rublas 0.2.8

Linear algebra and matrix multiplication for Ruda.
Documentation 
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
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244

use ruda_core::tensor::{DType, Shape};
use ruda_kernel::dsl as kernel_dsl;
use ruda_kernel::dsl::prelude::*;
use ruda_kernel::tensor::{RudaTensor, allocation::empty_device_contiguous_dtype, permutation::swap_dims};

// Split K keeps low-rank FP32 projections parallel even with a tiny M or N.
// No TF32 conversion or floating-point atomic accumulation is used.
#[allow(unsafe_code)]
pub(super) fn try_launch<R: Runtime>(a: &RudaTensor<R>, b: &RudaTensor<R>, out: &RudaTensor<R>) -> bool {
    if a.dtype != DType::F32 || b.dtype != DType::F32 || out.dtype != DType::F32
        || a.qparams.is_some() || b.qparams.is_some() || out.qparams.is_some()
        || a.meta.num_dims() != 2 || b.meta.num_dims() != 2 || out.meta.num_dims() != 2
        || !a.client.same_execution_queue(&b.client) || !a.client.same_execution_queue(&out.client)
    { return false; }
    let (m, k, n) = (a.meta.shape()[0], a.meta.shape()[1], b.meta.shape()[1]);
    let hw = &a.client.properties().hardware;
    if k == 16 && m > 0 && m % 16 == 0 && n >= 128 && n % 128 == 0
        && a.meta.strides()[..] == [16, 1] && b.meta.shape()[0] == 16
        && b.meta.strides()[..] == [1, 16] && out.meta.shape()[..] == [m, n]
        && out.meta.strides()[..] == [n, 1]
        && b.handle.offset_start.unwrap_or(0) % 16 == 0
        && m.checked_mul(n).is_some_and(|v| v <= u32::MAX as usize)
        && n <= u32::MAX as usize / 16 && m <= u32::MAX as usize / 16
        && hw.plane_size_min == 32 && hw.plane_size_max == 32
        && n / 128 <= hw.max_ruda_count.0 as usize && m / 16 <= hw.max_ruda_count.1 as usize
    {
        unsafe { rank16::launch_unchecked::<R>(&a.client, RudaCount::Static((n / 128) as u32, (m / 16) as u32, 1),
            RudaDim::new_1d(128), a.clone().into_array_arg(), b.clone().into_array_arg(),
            out.clone().into_array_arg(), n); }
        return true;
    }
    if m > 0 && m <= 32 && n > 32 && k >= 512 && b.meta.shape()[0] == k
        && out.meta.shape()[..] == [m, n]
    {
        return try_launch(&swap_dims(b.clone(), 0, 1), &swap_dims(a.clone(), 0, 1),
                          &swap_dims(out.clone(), 0, 1));
    }
    if m == 0 || k < 512 || n == 0 || n > 32 || b.meta.shape()[0] != k
        || out.meta.shape()[..] != [m, n] || (a.meta.strides()[1] != 1 && a.meta.strides()[0] != 1)
        || !((out.meta.strides()[1] == 1 && (m == 1 || out.meta.strides()[0] == n))
            || (out.meta.strides()[0] == 1 && (n == 1 || out.meta.strides()[1] == m)))
        || hw.plane_size_min != 32 || hw.plane_size_max != 32
    { return false; }
    let column_major = b.meta.strides()[0] == 1;
    if !column_major && b.meta.strides()[1] != 1 { return false; }
    for tensor in [a, b, out] {
        let span = tensor.meta.shape().iter().zip(tensor.meta.strides().iter()).try_fold(1usize,
            |sum, (&dim, &stride)| sum.checked_add((dim - 1).checked_mul(stride)?));
        if span.is_none_or(|span| span > u32::MAX as usize) { return false; }
    }
    let tile = if column_major { 1024 } else { 512 };
    let splits = k.div_ceil(tile);
    let Some(count) = m.checked_mul(n).and_then(|v| v.checked_mul(splits)) else { return false; };
    if count > 16 * 1024 * 1024 { return false; }
    let warps = if column_major { count } else { m * splits };
    if warps.div_ceil(4) > hw.max_ruda_count.0 as usize { return false; }
    let partial = empty_device_contiguous_dtype(a.client.clone(), a.device.clone(), Shape::from([count]), DType::F32);
    let grid = RudaCount::Static(warps.div_ceil(4) as u32, 1, 1);
    if column_major && n == 16 && m % 4 == 0 && a.meta.strides()[1] == 1 {
        unsafe { columns_reuse::launch_unchecked::<R>(&a.client,
            RudaCount::Static((m / 4 * 2 * splits) as u32, 1, 1), RudaDim::new_1d(128),
            a.clone().into_tensor_arg(), b.clone().into_tensor_arg(), partial.clone().into_array_arg(),
            m, k, splits, a.meta.strides()[0], b.meta.strides()[1]); }
    } else if column_major {
        unsafe { columns::launch_unchecked::<R>(&a.client, grid, RudaDim::new_1d(128),
            a.clone().into_tensor_arg(), b.clone().into_tensor_arg(), partial.clone().into_array_arg(),
            m, k, n, splits, a.meta.strides()[0], a.meta.strides()[1], b.meta.strides()[1]); }
    } else if n == 16 && m % 16 == 0 && k % 512 == 0 {
        unsafe { rows_transposed::launch_unchecked::<R>(&a.client,
            RudaCount::Static((m / 16 * splits) as u32, 1, 1), RudaDim::new_1d(128),
            a.clone().into_tensor_arg(), b.clone().into_tensor_arg(), partial.clone().into_array_arg(),
            m, splits, a.meta.strides()[0], a.meta.strides()[1], b.meta.strides()[0]); }
    } else {
        unsafe { rows::launch_unchecked::<R>(&a.client, grid, RudaDim::new_1d(128),
            a.clone().into_tensor_arg(), b.clone().into_tensor_arg(), partial.clone().into_array_arg(),
            m, k, n, splits, n.next_power_of_two(), a.meta.strides()[0], a.meta.strides()[1], b.meta.strides()[0]); }
    }
    reduce::launch::<R>(&a.client, RudaCount::Static((m * n).div_ceil(128) as u32, 1, 1),
        RudaDim::new_1d(128), partial.into_array_arg(), out.clone().into_tensor_arg(), splits as u32);
    true
}

