// Multi-query flash vision attention Q = 16. Doubles queries-per-workgroup
// over Q=8 to amortise K/V loads and reduction barriers further.
//
// Workgroup storage:
// q_shared (16 × 64) 4 KB
// kv_tile (32 × 64) 8 KB
// tile_scores (16 × WG) 4 KB
// rbuf, sum_buf 0.5 KB
// total ~16 KB (right at the WebGPU minimum)
//
// Per-thread state: 16 sets of (m, l, o) online-softmax accumulators.
//
// **NOT ROUTED BY DEFAULT** — Q=16 measured 1.75 s/iter on Radeon Pro 555 /
// Metal at the full 2304-patch vision shape, vs 1.26 s for Q=8. The
// 16 KB workgroup storage drops per-CU occupancy below the threshold
// where memory latency stays hidden. Kept around as a reference variant;
// the router stops at Q=8.
struct Params {
head_dim: u32,
n_heads: u32,
n_patches: u32,
_pad: u32,
}
@group(0) @binding(0) var<uniform> params: Params;
@group(0) @binding(1) var<storage, read> q: array<f32>;
@group(0) @binding(2) var<storage, read> k: array<f32>;
@group(0) @binding(3) var<storage, read> v: array<f32>;
@group(0) @binding(4) var<storage, read_write> out: array<f32>;
const WG: u32 = 64u;
const HEAD_DIM_MAX: u32 = 64u;
const TILE_T: u32 = 32u;
const Q_PER_WG: u32 = 16u;
var<workgroup> q_shared: array<f32, 1024>; // Q_PER_WG × HEAD_DIM_MAX
var<workgroup> kv_tile: array<f32, 2048>; // TILE_T × HEAD_DIM_MAX
var<workgroup> tile_scores: array<f32, 1024>; // Q_PER_WG × WG
var<workgroup> rbuf: array<f32, WG>;
var<workgroup> sum_buf: array<f32, WG>;
@compute @workgroup_size(64)
fn main(
@builtin(workgroup_id) wid: vec3<u32>,
@builtin(local_invocation_index) tid: u32,
) {
let qh: u32 = wid.y;
if (qh >= params.n_heads) { return; }
let head_dim: u32 = params.head_dim;
let n_patches: u32 = params.n_patches;
let n_heads: u32 = params.n_heads;
let bq_base: u32 = wid.x * Q_PER_WG;
let q_count: u32 = min(Q_PER_WG, n_patches - bq_base);
if (q_count == 0u) { return; }
for (var i: u32 = 0u; i < Q_PER_WG; i = i + 1u) {
let bq = bq_base + i;
if (bq < n_patches && tid < head_dim) {
let q_off = (bq * n_heads + qh) * head_dim + tid;
q_shared[i * head_dim + tid] = q[q_off];
}
}
workgroupBarrier();
var m_arr: array<f32, 16>;
var l_arr: array<f32, 16>;
var o_arr: array<f32, 16>;
for (var i: u32 = 0u; i < Q_PER_WG; i = i + 1u) {
m_arr[i] = -1.0e30;
l_arr[i] = 0.0;
o_arr[i] = 0.0;
}
let n_tiles = (n_patches + TILE_T - 1u) / TILE_T;
for (var tile: u32 = 0u; tile < n_tiles; tile = tile + 1u) {
let t0 = tile * TILE_T;
let tile_size = min(TILE_T, n_patches - t0);
let total_k = tile_size * head_dim;
var lk = tid;
loop {
if (lk >= total_k) { break; }
let t_local = lk / head_dim;
let d_local = lk % head_dim;
let g_off = ((t0 + t_local) * n_heads + qh) * head_dim + d_local;
kv_tile[lk] = k[g_off];
lk = lk + WG;
}
workgroupBarrier();
for (var q_idx: u32 = 0u; q_idx < Q_PER_WG; q_idx = q_idx + 1u) {
if (q_idx >= q_count) { break; }
var s_t: f32 = -1.0e30;
