1  #version 450
 2  
 3  #include "mul_mat_vec_base.comp"
 4  
 5  layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
 6  
 7  shared FLOAT_TYPE tmp[32];
 8  
 9  void main() {
10      const uint row = gl_WorkGroupID.x;
11  
12      uint a_offset, b_offset, d_offset;
13      get_offsets(a_offset, b_offset, d_offset);
14  
15      const uint num_blocks_per_row = p.ncols / QUANT_K;
16      const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
17  
18      const uint tid = gl_LocalInvocationID.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
19      const uint ix  = gl_LocalInvocationID.x%K_QUANTS_PER_ITERATION;  // 0 or 0, 1
20  
21      const uint step = 16/K_QUANTS_PER_ITERATION;            // 16 or 8
22  
23      const uint v_im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
24      const uint v_in = tid - step*v_im;                      // 0...15 or 0...7
25  
26      const uint8_t m = uint8_t(1 << (4 * v_im));
27  
28      const uint l0 = K_QUANTS_PER_ITERATION*v_in;            // 0...15
29      const uint q_offset = 32*v_im + l0;
30      const uint y_offset = 128*v_im + l0;
31  
32      tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp
33  
34      const uint s_shift = 4 * v_im;
35  
36      [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
37          const uint y_idx = i * QUANT_K + y_offset;
38  
39          const FLOAT_TYPE d = FLOAT_TYPE(data_a[ib0 + i].d);
40  
41          FLOAT_TYPE sum = FLOAT_TYPE(0.0);
42          for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
43              sum += FLOAT_TYPE(data_b[b_offset + y_idx + l +  0]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[0] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 8] >> (s_shift + 0) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 0)) != 0) ? 0 : 4))
44                   + FLOAT_TYPE(data_b[b_offset + y_idx + l + 32]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[2] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[10] >> (s_shift + 0) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 1)) != 0) ? 0 : 4))
45                   + FLOAT_TYPE(data_b[b_offset + y_idx + l + 64]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[4] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 8] >> (s_shift + 2) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 2)) != 0) ? 0 : 4))
46                   + FLOAT_TYPE(data_b[b_offset + y_idx + l + 96]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[6] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[10] >> (s_shift + 2) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l   ] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l   ] & (m << 3)) != 0) ? 0 : 4))
47                   + FLOAT_TYPE(data_b[b_offset + y_idx + l + 16]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[1] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 9] >> (s_shift + 0) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16]     ) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 0)) != 0) ? 0 : 4))
48                   + FLOAT_TYPE(data_b[b_offset + y_idx + l + 48]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[3] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[11] >> (s_shift + 0) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 2) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 1)) != 0) ? 0 : 4))
49                   + FLOAT_TYPE(data_b[b_offset + y_idx + l + 80]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[5] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[ 9] >> (s_shift + 2) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 4) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 2)) != 0) ? 0 : 4))
50                   + FLOAT_TYPE(data_b[b_offset + y_idx + l +112]) * FLOAT_TYPE(int8_t(((data_a[ib0 + i].scales[7] >> s_shift) & 0xF) | ((data_a[ib0 + i].scales[11] >> (s_shift + 2) & 0x3) << 4)) - 32) * FLOAT_TYPE(((data_a[ib0 + i].qs[q_offset + l+16] >> 6) & 3) - (((data_a[ib0 + i].hmask[l0 + l+16] & (m << 3)) != 0) ? 0 : 4));
51          }
52          tmp[16 * ix + tid] += d * sum;
53      }
54  
55      // sum up partial sums and write back result
56      barrier();
57      [[unroll]] for (uint s = 16; s > 0; s >>= 1) {
58          if (tid < s) {
59              tmp[tid] += tmp[tid + s];
60          }
61          barrier();
62      }
63      if (tid == 0) {
64          data_d[d_offset + row] = D_TYPE(tmp[0]);
65      }
66  }