aprender-contracts 0.34.0

Papers to Math to Contracts in Code — YAML contract parsing, validation, scaffold generation, and Kani harness codegen for provable Rust kernels
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

/// PTX assembly for the `ReLU` kernel (elementwise, 1 thread per element).
pub fn relu_ptx() -> &'static str {
    r".version 8.5
.target sm_90
.address_size 64
.visible .entry relu_kernel(
    .param .u64 input,
    .param .u64 output,
    .param .u32 n
) {
    .reg .u32 %tid, %ntid, %ctaid, %idx, %n;
    .reg .u64 %in_ptr, %out_ptr, %off;
    .reg .f32 %x, %zero, %y;
    .reg .pred %p;

    mov.u32 %tid, %tid.x;
    mov.u32 %ntid, %ntid.x;
    mov.u32 %ctaid, %ctaid.x;
    mad.lo.u32 %idx, %ctaid, %ntid, %tid;
    ld.param.u32 %n, [n];
    setp.ge.u32 %p, %idx, %n;
    @%p bra DONE;

    ld.param.u64 %in_ptr, [input];
    ld.param.u64 %out_ptr, [output];
    mul.wide.u32 %off, %idx, 4;
    add.u64 %in_ptr, %in_ptr, %off;
    add.u64 %out_ptr, %out_ptr, %off;
    ld.global.f32 %x, [%in_ptr];
    mov.f32 %zero, 0f00000000;
    max.f32 %y, %x, %zero;
    st.global.f32 [%out_ptr], %y;

DONE:
    ret;
}
"
}

/// PTX assembly for the `GELU` kernel.
///
/// Uses the approximation: GELU(x) = 0.5 * x * (1 + tanh(sqrt(2/pi) * (x + 0.044715 * x^3)))
/// where tanh(a) is computed as 1 - 2/(1 + exp(2a)) and
/// exp(a) = 2^(a / ln2) via `ex2.approx.f32`.
pub fn gelu_ptx() -> &'static str {
    r".version 8.5
.target sm_90
.address_size 64
.visible .entry gelu_kernel(
    .param .u64 input,
    .param .u64 output,
    .param .u32 n
) {
    .reg .u32 %tid, %ntid, %ctaid, %idx, %n;
    .reg .u64 %in_ptr, %out_ptr, %off;
    .reg .f32 %x, %y, %x3, %inner, %two_inner, %scaled;
    .reg .f32 %exp_val, %denom, %tanh_val, %one_plus_tanh, %half_x;
    .reg .f32 %k_sqrt2pi, %k_coeff, %k_half, %k_one, %k_two, %k_rcp_ln2;
    .reg .pred %p;

    mov.u32 %tid, %tid.x;
    mov.u32 %ntid, %ntid.x;
    mov.u32 %ctaid, %ctaid.x;
    mad.lo.u32 %idx, %ctaid, %ntid, %tid;
    ld.param.u32 %n, [n];
    setp.ge.u32 %p, %idx, %n;
    @%p bra DONE;

    ld.param.u64 %in_ptr, [input];
    ld.param.u64 %out_ptr, [output];
    mul.wide.u32 %off, %idx, 4;
    add.u64 %in_ptr, %in_ptr, %off;
    add.u64 %out_ptr, %out_ptr, %off;
    ld.global.f32 %x, [%in_ptr];

    // Constants
    mov.f32 %k_sqrt2pi, 0f3F4C422A;   // sqrt(2/pi) ~ 0.7978845608
    mov.f32 %k_coeff, 0f3D372713;     // 0.044715
    mov.f32 %k_half, 0f3F000000;      // 0.5
    mov.f32 %k_one, 0f3F800000;       // 1.0
    mov.f32 %k_two, 0f40000000;       // 2.0
    mov.f32 %k_rcp_ln2, 0f3FB8AA3B;   // 1/ln(2) ~ 1.442695

    // x^3
    mul.f32 %x3, %x, %x;
    mul.f32 %x3, %x3, %x;

    // inner = sqrt(2/pi) * (x + 0.044715 * x^3)
    fma.rn.f32 %inner, %k_coeff, %x3, %x;
    mul.f32 %inner, %k_sqrt2pi, %inner;

    // tanh(inner) = 1 - 2/(1 + exp(2*inner))
    // exp(2*inner) via ex2: exp(a) = 2^(a/ln2)
    mul.f32 %two_inner, %k_two, %inner;
    mul.f32 %scaled, %two_inner, %k_rcp_ln2;
    ex2.approx.f32 %exp_val, %scaled;
    add.f32 %denom, %k_one, %exp_val;
    rcp.approx.f32 %denom, %denom;
    // tanh = 1 - 2*rcp(1+exp(2*inner))
    mul.f32 %tanh_val, %k_two, %denom;
    sub.f32 %tanh_val, %k_one, %tanh_val;

    // 0.5 * x * (1 + tanh)
    add.f32 %one_plus_tanh, %k_one, %tanh_val;
    mul.f32 %half_x, %k_half, %x;
    mul.f32 %y, %half_x, %one_plus_tanh;

    st.global.f32 [%out_ptr], %y;

DONE:
    ret;
}
"
}

/// PTX assembly for the `SiLU` (Swish) kernel.
///
/// `SiLU`(x) = x / (1 + exp(-x)), where exp(-x) = 2^(-x / ln2) via `ex2.approx.f32`.
pub fn silu_ptx() -> &'static str {
    r".version 8.5
.target sm_90
.address_size 64
.visible .entry silu_kernel(
    .param .u64 input,
    .param .u64 output,
    .param .u32 n
) {
    .reg .u32 %tid, %ntid, %ctaid, %idx, %n;
    .reg .u64 %in_ptr, %out_ptr, %off;
    .reg .f32 %x, %y, %neg_x, %scaled, %exp_val, %denom, %rcp_denom;
    .reg .f32 %k_one, %k_rcp_ln2;
    .reg .pred %p;

    mov.u32 %tid, %tid.x;
    mov.u32 %ntid, %ntid.x;
    mov.u32 %ctaid, %ctaid.x;
    mad.lo.u32 %idx, %ctaid, %ntid, %tid;
    ld.param.u32 %n, [n];
    setp.ge.u32 %p, %idx, %n;
    @%p bra DONE;

    ld.param.u64 %in_ptr, [input];
    ld.param.u64 %out_ptr, [output];
    mul.wide.u32 %off, %idx, 4;
    add.u64 %in_ptr, %in_ptr, %off;
    add.u64 %out_ptr, %out_ptr, %off;
    ld.global.f32 %x, [%in_ptr];

    // Constants
    mov.f32 %k_one, 0f3F800000;       // 1.0
    mov.f32 %k_rcp_ln2, 0f3FB8AA3B;   // 1/ln(2) ~ 1.442695

    // exp(-x) = 2^(-x / ln2) = 2^(-x * (1/ln2))
    neg.f32 %neg_x, %x;
    mul.f32 %scaled, %neg_x, %k_rcp_ln2;
    ex2.approx.f32 %exp_val, %scaled;

    // silu = x / (1 + exp(-x))
    add.f32 %denom, %k_one, %exp_val;
    rcp.approx.f32 %rcp_denom, %denom;
    mul.f32 %y, %x, %rcp_denom;

    st.global.f32 [%out_ptr], %y;

DONE:
    ret;
}
"
}