cera 0.5.3

Rust-native LLM inference engine
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

@binding(4) @group(0) var<storage, read> par_buf_0 : array<u32>;

@binding(2) @group(0) var<storage, read> w_buf_0 : array<f32>;

@binding(1) @group(0) var<storage, read_write> rbuf_0 : array<f32>;

@binding(0) @group(0) var<storage, read> proj_buf_0 : array<f32>;

@binding(3) @group(0) var<storage, read_write> out_buf_0 : array<f32>;

@compute
@workgroup_size(256, 1, 1)
fn conv1d_fused_batch(@builtin(global_invocation_id) gid_0 : vec3<u32>)
{
    var ch_0 : u32 = gid_0.x;
    var hs_0 : u32 = par_buf_0[i32(0)];
    var ks_0 : u32 = par_buf_0[i32(1)];
    var d_conv_0 : u32 = par_buf_0[i32(2)];
    var _S1 : u32 = par_buf_0[i32(3)];
    var _S2 : u32 = par_buf_0[i32(4)];
    var _S3 : u32 = par_buf_0[i32(5)];
    if(ch_0 >= hs_0)
    {
        return;
    }
    var _S4 : bool;
    if(ks_0 > u32(4))
    {
        _S4 = true;
    }
    else
    {
        _S4 = d_conv_0 > u32(3);
    }
    if(_S4)
    {
        return;
    }
    var w_local_0 : array<f32, i32(4)>;
    w_local_0[i32(0)] = 0.0f;
    w_local_0[i32(1)] = 0.0f;
    w_local_0[i32(2)] = 0.0f;
    w_local_0[i32(3)] = 0.0f;
    if(u32(0) <= d_conv_0)
    {
        _S4 = u32(0) < ks_0;
    }
    else
    {
        _S4 = false;
    }
    if(_S4)
    {
        w_local_0[u32(0)] = w_buf_0[ch_0 * ks_0];
    }
    if(u32(1) <= d_conv_0)
    {
        _S4 = u32(1) < ks_0;
    }
    else
    {
        _S4 = false;
    }
    if(_S4)
    {
        w_local_0[u32(1)] = w_buf_0[ch_0 * ks_0 + u32(1)];
    }
    if(u32(2) <= d_conv_0)
    {
        _S4 = u32(2) < ks_0;
    }
    else
    {
        _S4 = false;
    }
    if(_S4)
    {
        w_local_0[u32(2)] = w_buf_0[ch_0 * ks_0 + u32(2)];
    }
    if(u32(3) <= d_conv_0)
    {
        _S4 = u32(3) < ks_0;
    }
    else
    {
        _S4 = false;
    }
    if(_S4)
    {
        w_local_0[u32(3)] = w_buf_0[ch_0 * ks_0 + u32(3)];
    }
    var rb_0 : array<f32, i32(3)>;
    rb_0[i32(0)] = 0.0f;
    rb_0[i32(1)] = 0.0f;
    rb_0[i32(2)] = 0.0f;
    var _S5 : bool = u32(0) < d_conv_0;
    if(_S5)
    {
        rb_0[u32(0)] = rbuf_0[ch_0];
    }
    var _S6 : bool = u32(1) < d_conv_0;
    if(_S6)
    {
        rb_0[u32(1)] = rbuf_0[hs_0 + ch_0];
    }
    var _S7 : bool = u32(2) < d_conv_0;
    if(_S7)
    {
        rb_0[u32(2)] = rbuf_0[u32(2) * hs_0 + ch_0];
    }
    var t_0 : u32 = u32(0);
    for(;;)
    {
        if(t_0 < _S1)
        {
        }
        else
        {
            break;
        }
        var base_0 : u32 = t_0 * _S2;
        var c_val_0 : f32 = proj_buf_0[base_0 + hs_0 + ch_0];
        var bx_0 : f32 = proj_buf_0[base_0 + ch_0] * proj_buf_0[base_0 + u32(2) * hs_0 + ch_0];
        var sum_0 : f32;
        if(_S5)
        {
            sum_0 = rb_0[u32(0)] * w_local_0[u32(0)];
        }
        else
        {
            sum_0 = 0.0f;
        }
        var sum_1 : f32;
        if(_S6)
        {
            sum_1 = sum_0 + rb_0[u32(1)] * w_local_0[u32(1)];
        }
        else
        {
            sum_1 = sum_0;
        }
        var sum_2 : f32;
        if(_S7)
        {
            sum_2 = sum_1 + rb_0[u32(2)] * w_local_0[u32(2)];
        }
        else
        {
            sum_2 = sum_1;
        }
        var sum_3 : f32 = sum_2 + bx_0 * w_local_0[d_conv_0];
        if(_S6)
        {
            rb_0[u32(0)] = rb_0[u32(1)];
        }
        if(_S7)
        {
            rb_0[u32(1)] = rb_0[u32(2)];
        }
        if(d_conv_0 > u32(0))
        {
            rb_0[d_conv_0 - u32(1)] = bx_0;
        }
        out_buf_0[t_0 * _S3 + ch_0] = c_val_0 * sum_3;
        t_0 = t_0 + u32(1);
    }
    if(_S5)
    {
        rbuf_0[ch_0] = rb_0[u32(0)];
    }
    if(_S6)
    {
        rbuf_0[hs_0 + ch_0] = rb_0[u32(1)];
    }
    if(_S7)
    {
        rbuf_0[u32(2) * hs_0 + ch_0] = rb_0[u32(2)];
    }
    return;
}