/// An adapter for [`MultilinearExtension`] that implements [`MultilinearPoly`] over the same
/// packed field that the [`MultilinearExtension`] stores evaluations in.
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct MLEDirectAdapter<P, Data = Vec<P>>(MultilinearExtension<P, Data>)
where
	P: PackedField,
	Data: Deref<Target = [P]>;

impl<'a, P, Data> MLEDirectAdapter<P, Data>
where
	P: PackedField,
	Data: Deref<Target = [P]> + Send + Sync + Debug + 'a,
{
	pub fn upcast_arc_dyn(self) -> Arc<dyn MultilinearPoly<P> + Send + Sync + 'a> {
		Arc::new(self)
	}

	/// Given a ($mu$-variate) multilinear function $f$ and an element $r$,
	/// return the multilinear function $f(r, X_1, ..., X_{\mu - 1})$.
	pub fn evaluate_zeroth_variable(&self, r: P::Scalar) -> Result<MultilinearExtension<P>, Error> {
		let multilin = &self.0;
		let mu = multilin.n_vars();
		if mu == 0 {
			bail!(Error::ConstantFold);
		}
		let packed_length = 1 << mu.saturating_sub(P::LOG_WIDTH + 1);
		// in general, the formula is: f(r||w) = r * (f(1||w) - f(0||w)) + f(0||w).
		let result = (0..packed_length)
			.into_par_iter()
			.map(|i| {
				let eval0_minus_eval1 = P::from_fn(|j| {
					let index = (i << P::LOG_WIDTH) | j;
					// necessary if `mu_minus_one` < `P::LOG_WIDTH`
					if index >= 1 << (mu - 1) {
						return P::Scalar::ZERO;
					}
					let eval0 = get_packed_slice(multilin.evals(), index << 1);
					let eval1 = get_packed_slice(multilin.evals(), (index << 1) | 1);
					eval0 - eval1
				});
				let eval0 = P::from_fn(|j| {
					let index = (i << P::LOG_WIDTH) | j;
					// necessary if `mu_minus_one` < `P::LOG_WIDTH`
					if index >= 1 << (mu - 1) {
						return P::Scalar::ZERO;
					}
					get_packed_slice(multilin.evals(), index << 1)
				});
				eval0_minus_eval1 * r + eval0
			})
			.collect::<Vec<_>>();

		MultilinearExtension::new(mu - 1, result)
	}
}

impl<P, Data> From<MultilinearExtension<P, Data>> for MLEDirectAdapter<P, Data>
where
	P: PackedField,
	Data: Deref<Target = [P]>,
{
	fn from(inner: MultilinearExtension<P, Data>) -> Self {
		Self(inner)
	}
}

impl<P, Data> AsRef<MultilinearExtension<P, Data>> for MLEDirectAdapter<P, Data>
where
	P: PackedField,
	Data: Deref<Target = [P]>,
{
	fn as_ref(&self) -> &MultilinearExtension<P, Data> {
		&self.0
	}
}

impl<F, P, Data> MultilinearPoly<P> for MLEDirectAdapter<P, Data>
where
	F: Field,
	P: PackedField<Scalar = F> + Debug,
	Data: Deref<Target = [P]> + Send + Sync + Debug,
{
	#[inline]
	fn n_vars(&self) -> usize {
		self.0.n_vars()
	}

	#[inline]
	fn log_extension_degree(&self) -> usize {
		0
	}

	fn evaluate_on_hypercube(&self, index: usize) -> Result<F, Error> {
		self.0.evaluate_on_hypercube(index)
	}

	fn evaluate_on_hypercube_and_scale(&self, index: usize, scalar: F) -> Result<F, Error> {
		let eval = self.0.evaluate_on_hypercube(index)?;
		Ok(scalar * eval)
	}

	fn evaluate(&self, query: MultilinearQueryRef<P>) -> Result<F, Error> {
		self.0.evaluate(query)
	}

	fn evaluate_partial_low(
		&self,
		query: MultilinearQueryRef<P>,
	) -> Result<MultilinearExtension<P>, Error> {
		self.0.evaluate_partial_low(query)
	}

