struct View {
shape: vec4<u32>,
stride: vec4<u32>,
offset: vec4<u32>,
};
@group(0) @binding(0) var<uniform> shape: vec4<u32>; // [C, R, B]
@group(0) @binding(1) var<uniform> source: View; // [R, T, B]
@group(0) @binding(2) var<uniform> destination: View; // [R, T, B]
@group(0) @binding(3) var<uniform> quant: array<vec4<f32>, 4u>;
@group(0) @binding(4) var<storage, read> matrix: array<u32>; // (B, R, C)
@group(0) @binding(5) var<storage, read> absmax: array<u32>;
#ifdef IN_FP16
@group(0) @binding(6) var<storage, read> input: array<vec4<u32>>; // (B, T, C)
#else
@group(0) @binding(6) var<storage, read> input: array<mat2x4<f32>>; // (B, T, C)
#endif
#ifdef OUT_FP16
@group(0) @binding(7) var<storage, read_write> output: array<vec2<u32>>; // (B, T, R)
#else
@group(0) @binding(7) var<storage, read_write> output: array<vec4<f32>>; // (B, T, R)
#endif
const NF4_BLOCK_STEP: u32 = NF4_BLOCK_SIZE / 8u;
const NUM_SUBGROUPS: u32 = BLOCK_SIZE / MIN_SUBGROUP_SIZE;
var<workgroup> sketch: array<vec4<f32>, NUM_SUBGROUPS>;
var<workgroup> q: array<vec4<f32>, 4u>;
// ACTIVATION_DEFINE
fn compute_index(view: View, batch: u32, token: u32, index: u32, step: u32) -> u32 {
let stride = view.stride.x >> step;
let offset = vec3<u32>(view.offset.zy, view.offset.x >> step);
return dot(vec3<u32>(batch, token, index) + offset, vec3<u32>(view.stride.y * stride, stride, 1u));
}
fn pack4x16float(x: vec4<f32>) -> vec2<u32> {
return vec2<u32>(pack2x16float(x.xy), pack2x16float(x.zw));
}
fn unpack4x16float(x: vec2<u32>) -> vec4<f32> {
return vec4<f32>(unpack2x16float(x.x), unpack2x16float(x.y));
}
fn unpack_absmax(index: u32) -> f32 {
let i = index / NF4_BLOCK_STEP; // 1 block of absmax: NF4_BLOCK_SIZE / 8u entries in matrix
return unpack2x16float(absmax[i >> 1u])[i & 1u];
}
fn unpack_matrix_0(v: u32) -> vec4<f32> {
let i = vec4<u32>(
(v & 0x0000000fu),
(v & 0x000000f0u) >> 4u,
(v & 0x00000f00u) >> 8u,
(v & 0x0000f000u) >> 12u,
);
return vec4<f32>(
q[i.x >> 2u][i.x & 3u],
q[i.y >> 2u][i.y & 3u],
q[i.z >> 2u][i.z & 3u],
q[i.w >> 2u][i.w & 3u],
);
}
fn unpack_matrix_1(v: u32) -> vec4<f32> {
let i = vec4<u32>(
(v & 0x000f0000u) >> 16u,
(v & 0x00f00000u) >> 20u,
(v & 0x0f000000u) >> 24u,
(v & 0xf0000000u) >> 28u,
);
return vec4<f32>(
q[i.x >> 2u][i.x & 3u],
q[i.y >> 2u][i.y & 3u],
q[i.z >> 2u][i.z & 3u],
q[i.w >> 2u][i.w & 3u],
);
}
fn reduce_sum(index: u32, stride: u32) {
if index < stride {
sketch[index] += sketch[index + stride];
}
workgroupBarrier();
}
@compute @workgroup_size(BLOCK_SIZE, 1, 1)
fn matmul(
@builtin(global_invocation_id) invocation_id: vec3<u32>,
@builtin(num_subgroups) num_subgroups: u32,
@builtin(subgroup_id) subgroup_id: u32,
@builtin(subgroup_size) subgroup_size: u32,
@builtin(subgroup_invocation_id) subgroup_invocation_id: u32
) {
let stride = source.stride.x / 8u;
let index = invocation_id.x % BLOCK_SIZE;
let channel = invocation_id.x / BLOCK_SIZE; // 1 channel: 4 rows in matrix
let token = invocation_id.y;
let batch = invocation_id.z;
let bb = compute_index(source, batch, token, 0u, 3u);
let cb = batch * shape.y * stride + channel * 4u * stride;
if index == 0u {
q = quant;
}
workgroupBarrier();
var local_sum = vec4<f32>(0.0);
for (var i = index; i < stride; i += BLOCK_SIZE) {
// read 8 elements from the input
let x = input[bb + i];
#ifdef SPARSE_INPUT
#ifdef IN_FP16
if all(unpack4x16float(x.xy) == vec4<f32>(0.0)) && all(unpack4x16float(x.zw) == vec4<f32>(0.0)) {
continue;
}
#else
if all(x[0] == vec4<f32>(0.0)) && all(x[1] == vec4<f32>(0.0)) {
continue;
}
#endif
#endif
// read 4 rows from the matrix, each with 4x2 unpacked floats, forming 2 4x4 sub-blocks
var ci = cb + i;
var v: vec4<u32>;
var a: vec4<f32>;
v[0] = matrix[ci]; a[0] = unpack_absmax(ci); ci += stride;
v[1] = matrix[ci]; a[1] = unpack_absmax(ci); ci += stride;
v[2] = matrix[ci]; a[2] = unpack_absmax(ci); ci += stride;
v[3] = matrix[ci]; a[3] = unpack_absmax(ci);
var m: mat4x4<f32>;
m[0] = unpack_matrix_0(v[0]);
m[1] = unpack_matrix_0(v[1]);
m[2] = unpack_matrix_0(v[2]);
m[3] = unpack_matrix_0(v[3]);
m = transpose(m);
#ifdef IN_FP16
local_sum = fma(m * unpack4x16float(x.xy), a, local_sum);
#else
local_sum = fma(m * x[0], a, local_sum);
#endif
m[0] = unpack_matrix_1(v[0]);
m[1] = unpack_matrix_1(v[1]);
m[2] = unpack_matrix_1(v[2]);
m[3] = unpack_matrix_1(v[3]);
m = transpose(m);
#ifdef IN_FP16
local_sum = fma(m * unpack4x16float(x.zw), a, local_sum);
#else
local_sum = fma(m * x[1], a, local_sum);
#endif
}
// for (var step = subgroup_size >> 1u; step > 0u; step >>= 1u) {
// local_sum += subgroupShuffleDown(local_sum, step);
// }
local_sum = subgroupAdd(local_sum);
if subgroup_invocation_id == 0u { sketch[subgroup_id] = local_sum; }
workgroupBarrier();
#ifdef SUBGROUP_SIZE_32_32
reduce_sum(index, 2u);
reduce_sum(index, 1u);
#else
#ifdef SUBGROUP_SIZE_32_64
if subgroup_size == 32u { reduce_sum(index, 2u); }
reduce_sum(index, 1u);
#else
#ifdef SUBGROUP_SIZE_64_64
reduce_sum(index, 1u);
#else
for (var step = num_subgroups >> 1u; step > 0u; step >>= 1u) {
if index < step {
sketch[index] += sketch[index + step];
}
workgroupBarrier();
}
#endif
#endif
#endif
if index == 0u {
let btc = compute_index(destination, batch, token, channel, 2u);
let out = ACT(sketch[0]);
#ifdef OUT_FP16
output[btc] = pack4x16float(out);
#else
output[btc] = out;
#endif
}
}