1#version 450
  2#extension GL_EXT_shader_explicit_arithmetic_types_int32 : require
  3
  4#include "mul_mat_vec_base.glsl"
  5
  6layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
  7
  8FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
  9
 10void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32, const uint i,
 11                               const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
 12    // Compute starting index in matrix B for this superblock
 13    const uint y_idx = i * QUANT_K + 32 * ib32;
 14    uint ibi = a_offset + first_row * num_blocks_per_row + i;
 15
 16    // Precompute indices for quantization lookup tables
 17    const uint qh_base = 2 * ib32;
 18    const uint qs_base = 4 * ib32;
 19    const uint sc_index = ib32 / 2;
 20    const uint sc_shift = 6 * (ib32 & 1);
 21
 22    // Loop over rows in the superblock
 23    [[unroll]] for (uint n = 0; n < num_rows; ++n) {
 24        // Load per-block scales and shift for quantization
 25        const uint16_t[4] scales = data_a[ibi].scales;
 26        const u16vec4 s = u16vec4(scales[0], scales[1], scales[2], scales[3]) >> 12;
 27        const float d = float(unpackHalf2x16(s.x | (s.y << 4) | (s.z << 8) | (s.w << 12)).x);
 28        const uint sc = data_a[ibi].scales[sc_index] >> sc_shift;
 29
 30        // Temporary caches for decoding
 31        FLOAT_TYPE dl_cache[4];
 32        uint16_t gvf_cache[4];
 33        float delta_cache[4];
 34
 35        // Precompute the multiplier and lookup values for 4 sub-blocks
 36        [[unroll]] for (uint l = 0; l < 4; ++l) {
 37            dl_cache[l] = FLOAT_TYPE(d * (2 * bitfieldExtract(sc, 3 * int(l / 2), 3) + 1));
 38            const uint qh = data_a[ibi].qh[qh_base + l / 2] >> (4 * (l & 1));
 39            const uint qs = data_a[ibi].qs[qs_base + l];
 40            gvf_cache[l] = iq1s_grid[qs | ((qh & 7) << 8)];
 41            delta_cache[l] = ((qh & 8) != 0) ? -IQ1M_DELTA : IQ1M_DELTA;
 42        }
 43
 44        // Loop over columns of the output
 45        [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
 46            // Compute base index for matrix B
 47            const uint base_b_idx = (j * p.batch_stride_b + b_offset + y_idx) / 4;
 48            vec4 b_vals[8];
 49
 50            // Load 8 vec4 values from matrix B
 51            [[unroll]] for (int idx = 0; idx < 8; ++idx) {
 52                b_vals[idx] = vec4(data_b_v4[base_b_idx + idx]);
 53            }
 54
 55            FLOAT_TYPE col_sum = FLOAT_TYPE(0.0);
 56
 57            // Loop over sub-blocks
 58            [[unroll]] for (uint l = 0; l < 4; ++l) {
 59                const uint16_t grid = gvf_cache[l];
 60                const float dl = dl_cache[l];
 61
 62                // Decode 8 2-bit fbits from gvf_cache
 63                float f0 = float(bitfieldExtract(grid, 0, 2));
 64                float f1 = float(bitfieldExtract(grid, 2, 2));
 65                float f2 = float(bitfieldExtract(grid, 4, 2));
 66                float f3 = float(bitfieldExtract(grid, 6, 2));
 67                float f4 = float(bitfieldExtract(grid, 8, 2));
 68                float f5 = float(bitfieldExtract(grid, 10, 2));
 69                float f6 = float(bitfieldExtract(grid, 12, 2));
 70                float f7 = float(bitfieldExtract(grid, 14, 2));
 71
 72                // Pack into vec4 for vectorized FMA
 73                const vec4 fbits_v0 = vec4(f0, f1, f2, f3);
 74                const vec4 fbits_v1 = vec4(f4, f5, f6, f7);
 75                const vec4 delta_v = vec4(delta_cache[l]);
 76
 77                // Vectorized fused multiply-add
 78                vec4 sum_v = fma(b_vals[2*l + 0], fbits_v0 + delta_v, vec4(0.0));
 79                sum_v      = fma(b_vals[2*l + 1], fbits_v1 + delta_v, sum_v);
 80
 81                // Horizontal add to get scalar sum
 82                FLOAT_TYPE sum = sum_v.x + sum_v.y + sum_v.z + sum_v.w;
 83
 84                // Accumulate to column sum
 85                col_sum = fma(dl, sum, col_sum);
 86            }
 87            // Write result to temporary buffer
 88            temp[j][n] += col_sum;
 89        }
 90        ibi += num_blocks_per_row;
 91    }
 92}
 93
 94void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
 95    uint a_offset, b_offset, d_offset;
 96    get_offsets(a_offset, b_offset, d_offset);
 97
 98    const uint num_blocks_per_row = p.ncols / QUANT_K;
 99
100    // 8 threads are used to process each block
101    const uint blocks_per_wg = gl_WorkGroupSize.x/8;
102    const uint tid = gl_LocalInvocationID.x;
103    const uint itid = tid % 8;  // 0...7
104    const uint ix = tid / 8;
105
106    [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
107        [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
108            temp[j][i] = FLOAT_TYPE(0);
109        }
110    }
111
112    [[unroll]] for (uint i = ix; i < num_blocks_per_row; i += blocks_per_wg)
113        calc_superblock(a_offset, b_offset, itid, i, num_blocks_per_row, first_row, num_rows);
114
115    reduce_result(temp, d_offset, first_row, num_rows, tid);
116}
117
118void main() {
119    const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
120
121    init_iq_shmem(gl_WorkGroupSize);
122
123    // do NUM_ROWS at a time, unless there aren't enough remaining rows
124    if (first_row + NUM_ROWS <= p.stride_d) {
125        compute_outputs(first_row, NUM_ROWS);
126    } else {
127        if (first_row >= p.stride_d) {
128            return;
129        }
130        compute_outputs(first_row, p.stride_d - first_row);
131    }
132}