1#version 450
 2
 3#extension GL_EXT_control_flow_attributes : require
 4
 5#define BLOCK_SIZE 64
 6layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
 7
 8layout(push_constant) uniform Parameters {
 9    uint B;
10    uint T;
11    uint C;
12    uint H;
13};
14
15layout(binding = 0) readonly buffer KBuf { A_TYPE k[]; };
16layout(binding = 1) readonly buffer VBuf { A_TYPE v[]; };
17layout(binding = 2) readonly buffer RBuf { A_TYPE r[]; };
18layout(binding = 3) readonly buffer TimeFBuf { A_TYPE tf[]; };
19layout(binding = 4) readonly buffer TimeDBuf { A_TYPE td[]; };
20layout(binding = 5) readonly buffer StateBuf { A_TYPE state_in[]; };
21layout(binding = 6) buffer DstBuf { A_TYPE dst[]; };
22
23shared A_TYPE _k[BLOCK_SIZE], _r[BLOCK_SIZE], _tf[BLOCK_SIZE], _td[BLOCK_SIZE];
24
25void main() {
26    const uint head_size = BLOCK_SIZE;
27    const uint batch_id = gl_WorkGroupID.x / H;
28    const uint head_id = gl_WorkGroupID.x % H;
29    const uint tid = gl_LocalInvocationID.x;
30
31    const uint state_size = C * head_size;
32    const uint n_seq_tokens = T / B;
33
34    if (batch_id >= B || head_id >= H) {
35        return;
36    }
37
38    A_TYPE state[BLOCK_SIZE];
39    [[unroll]] for (uint i = 0; i < head_size; i++) {
40        state[i] = state_in[batch_id * state_size + head_id * head_size * head_size
41                          + i * head_size + tid];
42    }
43
44    barrier();
45    _tf[tid] = tf[head_id * head_size + tid];
46    barrier();
47
48    const uint start_t = batch_id * n_seq_tokens * C + head_id * head_size + tid;
49    const uint end_t = (batch_id + 1) * n_seq_tokens * C + head_id * head_size + tid;
50
51    for (uint t = start_t; t < end_t; t += C) {
52        barrier();
53        _k[tid] = k[t];
54        _r[tid] = r[t];
55        _td[tid] = td[t];
56        barrier();
57
58        const A_TYPE v_val = v[t];
59        A_TYPE y = 0.0;
60
61        [[unroll]] for (uint j = 0; j < head_size; j += 4) {
62            vec4 k_vec = vec4(_k[j], _k[j+1], _k[j+2], _k[j+3]);
63            vec4 r_vec = vec4(_r[j], _r[j+1], _r[j+2], _r[j+3]);
64            vec4 tf_vec = vec4(_tf[j], _tf[j+1], _tf[j+2], _tf[j+3]);
65            vec4 td_vec = vec4(_td[j], _td[j+1], _td[j+2], _td[j+3]);
66            vec4 s_vec = vec4(state[j], state[j+1], state[j+2], state[j+3]);
67
68            vec4 kv = k_vec * v_val;
69
70            vec4 temp = tf_vec * kv + s_vec;
71            y += dot(r_vec, temp);
72
73            s_vec = s_vec * td_vec + kv;
74            state[j] = s_vec.x;
75            state[j+1] = s_vec.y;
76            state[j+2] = s_vec.z;
77            state[j+3] = s_vec.w;
78        }
79
80        dst[t] = y;
81    }
82
83    [[unroll]] for (uint i = 0; i < head_size; i++) {
84        dst[T * C + batch_id * state_size + head_id * head_size * head_size
85            + i * head_size + tid] = state[i];
86    }
87}