oarust 0.8.6

GPU-first semantic computing engine written in Rust
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58

// Mechanical adaptation of OA's tiled rank-two/rank-three last-axis transpose.
// Donor: source/cpp/lib/oa/core/shader/compute/matrix/transposeBatched.slang

import attributes;
import storage;

static const uint TILE = 32;
static const uint TILE_PAD = TILE + 1;

groupshared float tile[TILE * TILE_PAD];

struct PushConstants {
	uint input_index;
	uint output_index;
	uint batch_size;
	uint rows;
	uint columns;
};

[[vk::push_constant]] PushConstants push;
[[vk::binding(0, 0)]] RWByteAddressBuffer storage_buffers[];

[kernel_name("transpose")]
[domain("matrix")]
[variant("generic")]
[dtype("f32")]
[status("experimental")]
[shader("compute")]
[numthreads(TILE, TILE, 1)]
void main(uint3 group_id : SV_GroupID, uint3 group_thread_id : SV_GroupThreadID) {
	uint batch = group_id.z;
	if (batch >= push.batch_size) {
		return;
	}
	uint input_row = group_id.y * TILE + group_thread_id.y;
	uint input_column = group_id.x * TILE + group_thread_id.x;
	uint input_base = batch * push.rows * push.columns;
	if (input_row < push.rows && input_column < push.columns) {
		tile[group_thread_id.y * TILE_PAD + group_thread_id.x] = load_f32(
			storage_buffers[push.input_index],
			input_base + input_row * push.columns + input_column
		);
	} else {
		tile[group_thread_id.y * TILE_PAD + group_thread_id.x] = 0.0f;
	}
	GroupMemoryBarrierWithGroupSync();

	uint output_row = group_id.x * TILE + group_thread_id.y;
	uint output_column = group_id.y * TILE + group_thread_id.x;
	if (output_row < push.columns && output_column < push.rows) {
		uint output_base = batch * push.columns * push.rows;
		store_f32(
			storage_buffers[push.output_index],
			output_base + output_row * push.rows + output_column,
			tile[group_thread_id.x * TILE_PAD + group_thread_id.y]
		);
	}
}