cera 0.5.4

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

#include <metal_stdlib>
#include <metal_math>
#include <metal_texture>
using namespace metal;

#line 45 "cera/src/backend/shaders/slang/per_head_rmsnorm.slang"
struct KernelContext_0
{
    packed_uint4 device* par_buf_0;
    float device* x_buf_0;
    float device* weight_buf_0;
    array<float, int(256)> threadgroup* scratch_0;
};


#line 39
float block_sum_0(uint tid_0, float v_0, KernelContext_0 thread* kernelContext_0)
{



    float sg_0 = simd_sum(v_0);
    if((tid_0 & 31U) == 0U)
    {

#line 45
        (*kernelContext_0->scratch_0)[tid_0 >> 5U] = sg_0;

#line 45
    }
    threadgroup_barrier(mem_flags::mem_threadgroup);

#line 46
    float lane_0;

    if(tid_0 < 8U)
    {

#line 48
        lane_0 = (*kernelContext_0->scratch_0)[tid_0];

#line 48
    }
    else
    {

#line 48
        lane_0 = 0.0f;

#line 48
    }
    float total_0 = simd_sum(lane_0);
    if(tid_0 == 0U)
    {

#line 50
        (*kernelContext_0->scratch_0)[int(0)] = total_0;

#line 50
    }
    threadgroup_barrier(mem_flags::mem_threadgroup);
    float _S1 = (*kernelContext_0->scratch_0)[int(0)];

#line 67
    return _S1;
}



[[kernel]] void per_head_rmsnorm(uint3 lid_0 [[thread_position_in_threadgroup]], uint3 wid_0 [[threadgroup_position_in_grid]], packed_uint4 device* par_buf_1 [[buffer(2)]], float device* x_buf_1 [[buffer(0)]], float device* weight_buf_1 [[buffer(1)]])
{

#line 72
    thread KernelContext_0 kernelContext_1;

#line 72
    (&kernelContext_1)->par_buf_0 = par_buf_1;

#line 72
    (&kernelContext_1)->x_buf_0 = x_buf_1;

#line 72
    (&kernelContext_1)->weight_buf_0 = weight_buf_1;

#line 72
    threadgroup array<float, int(256)> scratch_1;

#line 72
    (&kernelContext_1)->scratch_0 = &scratch_1;
    uint tid_1 = lid_0.x;

    uint head_dim_0 = (uint4(*(par_buf_1+int(0))) ).x;
    float eps_0 = (as_type<float>(((uint4(*(par_buf_1+int(0))) ).y)));
    uint _S2 = wid_0.x * head_dim_0;

#line 77
    uint i_0 = tid_1;

#line 77
    float partial_0 = 0.0f;



    for(;;)
    {

#line 81
        if(i_0 < head_dim_0)
        {
        }
        else
        {

#line 81
            break;
        }

#line 82
        float device* _S3 = (&kernelContext_1)->x_buf_0+(_S2 + i_0);
        float partial_1 = partial_0 + *_S3 * *_S3;

#line 81
        i_0 = i_0 + 256U;

#line 81
        partial_0 = partial_1;

#line 81
    }

#line 81
    float _S4 = block_sum_0(tid_1, partial_0, &kernelContext_1);

#line 87
    float _S5 = 1.0f / sqrt(_S4 / float(head_dim_0) + eps_0);

#line 87
    i_0 = tid_1;


    for(;;)
    {

#line 90
        if(i_0 < head_dim_0)
        {
        }
        else
        {

#line 90
            break;
        }

#line 91
        uint _S6 = _S2 + i_0;

#line 91
        *((&kernelContext_1)->x_buf_0+_S6) = *((&kernelContext_1)->x_buf_0+_S6) * _S5 * (&kernelContext_1)->weight_buf_0[i_0];

#line 90
        i_0 = i_0 + 256U;

#line 90
    }


    return;
}