diskann-wide 0.48.0

DiskANN is a fast approximate nearest neighbor search library for high dimensional data
Documentation
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
/*
 * Copyright (c) Microsoft Corporation.
 * Licensed under the MIT license.
 */

use crate::traits::{SIMDMask, SIMDVector};

// Helper macros to more quickly define intrinsics.
// The pattern used for all the x86 definitions is pretty uniform, so wrap it into
// a macro.
macro_rules! x86_define_register {
    ($type:ident, $impl:ty, $mask:ty, $scalar:ty, $lanes:literal, $arch:ty) => {
        #[derive(Debug, Clone, Copy)]
        #[allow(non_camel_case_types)]
        #[repr(transparent)]
        pub struct $type(pub $impl);

        impl $type {
            /// Convert `self` to its corresponding [`crate::Emulated`] type.
            #[inline(always)]
            pub fn emulated(self) -> $crate::Emulated<$scalar, $lanes> {
                $crate::Emulated::from_array($crate::arch::Scalar, self.to_array())
            }
        }

        impl $crate::AsSIMD<$type> for $crate::Emulated<$scalar, $lanes> {
            #[inline(always)]
            fn as_simd(self, arch: $arch) -> $type {
                $type::from_array(arch, self.to_array())
            }
        }

        impl SIMDVector for $type {
            type Arch = $arch;
            type Scalar = $scalar;
            type Underlying = $impl;

            type Mask = $mask;
            type ConstLanes = Const<$lanes>;
            const LANES: usize = $lanes;
            const EMULATED: bool = false;

            #[inline(always)]
            fn arch(self) -> $arch {
                // SAFETY: The existence of `self` provides a witness that it is safe to
                // instantiate its architecture.
                unsafe { <$arch>::new() }
            }

            #[inline(always)]
            fn default(arch: $arch) -> Self {
                <Self as X86Default>::x86_default(arch)
            }

            #[inline(always)]
            fn to_underlying(self) -> Self::Underlying {
                self.0
            }

            #[inline(always)]
            fn from_underlying(_: $arch, repr: Self::Underlying) -> Self {
                Self(repr)
            }

            #[inline(always)]
            fn to_array(self) -> [$scalar; $lanes] {
                // SAFETY: Provided the scalar type is an integer or floating point,
                // then all bit patterns are valid between source and destination types.
                // (provided an x86 intrinsic is one of the transmuted types).
                //
                // The source argument is taken by value (no reference conversion) and
                // as long as `T` is `[repr(C)]`, then `[T; N]` will be `[repr(C)]`.
                //
                // The intrinsic types are `[repr(simd)]` which amounts to `[repr(C)]` and
                // change.
                unsafe { std::mem::transmute::<Self, [$scalar; $lanes]>(self) }
            }

            #[inline(always)]
            fn from_array(_: $arch, x: [$scalar; $lanes]) -> Self {
                // SAFETY: Provided the scalar type is an integer or floating point,
                // then all bit patterns are valid between source and destination types.
                // (provided an x86 intrinsic is one of the transmuted types).
                //
                // The source argument is taken by value (no reference conversion) and
                // as long as `T` is `[repr(C)]`, then `[T; N]` will be `[repr(C)]`.
                //
                // The intrinsic types are `[repr(simd)]` which amounts to `[repr(C)]` and
                // change.
                unsafe { std::mem::transmute::<[$scalar; $lanes], Self>(x) }
            }

            #[inline(always)]
            fn splat(arch: $arch, value: Self::Scalar) -> Self {
                <Self as X86Splat>::x86_splat(arch, value)
            }

            #[inline(always)]
            unsafe fn load_simd(arch: $arch, ptr: *const $scalar) -> Self {
                // SAFETY: This has the same safety constraints as the caller.
                unsafe { <Self as X86LoadStore>::load_simd(arch, ptr) }
            }

            #[inline(always)]
            unsafe fn load_simd_masked_logical(
                arch: $arch,
                ptr: *const $scalar,
                mask: $mask,
            ) -> Self {
                // SAFETY: This has the same safety constraints as the caller.
                unsafe { <Self as X86LoadStore>::load_simd_masked_logical(arch, ptr, mask) }
            }

