prism-inference 0.1.18

Bounded Rust CPU inference and tokenization for Prism-compatible GGUF models
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

//! Baseline NEON over one predecoded PQ2/PTQ1 block, reused across token dots.

use fearless_simd::{Neon, f32x4, i8x16, i16x8, i32x4, prelude::*, u8x16, u16x8};

#[inline]
fn reduce4_i32x4_vec(
    r0: i32x4<Neon>,
    r1: i32x4<Neon>,
    r2: i32x4<Neon>,
    r3: i32x4<Neon>,
) -> i32x4<Neon> {
    let p01 = r0.unzip_low(r1) + r0.unzip_high(r1);
    let p23 = r2.unzip_low(r3) + r2.unzip_high(r3);
    p01.unzip_low(p23) + p01.unzip_high(p23)
}

#[inline]
fn reduce4_i32x4(r0: i32x4<Neon>, r1: i32x4<Neon>, r2: i32x4<Neon>, r3: i32x4<Neon>) -> [i32; 4] {
    reduce4_i32x4_vec(r0, r1, r2, r3).to_array()
}

pub(in crate::kernels::prefill) struct Block {
    neon: Neon,
    weights: [i16x8<Neon>; 16],
}

impl Block {
    #[cfg(test)]
    pub(super) fn new(weights: &[i16; 128]) -> Option<Self> {
        let neon = fearless_simd::Level::baseline().as_neon()?;
        Some(Self {
            neon,
            weights: std::array::from_fn(|index| {
                i16x8::from_slice(neon, &weights[index * 8..index * 8 + 8])
            }),
        })
    }

    pub(super) fn from_pq2(packed: &[u8; 32]) -> Option<Self> {
        let neon = fearless_simd::Level::baseline().as_neon()?;
        let mut weights = [i16x8::splat(neon, 0); 16];
        for (half, bytes) in packed.as_chunks::<16>().0.iter().enumerate() {
            let raw = u8x16::from_slice(neon, bytes);
            let a = raw & 3;
            let b = (raw >> 2) & 3;
            let c = (raw >> 4) & 3;
            let d = raw >> 6;
            let ab_low: u16x8<Neon> = a.zip_low(b).bitcast();
            let ab_high: u16x8<Neon> = a.zip_high(b).bitcast();
            let cd_low: u16x8<Neon> = c.zip_low(d).bitcast();
            let cd_high: u16x8<Neon> = c.zip_high(d).bitcast();
            let decoded = [
                ab_low.zip_low(cd_low).bitcast::<i8x16<Neon>>() - 1,
                ab_low.zip_high(cd_low).bitcast::<i8x16<Neon>>() - 1,
                ab_high.zip_low(cd_high).bitcast::<i8x16<Neon>>() - 1,
                ab_high.zip_high(cd_high).bitcast::<i8x16<Neon>>() - 1,
            ];
            for (group, vector) in decoded.into_iter().enumerate() {
                let (low, high) = vector.widen();
                weights[half * 8 + group * 2] = low;
                weights[half * 8 + group * 2 + 1] = high;
            }
        }
        Some(Self { neon, weights })
    }

    pub(super) fn from_ptq1(packed: &[u8; 28]) -> Option<Self> {
        let neon = fearless_simd::Level::baseline().as_neon()?;
        let weights = crate::kernels::neon::decode_ptq1_i16(neon, packed);
        Some(Self { neon, weights })
    }

    #[inline]
    pub(super) fn accumulate_lanes128(
        &self,
        weight_scale: f32,
        activations: &[crate::kernels::Q8Block; 4],
        lanes: &mut [f32; 4],
    ) {
        let mut acc = f32x4::from_slice(self.neon, lanes);
        for (group, act) in activations.iter().enumerate() {
            let halves = act.values.as_chunks::<16>().0;
            let (l0, h0) = i8x16::from_slice(self.neon, &halves[0]).widen();
            let (l1, h1) = i8x16::from_slice(self.neon, &halves[1]).widen();
            let sum_01 = self.weights[group * 4] * l0 + self.weights[group * 4 + 2] * l1;
            let sum_23 = self.weights[group * 4 + 1] * h0 + self.weights[group * 4 + 3] * h1;
            let (q0, q1) = sum_01.widen();
            let (q2, q3) = sum_23.widen();
            let floats = reduce4_i32x4_vec(q0, q1, q2, q3).to_float::<f32x4<Neon>>();
            let d = f32x4::splat(self.neon, weight_scale * act.scale);
            acc = floats.mul_add_precise(d, acc);
        }
        acc.store_slice(lanes);
    }