	fn evaluate_partial_high(
		&self,
		query: MultilinearQueryRef<P>,
	) -> Result<MultilinearExtension<P>, Error> {
		self.0.evaluate_partial_high(query)
	}

	fn subcube_inner_products(
		&self,
		query: MultilinearQueryRef<P>,
		subcube_vars: usize,
		subcube_index: usize,
		inner_products: &mut [P],
	) -> Result<(), Error> {
		let query_n_vars = query.n_vars();
		if query_n_vars + subcube_vars > self.n_vars() {
			bail!(Error::ArgumentRangeError {
				arg: "query.n_vars() + subcube_vars".into(),
				range: 0..self.n_vars(),
			});
		}

		let max_index = 1 << (self.n_vars() - query_n_vars - subcube_vars);
		if subcube_index >= max_index {
			bail!(Error::ArgumentRangeError {
				arg: "subcube_index".into(),
				range: 0..max_index,
			});
		}

		let correct_len = 1 << subcube_vars.saturating_sub(P::LOG_WIDTH);
		if inner_products.len() != correct_len {
			bail!(Error::ArgumentRangeError {
				arg: "evals.len()".to_string(),
				range: correct_len..correct_len + 1,
			});
		}

		// TODO: Maybe optimize me
		let subcube_start = subcube_index << (query_n_vars + subcube_vars);
		for scalar_index in 0..1 << subcube_vars {
			let evals_start = subcube_start + (scalar_index << query_n_vars);
			let mut inner_product = F::ZERO;
			for i in 0..1 << query_n_vars {
				inner_product += get_packed_slice(query.expansion(), i)
					* get_packed_slice(self.0.evals(), evals_start + i);
			}

			set_packed_slice(inner_products, scalar_index, inner_product);
		}

		Ok(())
	}

	fn subcube_evals(
		&self,
		subcube_vars: usize,
		subcube_index: usize,
		log_embedding_degree: usize,
		evals: &mut [P],
	) -> Result<(), Error> {
		let n_vars = self.n_vars();
		if subcube_vars > n_vars {
			bail!(Error::ArgumentRangeError {
				arg: "subcube_vars".to_string(),
				range: 0..n_vars + 1,
			});
		}

		if log_embedding_degree != 0 {
			bail!(Error::LogEmbeddingDegreeTooLarge {
				log_embedding_degree
			});
		}

		let correct_len = 1 << subcube_vars.saturating_sub(P::LOG_WIDTH);
		if evals.len() != correct_len {
			bail!(Error::ArgumentRangeError {
				arg: "evals.len()".to_string(),
				range: correct_len..correct_len + 1,
			});
		}

		let max_index = 1 << (n_vars - subcube_vars);
		if subcube_index >= max_index {
			bail!(Error::ArgumentRangeError {
				arg: "subcube_index".to_string(),
				range: 0..max_index,
			});
		}

		if subcube_vars < P::LOG_WIDTH {
			let subcube_start = subcube_index << subcube_vars;
			for i in 0..1 << subcube_vars {
				// Safety: subcube_index < max_index check
				let scalar =
					unsafe { get_packed_slice_unchecked(self.0.evals(), subcube_start + i) };

				// Safety: i < 1 << min(0, subcube_vars) <= correct_len * P::WIDTH
				unsafe {
					set_packed_slice_unchecked(evals, i, scalar);
				}
			}
		} else {
			let range = subcube_index << (subcube_vars - P::LOG_WIDTH)
				..(subcube_index + 1) << (subcube_vars - P::LOG_WIDTH);
			evals.copy_from_slice(&self.0.evals()[range]);
		}

