/ vulkan-shaders / mul_mat_vec_q4_k.comp
mul_mat_vec_q4_k.comp
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 = 8/K_QUANTS_PER_ITERATION; // 8 or 4 22 23 const uint il = tid/step; // 0...3 24 const uint ir = tid - step*il; // 0...7 or 0...3 25 const uint n = 2 * K_QUANTS_PER_ITERATION; // 2 or 4 26 27 const uint v_im = il / 2; // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224 28 const uint v_in = il % 2; 29 30 const uint l0 = n * (2 * ir + v_in); // 0...15 31 const uint q_offset = 32*v_im + l0; 32 const uint y_offset = 64*v_im + l0; 33 34 tmp[16 * ix + tid] = FLOAT_TYPE(0.0); // partial sum for thread in warp 35 36 [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) { 37 const uint y1_idx = i * QUANT_K + y_offset; 38 const uint y2_idx = y1_idx + 128; 39 40 const FLOAT_TYPE dall = FLOAT_TYPE(data_a[ib0 + i].d.x); 41 const FLOAT_TYPE dmin = FLOAT_TYPE(data_a[ib0 + i].d.y); 42 43 const uint8_t sc0 = uint8_t( data_a[ib0 + i].scales[v_im * 2 ] & 0x3f); 44 const uint8_t sc1 = uint8_t( data_a[ib0 + i].scales[v_im * 2 + 1] & 0x3f); 45 const uint8_t sc2 = uint8_t( data_a[ib0 + i].scales[v_im * 2 + 4] & 0x3f); 46 const uint8_t sc3 = uint8_t( data_a[ib0 + i].scales[v_im * 2 + 5] & 0x3f); 47 const uint8_t sc4 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 8] & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 ] & 0xc0) >> 2)); 48 const uint8_t sc5 = uint8_t(( data_a[ib0 + i].scales[v_im * 2 + 9] & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 1] & 0xc0) >> 2)); 49 const uint8_t sc6 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 8] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 4] & 0xc0) >> 2)); 50 const uint8_t sc7 = uint8_t(((data_a[ib0 + i].scales[v_im * 2 + 9] >> 4) & 0x0f) | ((data_a[ib0 + i].scales[v_im * 2 + 5] & 0xc0) >> 2)); 51 52 #if K_QUANTS_PER_ITERATION == 2 53 const uint8_t q4_0 = uint8_t(data_a[ib0 + i].qs[q_offset ] & 0xf); 54 const uint8_t q4_1 = uint8_t(data_a[ib0 + i].qs[q_offset + 1] & 0xf); 55 const uint8_t q4_2 = uint8_t(data_a[ib0 + i].qs[q_offset + 2] & 0xf); 56 const uint8_t q4_3 = uint8_t(data_a[ib0 + i].qs[q_offset + 3] & 0xf); 57 const uint8_t q4_4 = uint8_t(data_a[ib0 + i].qs[q_offset ] >> 4); 58 const uint8_t q4_5 = uint8_t(data_a[ib0 + i].qs[q_offset + 1] >> 4); 59 const uint8_t q4_6 = uint8_t(data_a[ib0 + i].qs[q_offset + 2] >> 4); 60 const uint8_t q4_7 = uint8_t(data_a[ib0 + i].qs[q_offset + 3] >> 4); 61 const uint8_t q4_8 = uint8_t(data_a[ib0 + i].qs[q_offset + 64] & 0xf); 62 const uint8_t q4_9 = uint8_t(data_a[ib0 + i].qs[q_offset + 65] & 0xf); 63 const uint8_t q4_10 = uint8_t(data_a[ib0 + i].qs[q_offset + 66] & 0xf); 64 const uint8_t q4_11 = uint8_t(data_a[ib0 + i].qs[q_offset + 67] & 0xf); 65 const uint8_t q4_12 = uint8_t(data_a[ib0 + i].qs[q_offset + 64] >> 4); 66 const uint8_t q4_13 = uint8_t(data_a[ib0 + i].qs[q_offset + 65] >> 4); 67 const uint8_t q4_14 = uint8_t(data_a[ib0 + i].qs[q_offset + 66] >> 4); 68 const uint8_t q4_15 = uint8_t(data_a[ib0 + i].qs[q_offset + 67] >> 4); 69 70 const FLOAT_TYPE sx = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y1_idx]) * q4_0 + FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) * q4_1 + FLOAT_TYPE(data_b[b_offset + y1_idx + 2]) * q4_2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 3]) * q4_3); 71 const FLOAT_TYPE sy = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * q4_4 + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * q4_5 + FLOAT_TYPE(data_b[b_offset + y1_idx + 34]) * q4_6 + FLOAT_TYPE(data_b[b_offset + y1_idx + 35]) * q4_7); 72 const FLOAT_TYPE sz = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y2_idx]) * q4_8 + FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) * q4_9 + FLOAT_TYPE(data_b[b_offset + y2_idx + 2]) * q4_10 + FLOAT_TYPE(data_b[b_offset + y2_idx + 3]) * q4_11); 73 const FLOAT_TYPE sw = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * q4_12 + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * q4_13 + FLOAT_TYPE(data_b[b_offset + y2_idx + 34]) * q4_14 + FLOAT_TYPE(data_b[b_offset + y2_idx + 35]) * q4_15); 74 const FLOAT_TYPE smin = FLOAT_TYPE( 75 FLOAT_TYPE(data_b[b_offset + y1_idx ]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx ]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * sc7 76 + FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * sc7 77 + FLOAT_TYPE(data_b[b_offset + y1_idx + 2]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 34]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx + 2]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 34]) * sc7 78 + FLOAT_TYPE(data_b[b_offset + y1_idx + 3]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 35]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx + 3]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 35]) * sc7 79 ); 80 tmp[16 * ix + tid] += FLOAT_TYPE(dall * (sx * sc0 + sy * sc1 + sz * sc4 + sw * sc5) - dmin * smin); 81 #else 82 const uint8_t q4_0 = uint8_t(data_a[ib0 + i].qs[q_offset ] & 0xf); 83 const uint8_t q4_1 = uint8_t(data_a[ib0 + i].qs[q_offset + 1] & 0xf); 84 const uint8_t q4_2 = uint8_t(data_a[ib0 + i].qs[q_offset ] >> 4); 85 const uint8_t q4_3 = uint8_t(data_a[ib0 + i].qs[q_offset + 1] >> 4); 86 const uint8_t q4_4 = uint8_t(data_a[ib0 + i].qs[q_offset + 64] & 0xf); 87 const uint8_t q4_5 = uint8_t(data_a[ib0 + i].qs[q_offset + 65] & 0xf); 88 const uint8_t q4_6 = uint8_t(data_a[ib0 + i].qs[q_offset + 64] >> 4); 89 const uint8_t q4_7 = uint8_t(data_a[ib0 + i].qs[q_offset + 65] >> 4); 90 91 const FLOAT_TYPE sx = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y1_idx ]) * q4_0 + FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) * q4_1); 92 const FLOAT_TYPE sy = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * q4_2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * q4_3); 93 const FLOAT_TYPE sz = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y2_idx ]) * q4_4 + FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) * q4_5); 94 const FLOAT_TYPE sw = FLOAT_TYPE(FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * q4_6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * q4_7); 95 const FLOAT_TYPE smin = FLOAT_TYPE( 96 FLOAT_TYPE(data_b[b_offset + y1_idx]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 32]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 32]) * sc7 97 + FLOAT_TYPE(data_b[b_offset + y1_idx + 1]) * sc2 + FLOAT_TYPE(data_b[b_offset + y1_idx + 33]) * sc3 + FLOAT_TYPE(data_b[b_offset + y2_idx + 1]) * sc6 + FLOAT_TYPE(data_b[b_offset + y2_idx + 33]) * sc7 98 ); 99 100 tmp[16 * ix + tid] += FLOAT_TYPE(dall * (sx * FLOAT_TYPE(data_a[ib0 + i].scales[v_im] & 0x3f) + sy * FLOAT_TYPE(data_a[ib0 + i].scales[v_im + 1] & 0x3f) + sz * FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 4] & 0x0f) | ((data_a[ib0 + i].scales[v_im] & 0xc0) >> 2)) + sw * FLOAT_TYPE((data_a[ib0 + i].scales[v_im + 5] & 0x0f) | ((data_a[ib0 + i].scales[v_im + 1] & 0xc0) >> 2))) - dmin * smin); 101 #endif 102 } 103 104 // sum up partial sums and write back result 105 barrier(); 106 [[unroll]] for (uint s = 16; s > 0; s >>= 1) { 107 if (tid < s) { 108 tmp[tid] += tmp[tid + s]; 109 } 110 barrier(); 111 } 112 if (tid == 0) { 113 data_d[d_offset + row] = D_TYPE(tmp[0]); 114 } 115 }