ruDNN 0.1.95

Neural network operators for Ruda.
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

//! Dense SDPA with plane-cooperative online softmax and recomputed backward.
//! Scores, probabilities and dropout masks are never stored as L x S tensors.
use ruda_kernel::dsl as kernel_dsl;
use ruda_kernel::dsl::prelude::*;

#[ruda]
fn mask_offset(mask: &Tensor<f32>, position: usize) -> usize {
    let mut remaining = position;
    let mut offset = 0usize;
    let mut dim = mask.rank();
    while dim > 0 {
        dim -= 1;
        offset += (remaining % mask.shape(dim)) * mask.stride(dim);
        remaining /= mask.shape(dim);
    }
    offset
}

#[ruda]
fn keep(position: usize, seed_low: u32, seed_high: u32, base_low: u32, base_high: u32, p: f32) -> f32 {
    let block = u32::cast_from(position / 4);
    let mut a = base_low + block;
    let mut b = base_high;
    if a < base_low { b += 1; }
    let mut c = 0u32;
    let mut d = 0u32;
    let mut ka = seed_low;
    let mut kb = seed_high;
    for _round in 0..10 {
        let ha = a.mul_hi(0xD2511F53u32);
        let hc = c.mul_hi(0xCD9E8D57u32);
        let la = a * 0xD2511F53u32;
        let lc = c * 0xCD9E8D57u32;
        a = hc ^ b ^ ka; b = lc; c = ha ^ d ^ kb; d = la;
        ka += 0x9E3779B9u32; kb += 0xBB67AE85u32;
    }
    let lane = position % 4;
    let mut word = a;
    if lane == 1 { word = b; } else if lane == 2 { word = c; } else if lane == 3 { word = d; }
    let unit = f32::cast_from(word >> 8) * (1.0f32 / 16777216.0f32);
    let mut result = 0.0f32;
    if p < 1.0 && unit >= p { result = 1.0 / (1.0 - p); }
    result
}

