1@group(0) @binding(0)
2var<storage, read_write> src: array<f32>;
3
4@group(0) @binding(1)
5var<storage, read_write> dst: array<f32>;
6
7struct Params {
8 offset_src: u32, // in elements
9 offset_dst: u32, // in elements
10 ne0: u32,
11};
12
13@group(0) @binding(2)
14var<uniform> params: Params;
15
16var<workgroup> shared_sum: array<f32, WG_SIZE>;
17
18@compute @workgroup_size(WG_SIZE)
19fn main(@builtin(workgroup_id) wid: vec3<u32>,
20 @builtin(local_invocation_id) lid: vec3<u32>) {
21 let row_idx = params.offset_src + wid.x * params.ne0;
22 let elems = (params.ne0 + WG_SIZE - 1) / WG_SIZE;
23 var local_sum: f32 = 0.0;
24 for (var col = lid.x * elems; col < (lid.x + 1) * elems && col < params.ne0; col ++) {
25 local_sum += src[row_idx + col];
26 }
27 shared_sum[lid.x] = local_sum;
28 workgroupBarrier();
29
30 // upsweep
31 var offset = 1u;
32 while (offset < WG_SIZE) {
33 let idx = (lid.x + 1) * offset * 2 - 1;
34 if (idx < WG_SIZE) {
35 shared_sum[idx] = shared_sum[idx] + shared_sum[idx - offset];
36 }
37 workgroupBarrier();
38 offset <<= 1;
39 }
40
41 // set last to 0 for exclusive sum
42 if (lid.x == 0) {
43 shared_sum[WG_SIZE - 1] = 0.0;
44 }
45 workgroupBarrier();
46
47 // downsweep
48 offset = WG_SIZE >> 1;
49 while (offset > 0) {
50 let idx = (lid.x + 1) * offset * 2 - 1;
51 if (idx < WG_SIZE) {
52 let t = shared_sum[idx - offset];
53 shared_sum[idx - offset] = shared_sum[idx];
54 shared_sum[idx] = shared_sum[idx] + t;
55 }
56 workgroupBarrier();
57 offset = offset >> 1;
58 }
59
60 // shared_sum[lid] is exclusive prefix sum up to this thread.
61 var running_sum = shared_sum[lid.x];
62 for (var col = lid.x * elems; col < (lid.x + 1) * elems && col < params.ne0; col ++) {
63 running_sum += src[row_idx + col];
64 dst[params.offset_dst + wid.x * params.ne0 + col] = running_sum;
65 }
66}