1  #version 450
 2  
 3  #include "dequant_head.comp"
 4  
 5  layout(local_size_x = 64, 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 ip = tid / 32;
19          const uint il = tid - 32 * ip;
20          const uint is = 8 * ip + il / 16;
21  
22          const uint y_idx = i * QUANT_K + 128 * ip + il;
23  
24          const uint ql_idx = 32 * ip + il;
25          const uint8_t qs = data_a[i].qs[32 * ip + il];
26  
27          FLOAT_TYPE dall = FLOAT_TYPE(data_a[i].d.x);
28          FLOAT_TYPE dmin = FLOAT_TYPE(data_a[i].d.y);
29          data_b[y_idx +  0] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+0] & 0xF) * ((qs >> 0) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+0] >> 4));
30          data_b[y_idx + 32] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+2] & 0xF) * ((qs >> 2) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+2] >> 4));
31          data_b[y_idx + 64] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+4] & 0xF) * ((qs >> 4) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+4] >> 4));
32          data_b[y_idx + 96] = D_TYPE(dall * FLOAT_TYPE((data_a[i].scales[is+6] & 0xF) * ((qs >> 6) & 3)) - dmin * FLOAT_TYPE(data_a[i].scales[is+6] >> 4));
33      }
34  }