    #[inline]
    pub(super) fn dot128_q8(&self, activations: &[crate::kernels::Q8Block; 4]) -> f32 {
        let group_sums: [i32x4<Neon>; 4] = std::array::from_fn(|group| {
            let halves = activations[group].values.as_chunks::<16>().0;
            let (l0, h0) = i8x16::from_slice(self.neon, &halves[0]).widen();
            let (l1, h1) = i8x16::from_slice(self.neon, &halves[1]).widen();
            let pairs = (self.weights[group * 4] * l0 + self.weights[group * 4 + 1] * h0)
                + (self.weights[group * 4 + 2] * l1 + self.weights[group * 4 + 3] * h1);
            let (sum_low, sum_high) = pairs.widen();
            sum_low + sum_high
        });
        let dots = reduce4_i32x4_vec(group_sums[0], group_sums[1], group_sums[2], group_sums[3])
            .to_float::<f32x4<Neon>>();
        let scales = f32x4::from_slice(
            self.neon,
            &[
                activations[0].scale,
                activations[1].scale,
                activations[2].scale,
                activations[3].scale,
            ],
        );
        let prod = (dots * scales).to_array();
        (((0.0 + prod[0]) + prod[1]) + prod[2]) + prod[3]
    }

    #[inline]
    pub(super) fn dot128(&self, values: &[i8; 128]) -> i32 {
        // Each signed product is in -256..=254 for accepted -1..=2 weights.
        // Sixteen products per i16 lane are bounded by 4,096 before widening.
        let mut acc = i16x8::splat(self.neon, 0);
        for (index, values) in values.as_chunks::<16>().0.iter().enumerate() {
            let (low, high) = i8x16::from_slice(self.neon, values).widen();
            acc += self.weights[index * 2] * low + self.weights[index * 2 + 1] * high;
        }
        // The complete dot's absolute bound is 128 * 2 * 128 = 32,768.
        let (sum_low, sum_high) = acc.widen();
        (sum_low + sum_high).reduce_sum()
    }

    #[inline]
    pub(super) fn dot128_four(&self, values: [&[i8; 128]; 4]) -> [i32; 4] {
        let mut acc = [i16x8::splat(self.neon, 0); 4];
        let v0 = values[0].as_chunks::<16>().0;
        let v1 = values[1].as_chunks::<16>().0;
        let v2 = values[2].as_chunks::<16>().0;
        let v3 = values[3].as_chunks::<16>().0;
        for index in 0..8 {
            let w_low = self.weights[index * 2];
            let w_high = self.weights[index * 2 + 1];
            let (l0, h0) = i8x16::from_slice(self.neon, &v0[index]).widen();
            let (l1, h1) = i8x16::from_slice(self.neon, &v1[index]).widen();
            let (l2, h2) = i8x16::from_slice(self.neon, &v2[index]).widen();
            let (l3, h3) = i8x16::from_slice(self.neon, &v3[index]).widen();
            acc[0] += w_low * l0 + w_high * h0;
            acc[1] += w_low * l1 + w_high * h1;
            acc[2] += w_low * l2 + w_high * h2;
            acc[3] += w_low * l3 + w_high * h3;
        }
        let (s0_l, s0_h) = acc[0].widen();
        let (s1_l, s1_h) = acc[1].widen();
        let (s2_l, s2_h) = acc[2].widen();
        let (s3_l, s3_h) = acc[3].widen();
        reduce4_i32x4(s0_l + s0_h, s1_l + s1_h, s2_l + s2_h, s3_l + s3_h)
    }

    #[inline]
    pub(super) fn dot128_eight(&self, values: [&[i8; 128]; 8]) -> [i32; 8] {
        let mut acc = [i16x8::splat(self.neon, 0); 8];
        let chunks = values.map(|token| token.as_chunks::<16>().0);
        for index in 0..8 {
            let w_low = self.weights[index * 2];
            let w_high = self.weights[index * 2 + 1];
            for token in 0..8 {
                let (low, high) = i8x16::from_slice(self.neon, &chunks[token][index]).widen();
                acc[token] += w_low * low + w_high * high;
            }
        }
        let (s0_l, s0_h) = acc[0].widen();
        let (s1_l, s1_h) = acc[1].widen();
        let (s2_l, s2_h) = acc[2].widen();
        let (s3_l, s3_h) = acc[3].widen();
        let (s4_l, s4_h) = acc[4].widen();
        let (s5_l, s5_h) = acc[5].widen();
        let (s6_l, s6_h) = acc[6].widen();
        let (s7_l, s7_h) = acc[7].widen();
        let first = reduce4_i32x4(s0_l + s0_h, s1_l + s1_h, s2_l + s2_h, s3_l + s3_h);
        let second = reduce4_i32x4(s4_l + s4_h, s5_l + s5_h, s6_l + s6_h, s7_l + s7_h);
        [
            first[0], first[1], first[2], first[3], second[0], second[1], second[2], second[3],
        ]
    }

    #[cfg(test)]
    #[inline]
    pub(super) fn dot32(&self, group: usize, values: &[i8; 32]) -> i32 {
        let halves = values.as_chunks::<16>().0;
        let (l0, h0) = i8x16::from_slice(self.neon, &halves[0]).widen();
        let (l1, h1) = i8x16::from_slice(self.neon, &halves[1]).widen();
        let pairs = (self.weights[group * 4] * l0 + self.weights[group * 4 + 1] * h0)
            + (self.weights[group * 4 + 2] * l1 + self.weights[group * 4 + 3] * h1);
        let (sum_low, sum_high) = pairs.widen();
        (sum_low + sum_high).reduce_sum()
    }