		Ok(())
	}

	fn packed_evals(&self) -> Option<&[P]> {
		Some(self.0.evals())
	}
}

#[cfg(test)]
mod tests {
	use super::*;
	use crate::{tensor_prod_eq_ind, MultilinearQuery};
	use binius_field::{
		arch::OptimalUnderlier256b, as_packed_field::PackedType, packed::iter_packed_slice,
		BinaryField128b, BinaryField16b, BinaryField32b, BinaryField8b, PackedBinaryField16x8b,
		PackedBinaryField1x128b, PackedBinaryField4x32b, PackedBinaryField8x16b, PackedExtension,
		PackedField, PackedFieldIndexable,
	};
	use rand::prelude::*;
	use std::iter::repeat_with;

	type F = BinaryField16b;
	type P = PackedBinaryField8x16b;

	fn multilinear_query<P: PackedField>(p: &[P::Scalar]) -> MultilinearQuery<P, Vec<P>> {
		let mut result = vec![P::default(); 1 << p.len().saturating_sub(P::LOG_WIDTH)];
		result[0] = P::set_single(P::Scalar::ONE);
		tensor_prod_eq_ind(0, &mut result, p).unwrap();
		MultilinearQuery::with_expansion(p.len(), result).unwrap()
	}

	#[test]
	fn test_evaluate_subcube_and_evaluate_partial_low_consistent() {
		let mut rng = StdRng::seed_from_u64(0);
		let poly = MultilinearExtension::from_values(
			repeat_with(|| PackedBinaryField4x32b::random(&mut rng))
				.take(1 << 8)
				.collect(),
		)
		.unwrap()
		.specialize::<PackedBinaryField1x128b>();

		let q = repeat_with(|| <BinaryField128b as PackedField>::random(&mut rng))
			.take(6)
			.collect::<Vec<_>>();
		let query = multilinear_query(&q);

		let partial_low = poly.evaluate_partial_low(query.to_ref()).unwrap();

		let mut inner_products = vec![PackedBinaryField1x128b::zero(); 16];
		poly.subcube_inner_products(query.to_ref(), 4, 0, inner_products.as_mut_slice())
			.unwrap();

		for (idx, inner_product) in iter_packed_slice(&inner_products).enumerate() {
			assert_eq!(inner_product, partial_low.evaluate_on_hypercube(idx).unwrap(),);
		}
	}

	#[test]
	fn test_evaluate_subcube_smaller_than_packed_width() {
		let mut rng = StdRng::seed_from_u64(0);
		let poly = MultilinearExtension::new(
			2,
			vec![PackedBinaryField16x8b::from_scalars(
				[2, 2, 9, 9].map(BinaryField8b::new),
			)],
		)
		.unwrap()
		.specialize::<PackedBinaryField1x128b>();

		let q = repeat_with(|| <BinaryField128b as PackedField>::random(&mut rng))
			.take(1)
			.collect::<Vec<_>>();
		let query = multilinear_query(&q);

		let mut inner_products = vec![PackedBinaryField1x128b::zero(); 2];
		poly.subcube_inner_products(query.to_ref(), 1, 0, inner_products.as_mut_slice())
			.unwrap();

		assert_eq!(get_packed_slice(&inner_products, 0), BinaryField128b::new(2));
		assert_eq!(get_packed_slice(&inner_products, 1), BinaryField128b::new(9));
	}

	#[test]
	fn test_subcube_evals_embeds_correctly() {
		let mut rng = StdRng::seed_from_u64(0);

		type P = PackedBinaryField16x8b;
		type PE = PackedBinaryField1x128b;

		let packed_count = 4;
		let values: Vec<_> = repeat_with(|| P::random(&mut rng))
			.take(1 << packed_count)
			.collect();

		let mle = MultilinearExtension::from_values(values).unwrap();
		let mles = MLEEmbeddingAdapter::<P, PE, _>::from(mle);

		let bytes_values = P::unpack_scalars(mles.0.evals());

		let n_vars = packed_count + P::LOG_WIDTH;
		let mut evals = vec![PE::zero(); 1 << n_vars];
		for subcube_vars in 0..n_vars {
			for subcube_index in 0..1 << (n_vars - subcube_vars) {
				for log_embedding_degree in 0..=4 {
					let evals_subcube = &mut evals
						[0..1 << subcube_vars.saturating_sub(log_embedding_degree + PE::LOG_WIDTH)];

