1#version 450
 2
 3#include "dequant_head.glsl"
 4
 5layout(local_size_x = 256, local_size_y = 1, local_size_z = 1) in;
 6
 7layout (binding = 0) readonly buffer A {block_iq2_s data_a[];};
 8layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
 9
10void main() {
11    // Each thread handles 1 subblock (32 values with 2 scales)
12    const uint ib = gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x / 8;
13
14    init_iq_shmem(gl_WorkGroupSize);
15
16    if (ib >= p.nel / 256) {
17        return;
18    }
19
20    const uint ib32 = gl_LocalInvocationID.x % 8;
21    const uint b_idx = 256 * ib + 32 * ib32;
22
23    const float d = float(data_a[ib].d);
24    const vec2 scale = vec2(data_a[ib].scales[ib32] & 0xf, data_a[ib].scales[ib32] >> 4);
25    const vec2 db = d * (0.5 + scale) * 0.25;
26
27    uint qh = data_a[ib].qh[ib32];
28    [[unroll]] for (uint l = 0; l < 4; ++l) {
29        uint qs = data_a[ib].qs[4 * ib32 + l];
30        const uint8_t sign = data_a[ib].qs[QUANT_K / 8 + 4 * ib32 + l];
31        qs |= (qh << (8 - 2 * l)) & 0x300;
32        const uvec2 grid = iq2s_grid[qs];
33        const u8vec4 grid0 = unpack8(grid.x);
34        const u8vec4 grid1 = unpack8(grid.y);
35        data_b[b_idx + 8 * l + 0] = D_TYPE(db[l/2] * grid0.x * ((sign & 1) != 0 ? -1.0 : 1.0));
36        data_b[b_idx + 8 * l + 1] = D_TYPE(db[l/2] * grid0.y * ((sign & 2) != 0 ? -1.0 : 1.0));
37        data_b[b_idx + 8 * l + 2] = D_TYPE(db[l/2] * grid0.z * ((sign & 4) != 0 ? -1.0 : 1.0));
38        data_b[b_idx + 8 * l + 3] = D_TYPE(db[l/2] * grid0.w * ((sign & 8) != 0 ? -1.0 : 1.0));
39        data_b[b_idx + 8 * l + 4] = D_TYPE(db[l/2] * grid1.x * ((sign & 16) != 0 ? -1.0 : 1.0));
40        data_b[b_idx + 8 * l + 5] = D_TYPE(db[l/2] * grid1.y * ((sign & 32) != 0 ? -1.0 : 1.0));
41        data_b[b_idx + 8 * l + 6] = D_TYPE(db[l/2] * grid1.z * ((sign & 64) != 0 ? -1.0 : 1.0));
42        data_b[b_idx + 8 * l + 7] = D_TYPE(db[l/2] * grid1.w * ((sign & 128) != 0 ? -1.0 : 1.0));
43    }
44}