    #[cfg(test)]
    #[inline]
    pub(super) fn dot32_four(&self, group: usize, values: [&[i8; 32]; 4]) -> [i32; 4] {
        let w0 = self.weights[group * 4];
        let w1 = self.weights[group * 4 + 1];
        let w2 = self.weights[group * 4 + 2];
        let w3 = self.weights[group * 4 + 3];
        let sums = values.map(|token| {
            let halves = token.as_chunks::<16>().0;
            let (l0, h0) = i8x16::from_slice(self.neon, &halves[0]).widen();
            let (l1, h1) = i8x16::from_slice(self.neon, &halves[1]).widen();
            let pairs = (w0 * l0 + w1 * h0) + (w2 * l1 + w3 * h1);
            let (sum_low, sum_high) = pairs.widen();
            sum_low + sum_high
        });
        reduce4_i32x4(sums[0], sums[1], sums[2], sums[3])
    }

    #[cfg(test)]
    #[inline]
    pub(super) fn dot32_eight(&self, group: usize, values: [&[i8; 32]; 8]) -> [i32; 8] {
        let w0 = self.weights[group * 4];
        let w1 = self.weights[group * 4 + 1];
        let w2 = self.weights[group * 4 + 2];
        let w3 = self.weights[group * 4 + 3];
        let sums = values.map(|token| {
            let halves = token.as_chunks::<16>().0;
            let (l0, h0) = i8x16::from_slice(self.neon, &halves[0]).widen();
            let (l1, h1) = i8x16::from_slice(self.neon, &halves[1]).widen();
            let pairs = (w0 * l0 + w1 * h0) + (w2 * l1 + w3 * h1);
            let (sum_low, sum_high) = pairs.widen();
            sum_low + sum_high
        });
        let first = reduce4_i32x4(sums[0], sums[1], sums[2], sums[3]);
        let second = reduce4_i32x4(sums[4], sums[5], sums[6], sums[7]);
        [
            first[0], first[1], first[2], first[3], second[0], second[1], second[2], second[3],
        ]
    }

    #[cfg(test)]
    #[inline]
    pub(super) fn lanes32(&self, group: usize, values: &[i8; 32]) -> [i32; 4] {
        let halves = values.as_chunks::<16>().0;
        let (l0, h0) = i8x16::from_slice(self.neon, &halves[0]).widen();
        let (l1, h1) = i8x16::from_slice(self.neon, &halves[1]).widen();
        // Native DOTPROD lane n sums adjacent groups 4*n and 16+4*n.
        // Keep these four integer sums separate for the original FP32 FMAs.
        let sum_01 = self.weights[group * 4] * l0 + self.weights[group * 4 + 2] * l1;
        let sum_23 = self.weights[group * 4 + 1] * h0 + self.weights[group * 4 + 3] * h1;
        let (lane0, lane1) = sum_01.widen();
        let (lane2, lane3) = sum_23.widen();
        reduce4_i32x4(lane0, lane1, lane2, lane3)
    }

    #[cfg(test)]
    #[inline]
    pub(super) fn lanes32_four(&self, group: usize, values: [&[i8; 32]; 4]) -> [[i32; 4]; 4] {
        let w0 = self.weights[group * 4];
        let w1 = self.weights[group * 4 + 1];
        let w2 = self.weights[group * 4 + 2];
        let w3 = self.weights[group * 4 + 3];
        values.map(|token| {
            let halves = token.as_chunks::<16>().0;
            let (l0, h0) = i8x16::from_slice(self.neon, &halves[0]).widen();
            let (l1, h1) = i8x16::from_slice(self.neon, &halves[1]).widen();
            let sum_01 = w0 * l0 + w2 * l1;
            let sum_23 = w1 * h0 + w3 * h1;
            let (lane0, lane1) = sum_01.widen();
            let (lane2, lane3) = sum_23.widen();
            reduce4_i32x4(lane0, lane1, lane2, lane3)
        })
    }

    #[cfg(test)]
    #[inline]
    pub(super) fn lanes32_eight(&self, group: usize, values: [&[i8; 32]; 8]) -> [[i32; 4]; 8] {
        let w0 = self.weights[group * 4];
        let w1 = self.weights[group * 4 + 1];
        let w2 = self.weights[group * 4 + 2];
        let w3 = self.weights[group * 4 + 3];
        values.map(|token| {
            let halves = token.as_chunks::<16>().0;
            let (l0, h0) = i8x16::from_slice(self.neon, &halves[0]).widen();
            let (l1, h1) = i8x16::from_slice(self.neon, &halves[1]).widen();
            let sum_01 = w0 * l0 + w2 * l1;
            let sum_23 = w1 * h0 + w3 * h1;
            let (lane0, lane1) = sum_01.widen();
            let (lane2, lane3) = sum_23.widen();
            reduce4_i32x4(lane0, lane1, lane2, lane3)
        })
    }
}