if (tid < tile_size) {
var sum: f32 = 0.0;
let row_off = tid * head_dim;
let q_row_off = q_idx * head_dim;
let n_vec = head_dim / 4u;
for (var dv: u32 = 0u; dv < n_vec; dv = dv + 1u) {
let dv4 = dv * 4u;
let qv = vec4<f32>(
q_shared[q_row_off + dv4],
q_shared[q_row_off + dv4 + 1u],
q_shared[q_row_off + dv4 + 2u],
q_shared[q_row_off + dv4 + 3u],
);
let kv = vec4<f32>(
kv_tile[row_off + dv4],
kv_tile[row_off + dv4 + 1u],
kv_tile[row_off + dv4 + 2u],
kv_tile[row_off + dv4 + 3u],
);
sum = sum + dot(qv, kv);
}
for (var d: u32 = n_vec * 4u; d < head_dim; d = d + 1u) {
sum = sum + q_shared[q_row_off + d] * kv_tile[row_off + d];
}
s_t = sum;
}
rbuf[tid] = s_t;
sum_buf[tid] = select(0.0, 1.0, tid < tile_size);
workgroupBarrier();
var stride: u32 = WG / 2u;
loop {
if (stride == 0u) { break; }
if (tid < stride) {
let m_a = rbuf[tid];
let m_b = rbuf[tid + stride];
let l_a = sum_buf[tid];
let l_b = sum_buf[tid + stride];
let m_n = max(m_a, m_b);
rbuf[tid] = m_n;
sum_buf[tid] = l_a * exp(m_a - m_n) + l_b * exp(m_b - m_n);
}
workgroupBarrier();
stride = stride / 2u;
}
let tile_m = rbuf[0];
let tile_l = sum_buf[0];
let m_cur = m_arr[q_idx];
let l_cur = l_arr[q_idx];
let o_cur = o_arr[q_idx];
let m_new = max(m_cur, tile_m);
let alpha = exp(m_cur - m_new);
var p_t: f32 = 0.0;
if (tid < tile_size) {
p_t = exp(s_t - m_new);
}
tile_scores[q_idx * WG + tid] = p_t;
m_arr[q_idx] = m_new;
l_arr[q_idx] = l_cur * alpha + tile_l * exp(tile_m - m_new);
o_arr[q_idx] = o_cur * alpha;
}
workgroupBarrier();
var lv = tid;
loop {
if (lv >= total_k) { break; }
let t_local = lv / head_dim;
let d_local = lv % head_dim;
let g_off = ((t0 + t_local) * n_heads + qh) * head_dim + d_local;
kv_tile[lv] = v[g_off];
lv = lv + WG;
}
workgroupBarrier();
if (tid < head_dim) {
for (var q_idx: u32 = 0u; q_idx < Q_PER_WG; q_idx = q_idx + 1u) {
if (q_idx >= q_count) { break; }
let s_off = q_idx * WG;
var contrib: f32 = 0.0;
let n_vec = tile_size / 4u;
for (var tv: u32 = 0u; tv < n_vec; tv = tv + 1u) {
let t0_l = tv * 4u;
let sv = vec4<f32>(
tile_scores[s_off + t0_l], tile_scores[s_off + t0_l + 1u],
tile_scores[s_off + t0_l + 2u], tile_scores[s_off + t0_l + 3u],
);
let vv = vec4<f32>(
kv_tile[t0_l * head_dim + tid],
kv_tile[(t0_l + 1u) * head_dim + tid],
kv_tile[(t0_l + 2u) * head_dim + tid],
kv_tile[(t0_l + 3u) * head_dim + tid],
);
contrib = contrib + dot(sv, vv);
}
for (var t_local: u32 = n_vec * 4u; t_local < tile_size; t_local = t_local + 1u) {
contrib = contrib + tile_scores[s_off + t_local] * kv_tile[t_local * head_dim + tid];
}
o_arr[q_idx] = o_arr[q_idx] + contrib;
}
}
workgroupBarrier();
}
if (tid < head_dim) {
for (var q_idx: u32 = 0u; q_idx < Q_PER_WG; q_idx = q_idx + 1u) {
if (q_idx >= q_count) { break; }
let bq = bq_base + q_idx;
let out_off = (bq * n_heads + qh) * head_dim + tid;
out[out_off] = o_arr[q_idx] / l_arr[q_idx];
}
}
}