					mles.subcube_evals(
						subcube_vars,
						subcube_index,
						log_embedding_degree,
						evals_subcube,
					)
					.unwrap();

					let bytes_evals = P::unpack_scalars(
						<PE as PackedExtension<BinaryField8b>>::cast_bases(evals_subcube),
					);

					let shift = 4 - log_embedding_degree;
					let skip_mask = (1 << shift) - 1;
					for (i, &b_evals) in bytes_evals.iter().enumerate() {
						let b_values = if i & skip_mask == 0 && i < 1 << (subcube_vars + shift) {
							bytes_values[(subcube_index << subcube_vars) + (i >> shift)]
						} else {
							BinaryField8b::ZERO
						};
						assert_eq!(b_evals, b_values);
					}
				}
			}
		}
	}

	#[test]
	fn test_subcube_inner_products_and_evaluate_partial_low_conform() {
		let mut rng = StdRng::seed_from_u64(0);
		let n_vars = 12;
		let evals = repeat_with(|| P::random(&mut rng))
			.take(1 << (n_vars - P::LOG_WIDTH))
			.collect::<Vec<_>>();
		let mle = MultilinearExtension::from_values(evals).unwrap();
		let mles = MLEDirectAdapter::from(mle);
		let q = repeat_with(|| Field::random(&mut rng))
			.take(6)
			.collect::<Vec<F>>();
		let query = multilinear_query(&q);
		let partial_eval = mles.evaluate_partial_low(query.to_ref()).unwrap();

		let subcube_vars = 4;
		let mut evals = vec![P::default(); 1 << (subcube_vars - P::LOG_WIDTH)];
		for subcube_index in 0..(n_vars - query.n_vars() - subcube_vars) {
			mles.subcube_inner_products(
				query.to_ref(),
				subcube_vars,
				subcube_index,
				evals.as_mut_slice(),
			)
			.unwrap();
			for hypercube_idx in 0..(1 << subcube_vars) {
				assert_eq!(
					get_packed_slice(&evals, hypercube_idx),
					partial_eval
						.evaluate_on_hypercube(hypercube_idx + (subcube_index << subcube_vars))
						.unwrap()
				);
			}
		}
	}

	#[test]
	fn test_packed_evals_against_subcube_evals() {
		type U = OptimalUnderlier256b;
		type P = PackedType<U, BinaryField32b>;
		type PExt = PackedType<U, BinaryField128b>;

		let mut rng = StdRng::seed_from_u64(0);
		let evals = repeat_with(|| P::random(&mut rng))
			.take(2)
			.collect::<Vec<_>>();
		let mle = MultilinearExtension::from_values(evals.clone()).unwrap();
		let poly = MLEEmbeddingAdapter::from(mle);
		assert_eq!(
			<PExt as PackedExtension<BinaryField32b>>::cast_bases(poly.packed_evals().unwrap()),
			&evals
		);

		let mut evals_out = vec![PExt::zero(); 2];
		poly.subcube_evals(poly.n_vars(), 0, poly.log_extension_degree(), evals_out.as_mut_slice())
			.unwrap();
		assert_eq!(evals_out, poly.packed_evals().unwrap());
	}

	#[test]
	fn test_evaluate_zeroth_evaluate_partial_low_consistent() {
		let mut rng = StdRng::seed_from_u64(0);
		let values: Vec<_> = repeat_with(|| PackedBinaryField4x32b::random(&mut rng))
			.take(1 << 8)
			.collect();

		let me = MultilinearExtension::from_values(values).unwrap();
		let mled = MLEDirectAdapter::from(me);
		let r = <BinaryField32b as PackedField>::random(&mut rng);

		let eval_1: MultilinearExtension<PackedBinaryField4x32b> =
			mled.evaluate_zeroth_variable(r).unwrap();
		let eval_2 = mled
			.evaluate_partial_low(multilinear_query(&[r]).to_ref())
			.unwrap();
		assert_eq!(eval_1, eval_2);
	}
}