            #[inline(always)]
            unsafe fn load_simd_first(arch: $arch, ptr: *const $scalar, first: usize) -> Self {
                // SAFETY: This has the same safety constraints as the caller.
                unsafe { <Self as X86LoadStore>::load_simd_first(arch, ptr, first) }
            }

            #[inline(always)]
            unsafe fn store_simd(self, ptr: *mut $scalar) {
                // SAFETY: This has the same safety constraints as the caller.
                unsafe { <Self as X86LoadStore>::store_simd(self, ptr) }
            }

            #[inline(always)]
            unsafe fn store_simd_masked_logical(self, ptr: *mut $scalar, mask: $mask) {
                // SAFETY: This has the same safety constraints as the caller.
                unsafe { <Self as X86LoadStore>::store_simd_masked_logical(self, ptr, mask) }
            }

            #[inline(always)]
            unsafe fn store_simd_first(self, ptr: *mut $scalar, first: usize) {
                // SAFETY: This has the same safety constraints as the caller.
                unsafe { <Self as X86LoadStore>::store_simd_first(self, ptr, first) }
            }
        }
    };
}

// Externalize splat implementations to enable fine-grained overloading.
pub(super) trait X86Splat: SIMDVector {
    fn x86_splat(arch: <Self as SIMDVector>::Arch, value: <Self as SIMDVector>::Scalar) -> Self;
}

pub(super) trait X86Default: SIMDVector {
    fn x86_default(arch: <Self as SIMDVector>::Arch) -> Self;
}

pub(super) trait X86LoadStore: SIMDVector {
    unsafe fn load_simd(
        arch: <Self as SIMDVector>::Arch,
        ptr: *const <Self as SIMDVector>::Scalar,
    ) -> Self;
    unsafe fn load_simd_masked_logical(
        arch: <Self as SIMDVector>::Arch,
        ptr: *const <Self as SIMDVector>::Scalar,
        mask: Self::Mask,
    ) -> Self;
    unsafe fn load_simd_first(
        arch: <Self as SIMDVector>::Arch,
        ptr: *const <Self as SIMDVector>::Scalar,
        first: usize,
    ) -> Self {
        // SAFETY: The implementation of `X86LoadStore` is trusted.
        unsafe {
            <Self as X86LoadStore>::load_simd_masked_logical(
                arch,
                ptr,
                Self::Mask::keep_first(arch, first),
            )
        }
    }

    unsafe fn store_simd(self, ptr: *mut <Self as SIMDVector>::Scalar);
    unsafe fn store_simd_masked_logical(
        self,
        ptr: *mut <Self as SIMDVector>::Scalar,
        mask: Self::Mask,
    );
    unsafe fn store_simd_first(self, ptr: *mut <Self as SIMDVector>::Scalar, first: usize) {
        // SAFETY: The implementation of `X86LoadStore` is trusted.
        unsafe {
            <Self as X86LoadStore>::store_simd_masked_logical(
                self,
                ptr,
                Self::Mask::keep_first(self.arch(), first),
            )
        }
    }
}

macro_rules! x86_retarget {
    ($T:path => $U:path) => {
        impl $T {
            #[inline(always)]
            pub fn retarget(self) -> $U {
                <$U>::from_underlying(self.arch().into(), self.to_underlying())
            }

            pub fn from(self, other: $U) -> Self {
                Self::from_underlying(self.arch(), other.to_underlying())
            }
        }
    };
}