#[ruda(launch)]
pub fn forward<F: Float>(q: &Array<F>, k: &Array<F>, v: &Array<F>, mask: &Tensor<f32>,
    out: &mut Array<F>, stats: &mut Array<f32>,
    heads: u32, kv_heads: u32, queries: u32, keys: u32, scale: f32, p: f32,
    seed_low: u32, seed_high: u32, base_low: u32, base_high: u32,
    #[comptime] key_dim: usize, #[comptime] value_dim: usize,
    #[comptime] lanes: usize, #[comptime] causal: bool, #[comptime] mask_kind: u32,
    #[comptime] dropout: bool, #[define(F)] _dtype: StorageType) {
    let row = RUDA_POS_X as usize;
    let lane = UNIT_POS as usize;
    let head = (row / queries as usize) % heads as usize;
    let batch = row / (heads as usize * queries as usize);
    let kh = head / (heads / kv_heads) as usize;
    let mut query = Array::<f32>::new((key_dim + lanes - 1) / lanes);
    let mut acc = Array::<f32>::new((value_dim + lanes - 1) / lanes);
    #[unroll]
    for j in 0usize..((key_dim + lanes - 1) / lanes) {
        let d = lane + j * lanes;
        query[j] = 0.0;
        if d < key_dim { query[j] = f32::cast_from(q[row * key_dim + d]); }
    }
    #[unroll]
    for j in 0usize..((value_dim + lanes - 1) / lanes) { acc[j] = 0.0; }
    let mut maximum = 0.0f32;
    let mut denominator = 0.0f32;
    for token in 0usize..keys as usize {
        let cache_row = (batch * kv_heads as usize + kh) * keys as usize + token;
        let mut partial = 0.0f32;
        #[unroll]
        for j in 0usize..((key_dim + lanes - 1) / lanes) {
            let d = lane + j * lanes;
            if d < key_dim { partial = fma(query[j], f32::cast_from(k[cache_row * key_dim + d]), partial); }
        }
        let mut score = plane_sum(partial) * scale;
        let mut allowed = true;
        if comptime!(causal) { allowed = token <= row % queries as usize; }
        if comptime!(mask_kind != 0) {
            let m = mask[mask_offset(mask, row * keys as usize + token)];
            if comptime!(mask_kind == 1) { allowed = m != 0.0; }
            else { score += m; }
        }
        // Additive -inf excludes a score, matching the public masked_softmax.
        if comptime!(mask_kind == 2) { allowed = score != f32::NEG_INFINITY; }
        let mut weight = 0.0f32;
        let mut old_scale = 1.0f32;
        if allowed {
            let mut next_max = score;
            if denominator != 0.0 { next_max = f32::max(maximum, score); }
            old_scale = 0.0;
            if denominator != 0.0 { old_scale = (maximum - next_max).exp(); }
            weight = (score - next_max).exp();
            denominator = fma(denominator, old_scale, weight);
            maximum = next_max;
        }
        let mut multiplier = 1.0f32;
        if comptime!(dropout) { multiplier = keep(row * keys as usize + token, seed_low, seed_high, base_low, base_high, p); }
        let mut dropped_weight = weight * multiplier;
        if comptime!(dropout) { if p == 1.0 { dropped_weight = 0.0; } }
        let mut accumulator_scale = old_scale;
        if comptime!(dropout) { if p == 1.0 { accumulator_scale = 1.0; } }
        // Masked zero weights still multiply V, preserving IEEE NaN propagation.
        #[unroll]
        for j in 0usize..((value_dim + lanes - 1) / lanes) {
            let d = lane + j * lanes;
            if d < value_dim { acc[j] = fma(dropped_weight, f32::cast_from(v[cache_row * value_dim + d]), acc[j] * accumulator_scale); }
        }
    }
    #[unroll]
    for j in 0usize..((value_dim + lanes - 1) / lanes) {
        let d = lane + j * lanes;
        if d < value_dim {
            let mut result = acc[j];
            if denominator != 0.0 && p != 1.0 { result /= denominator; }
            out[row * value_dim + d] = F::cast_from(result);
        }
    }
    if lane == 0 { stats[row * 2] = maximum; stats[row * 2 + 1] = denominator; }
}

