1  #version 450
 2  
 3  #include "dequant_head.comp"
 4  
 5  layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
 6  
 7  layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
 8  layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
 9  
10  void main() {
11      [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
12          const uint i = gl_WorkGroupID.x * 256 + wgy;
13          if (i >= p.M * p.K / QUANT_K) {
14              return;
15          }
16  
17          const uint tid = gl_LocalInvocationID.x;
18          const uint il = tid / 8;
19          const uint ir = tid % 8;
20          const uint is = 2 * il;
21          const uint n = 4;
22  
23          const FLOAT_TYPE dall = FLOAT_TYPE(data_a[i].d.x);
24          const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[i].d.y);
25  
26          const uint y_idx = i * QUANT_K + 64 * il + n * ir;
27          const uint qs_idx = 32*il + n * ir;
28  
29          uint8_t sc;
30          uint8_t m;
31          if (is < 4) {
32              sc = uint8_t(data_a[i].scales[is] & 63);
33              m  = uint8_t(data_a[i].scales[is + 4] & 63);
34          } else {
35              sc = uint8_t((data_a[i].scales[is + 4] & 0xF) | ((data_a[i].scales[is - 4] >> 6) << 4));
36              m  = uint8_t((data_a[i].scales[is + 4] >>  4) | ((data_a[i].scales[is    ] >> 6) << 4));
37          }
38          const FLOAT_TYPE d1 = dall * sc;
39          const FLOAT_TYPE m1 = dmin * m;
40  
41          if (is < 4) {
42              sc = uint8_t(data_a[i].scales[is + 1] & 63);
43              m  = uint8_t(data_a[i].scales[is + 5] & 63);
44          } else {
45              sc = uint8_t((data_a[i].scales[is + 5] & 0xF) | ((data_a[i].scales[is - 3] >> 6) << 4));
46              m  = uint8_t((data_a[i].scales[is + 5] >>  4) | ((data_a[i].scales[is + 1] >> 6) << 4));
47          }
48          const FLOAT_TYPE d2 = dall * sc;
49          const FLOAT_TYPE m2 = dmin * m;
50  
51          [[unroll]] for (uint l = 0; l < n; ++l) {
52              data_b[y_idx + l     ] = D_TYPE(d1 * FLOAT_TYPE(data_a[i].qs[qs_idx + l] & 0xF) - m1);
53              data_b[y_idx + l + 32] = D_TYPE(d2 * FLOAT_TYPE(data_a[i].qs[qs_idx + l] >>  4) - m2);
54          }
55      }
56  }