/// Utility macro for defining `X86Splat`.
///
/// SAFETY: It is the invoker's responsibility to ensure that the intrinsic is safe to call.
/// That is - any intrinsics invoked must be compatible with `$type`'s associated architecture.
macro_rules! x86_define_splat {
    ($type:ty, $intrinsic:expr, $requires:literal) => {
        impl X86Splat for $type {
            #[inline(always)]
            fn x86_splat(
                _: <Self as SIMDVector>::Arch,
                value: <Self as SIMDVector>::Scalar,
            ) -> Self {
                // SAFETY: The presence of `Arch` proves that this function is safe to call.
                Self(unsafe { $intrinsic(value) })
            }
        }
    };
    // This variant of the macro performs a bitcast to the value that needs to be
    // broadcasted in order to get the types correct for the x86 intrinsic.
    ($type:ty as $cast:ty, $intrinsic:expr, $requires:literal) => {
        impl X86Splat for $type {
            #[inline(always)]
            fn x86_splat(
                _: <Self as SIMDVector>::Arch,
                value: <Self as SIMDVector>::Scalar,
            ) -> Self {
                // SAFETY: The presence of `Arch` proves that this function is safe to call.
                Self(unsafe { $intrinsic(value as $cast) })
            }
        }
    };
}

/// Utility macro for defining `X86Default`.
///
/// SAFETY: It is the invoker's responsibility to ensure that the intrinsic is safe to call.
/// That is - any intrinsics invoked must be compatible with `$type`'s associated architecture.
macro_rules! x86_define_default {
    ($type:ty, $intrinsic:expr, $requires:literal) => {
        impl X86Default for $type {
            #[inline(always)]
            fn x86_default(_: <Self as SIMDVector>::Arch) -> Self {
                // SAFETY: The invoker of this macro must pass the `target_feature`
                // requirement of the intrinsic.
                //
                // That way, if the intrinsic is not available, we get a compile-time error.
                Self(unsafe { $intrinsic() })
            }
        }
    };
}

/// SAFETY: It is the invoker's responsibility to ensure that the provided intrinsics are
/// safe to call.
///
/// That is - any intrinsics invoked must be compatible with `$type`'s associated architecture.
macro_rules! x86_splitjoin {
    (__m512i, $type:path, $half:path) => {
        impl $crate::SplitJoin for $type {
            type Halved = $half;

            #[inline(always)]
            fn split(self) -> crate::LoHi<Self::Halved> {
                // SAFETY: This must only be instantiated for architecture supporting AVX512DQ.
                unsafe {
                    crate::LoHi::new(
                        Self::Halved::from_underlying(
                            self.arch(),
                            _mm512_extracti32x8_epi32(self.0, 0),
                        ),
                        Self::Halved::from_underlying(
                            self.arch(),
                            _mm512_extracti32x8_epi32(self.0, 1),
                        ),
                    )
                }
            }

            #[inline(always)]
            fn join(lohi: crate::LoHi<Self::Halved>) -> Self {
                // SAFETY: Required by instantiator.
                let v = Self::default(lohi.lo.arch()).to_underlying();

                // SAFETY: `_mm512_inserti32x8` requires `AVX512DQ`.
                let v = unsafe {
                    _mm512_inserti32x8(_mm512_inserti32x8(v, lohi.lo.0, 0), lohi.hi.0, 1)
                };
                Self(v)
            }
        }
    };
    ($type:path, $half:path, $split:path, $join:path, $requires:literal) => {
        impl $crate::SplitJoin for $type {
            type Halved = $half;

            #[inline(always)]
            fn split(self) -> $crate::LoHi<$half> {
                // SAFETY: Required by instantiator.
                unsafe { $crate::LoHi::new($half($split(self.0, 0)), $half($split(self.0, 1))) }
            }

            #[inline(always)]
            fn join(lohi: $crate::LoHi<$half>) -> Self {
                // SAFETY: Required by instantiator.
                Self(unsafe { $join(lohi.hi.0, lohi.lo.0) })
            }
        }
    };
}

macro_rules! x86_avx512_int_comparisons {
    ($type:ty, $intrinsic:ident, $requires:literal) => {
        impl $crate::SIMDPartialEq for $type {
            #[inline(always)]
            fn eq_simd(self, other: Self) -> Self::Mask {
                // SAFETY: Caller asserts that this intrinsic is safe to call for the
                // architecture stored in `$type`.
                Self::Mask::from_underlying(self.arch(), unsafe {
                    $intrinsic::<_MM_CMPINT_EQ>(self.0, other.0)
                })
            }