#[ruda(launch)]
pub fn backward<F: Float>(q: &Array<F>, k: &Array<F>, v: &Array<F>, mask: &Tensor<f32>,
    grad: &Array<F>, stats: &Array<f32>, dq: &mut Array<F>,
    dk: &mut Array<Atomic<f32>>, dv: &mut Array<Atomic<f32>>, dm: &mut Array<Atomic<f32>>,
    heads: u32, kv_heads: u32, queries: u32, keys: u32, scale: f32, p: f32,
    seed_low: u32, seed_high: u32, base_low: u32, base_high: u32,
    #[comptime] key_dim: usize, #[comptime] value_dim: usize, #[comptime] lanes: usize,
    #[comptime] causal: bool, #[comptime] mask_kind: u32, #[comptime] dropout: bool,
    #[comptime] need_mask: bool, #[define(F)] _dtype: StorageType) {
    let row = RUDA_POS_X as usize;
    let lane = UNIT_POS as usize;
    let head = (row / queries as usize) % heads as usize;
    let batch = row / (heads as usize * queries as usize);
    let kh = head / (heads / kv_heads) as usize;
    let maximum = stats[row * 2];
    let denominator = stats[row * 2 + 1];
    let mut query = Array::<f32>::new((key_dim + lanes - 1) / lanes);
    let mut qgrad = Array::<f32>::new((key_dim + lanes - 1) / lanes);
    let mut dy = Array::<f32>::new((value_dim + lanes - 1) / lanes);
    #[unroll]
    for j in 0usize..((key_dim + lanes - 1) / lanes) {
        let d = lane + j * lanes; query[j] = 0.0; qgrad[j] = 0.0;
        if d < key_dim { query[j] = f32::cast_from(q[row * key_dim + d]); }
    }
    #[unroll]
    for j in 0usize..((value_dim + lanes - 1) / lanes) {
        let d = lane + j * lanes; dy[j] = 0.0;
        if d < value_dim { dy[j] = f32::cast_from(grad[row * value_dim + d]); }
    }
    let mut expected = 0.0f32;
    // First scan computes the exact FP32 softmax Jacobian contraction. Using
    // low-precision saved O here would introduce an additional rounding error.
    for pass in 0usize..2 {
        for token in 0usize..keys as usize {
            let cache_row = (batch * kv_heads as usize + kh) * keys as usize + token;
            let mut partial = 0.0f32;
            let mut dp_local = 0.0f32;
            #[unroll]
            for j in 0usize..((key_dim + lanes - 1) / lanes) {
                let d = lane + j * lanes;
                if d < key_dim { partial = fma(query[j], f32::cast_from(k[cache_row * key_dim + d]), partial); }
            }
            #[unroll]
            for j in 0usize..((value_dim + lanes - 1) / lanes) {
                let d = lane + j * lanes;
                if d < value_dim { dp_local = fma(dy[j], f32::cast_from(v[cache_row * value_dim + d]), dp_local); }
            }
            let mut score = plane_sum(partial) * scale;
            let mut allowed = true;
            if comptime!(causal) { allowed = token <= row % queries as usize; }
            let mut mi = 0usize;
            if comptime!(mask_kind != 0) {
                mi = mask_offset(mask, row * keys as usize + token);
                let m = mask[mi];
                if comptime!(mask_kind == 1) { allowed = m != 0.0; } else { score += m; }
            }
            if comptime!(mask_kind == 2) { allowed = score != f32::NEG_INFINITY; }
            let mut probability = 0.0f32;
            if allowed && denominator != 0.0 { probability = (score - maximum).exp() / denominator; }
            let mut multiplier = 1.0f32;
            if comptime!(dropout) { multiplier = keep(row * keys as usize + token, seed_low, seed_high, base_low, base_high, p); }
            let mut dp = plane_sum(dp_local) * multiplier;
            if comptime!(dropout) { if p == 1.0 { dp = 0.0; } }
            if pass == 0 { expected += probability * dp; }
            else {
                let mut ds = 0.0f32;
                if allowed { ds = probability * (dp - expected); }
                #[unroll]
                for j in 0usize..((key_dim + lanes - 1) / lanes) {
                    let d = lane + j * lanes;
                    if d < key_dim {
                        qgrad[j] = fma(ds * scale, f32::cast_from(k[cache_row * key_dim + d]), qgrad[j]);
                        dk[cache_row * key_dim + d].fetch_add(ds * scale * query[j]);
                    }
                }
                #[unroll]
                for j in 0usize..((value_dim + lanes - 1) / lanes) {
                    let d = lane + j * lanes;
                    let mut dropped_probability = probability * multiplier;
                    if comptime!(dropout) { if p == 1.0 { dropped_probability = 0.0; } }
                    if d < value_dim { dv[cache_row * value_dim + d].fetch_add(dropped_probability * dy[j]); }
                }
                if comptime!(need_mask) { if lane == 0 { dm[mi].fetch_add(ds); } }
            }
        }
    }
    #[unroll]
    for j in 0usize..((key_dim + lanes - 1) / lanes) {
        let d = lane + j * lanes;
        if d < key_dim { dq[row * key_dim + d] = F::cast_from(qgrad[j]); }
    }
}

#[ruda(launch)]
pub fn zero(out: &mut Array<f32>) {
    let pos = ABSOLUTE_POS as usize;
    if pos < out.len() { out[pos] = 0.0; }
}

#[ruda(launch)]
pub fn narrow<F: Float>(input: &Array<f32>, out: &mut Array<F>, #[define(F)] _dtype: StorageType) {
    let pos = ABSOLUTE_POS as usize;
    if pos < out.len() { out[pos] = F::cast_from(input[pos]); }
}