1  void main() {
 2      const uint i = gl_WorkGroupID.x;
 3      const int r = inB[i + pcs.inBOff];
 4  
 5      int z = 0;
 6      for (uint ind = gl_LocalInvocationID.x; ind < pcs.ne00/16; ind += gl_WorkGroupSize.x) {
 7          const uint inIndex = (r * pcs.nb01 + pcs.inAOff) + ind/NL * SIZE_OF_BLOCK;
 8          const mat4 result = dequantize_block(inIndex, ind%NL);
 9          for (uint j = 0; j < 4; ++j) {
10              for (uint k = 0; k < 4; ++k) {
11                  const uint outIndex = i * pcs.nb1/BYTES_FOR_TYPE + pcs.outOff + z;
12                  out_[outIndex] = result[j][k];
13                  ++z;
14              }
15          }
16      }
17  }