            #[inline(always)]
            fn ne_simd(self, other: Self) -> Self::Mask {
                // SAFETY: Caller asserts that this intrinsic is safe to call for the
                // architecture stored in `$type`.
                Self::Mask::from_underlying(self.arch(), unsafe {
                    $intrinsic::<_MM_CMPINT_NE>(self.0, other.0)
                })
            }
        }

        impl $crate::SIMDPartialOrd for $type {
            #[inline(always)]
            fn lt_simd(self, other: Self) -> Self::Mask {
                // SAFETY: Caller asserts that this intrinsic is safe to call for the
                // architecture stored in `$type`.
                Self::Mask::from_underlying(self.arch(), unsafe {
                    $intrinsic::<_MM_CMPINT_LT>(self.0, other.0)
                })
            }

            #[inline(always)]
            fn le_simd(self, other: Self) -> Self::Mask {
                // SAFETY: Caller asserts that this intrinsic is safe to call for the
                // architecture stored in `$type`.
                Self::Mask::from_underlying(self.arch(), unsafe {
                    $intrinsic::<_MM_CMPINT_LE>(self.0, other.0)
                })
            }

            #[inline(always)]
            fn gt_simd(self, other: Self) -> Self::Mask {
                // SAFETY: Caller asserts that this intrinsic is safe to call for the
                // architecture stored in `$type`.
                Self::Mask::from_underlying(self.arch(), unsafe {
                    $intrinsic::<_MM_CMPINT_NLE>(self.0, other.0)
                })
            }

            #[inline(always)]
            fn ge_simd(self, other: Self) -> Self::Mask {
                // SAFETY: Caller asserts that this intrinsic is safe to call for the
                // architecture stored in `$type`.
                Self::Mask::from_underlying(self.arch(), unsafe {
                    $intrinsic::<_MM_CMPINT_NLT>(self.0, other.0)
                })
            }
        }
    };
}

macro_rules! x86_avx512_load_store {
    ($T:ty,
     $load:ident,
     $mask_load:ident,
     $store:ident,
     $mask_store:ident,
     $cast:ty,
     $requires:literal
    ) => {
        impl $crate::arch::x86_64::macros::X86LoadStore for $T {
            #[inline(always)]
            unsafe fn load_simd(
                arch: <Self as $crate::SIMDVector>::Arch,
                ptr: *const <Self as $crate::SIMDVector>::Scalar,
            ) -> Self {
                // SAFETY: Instantiator asserts that `$load` is withihn the capabilities
                // of the associated `Arch`.
                Self::from_underlying(arch, unsafe { $load(ptr.cast::<$cast>()) })
            }

            #[inline(always)]
            unsafe fn load_simd_masked_logical(
                arch: <Self as $crate::SIMDVector>::Arch,
                ptr: *const <Self as $crate::SIMDVector>::Scalar,
                mask: <Self as $crate::SIMDVector>::Mask,
            ) -> Self {
                // SAFETY: Instantiator asserts that `$mask_load` is withihn the capabilities
                // of the associated `Arch`.
                Self::from_underlying(arch, unsafe { $mask_load(mask.0, ptr.cast::<$cast>()) })
            }

            #[inline(always)]
            unsafe fn store_simd(self, ptr: *mut <Self as $crate::SIMDVector>::Scalar) {
                // SAFETY: Instantiator asserts that `$store` is withihn the capabilities
                // of the associated `Arch`.
                unsafe { $store(ptr.cast::<$cast>(), self.0) }
            }

            #[inline(always)]
            unsafe fn store_simd_masked_logical(
                self,
                ptr: *mut <Self as $crate::SIMDVector>::Scalar,
                mask: <Self as $crate::SIMDVector>::Mask,
            ) {
                // SAFETY: Instantiator asserts that `$mask_store` is withihn the capabilities
                // of the associated `Arch`.
                unsafe { $mask_store(ptr.cast::<$cast>(), mask.0, self.0) }
            }
        }
    };
}

pub(crate) use x86_avx512_int_comparisons;
pub(crate) use x86_avx512_load_store;
pub(crate) use x86_define_default;
pub(crate) use x86_define_register;
pub(crate) use x86_define_splat;
pub(crate) use x86_retarget;
pub(crate) use x86_splitjoin;