#[ruda(launch_unchecked)]
fn columns(a: &Tensor<f32>, b: &Tensor<f32>, partial: &mut Array<f32>,
    #[comptime] m: usize, #[comptime] k: usize, #[comptime] n: usize, #[comptime] splits: usize,
    #[comptime] as0: usize, #[comptime] as1: usize, #[comptime] bs1: usize) {
    let warp = RUDA_POS_X as usize * 4 + UNIT_POS as usize / 32;
    let lane = UNIT_POS as usize % 32;
    let output = warp / splits as usize;
    let split = warp % splits as usize;
    if output < m as usize * n as usize {
        let row = output / n as usize;
        let col = output % n as usize;
        let mut acc = 0.0f32;
        for offset in 0usize..32 {
            let inner = split * 1024 + offset * 32 + lane;
            if inner < k as usize {
                acc = fma(a[row * as0 + inner * as1], b[inner + col * bs1], acc);
            }
        }
        let total = plane_sum(acc);
        if lane == 0 { partial[split * m as usize * n as usize + output] = total; }
    }
}

#[ruda(launch_unchecked)]
fn columns_reuse(a: &Tensor<f32>, b: &Tensor<f32>, partial: &mut Array<f32>,
    #[comptime] m: usize, #[comptime] k: usize, #[comptime] splits: usize,
    #[comptime] as0: usize, #[comptime] bs1: usize) {
    let tile = RUDA_POS_X as usize / 2;
    let row = tile / splits * 4 + UNIT_POS as usize / 32;
    let split = tile % splits;
    let col = RUDA_POS_X as usize % 2 * 8;
    let lane = UNIT_POS as usize % 32;
    let mut acc = Array::<f32>::new(8usize);
    #[unroll]
    for j in 0usize..8 { acc[j] = 0.0f32; }
    for offset in 0usize..32 {
        let inner = split * 1024 + offset * 32 + lane;
        if inner < k {
            let value = a[row * as0 + inner];
            #[unroll]
            for j in 0usize..8 { acc[j] = fma(value, b[inner + (col + j) * bs1], acc[j]); }
        }
    }
    #[unroll]
    for j in 0usize..8 {
        let sum = plane_sum(acc[j]);
        if lane == 0 { partial[(split * m + row) * 16 + col + j] = sum; }
    }
}

#[ruda(launch_unchecked)]
fn rows(a: &Tensor<f32>, b: &Tensor<f32>, partial: &mut Array<f32>,
    #[comptime] m: usize, #[comptime] k: usize, #[comptime] n: usize, #[comptime] splits: usize,
    #[comptime] cols: usize, #[comptime] as0: usize, #[comptime] as1: usize, #[comptime] bs0: usize) {
    let warp = RUDA_POS_X as usize * 4 + UNIT_POS as usize / 32;
    let lane = UNIT_POS as usize % 32;
    let row = warp / splits as usize;
    let split = warp % splits as usize;
    let col = lane % cols;
    if row < m as usize {
        let mut acc = 0.0f32;
        for offset in 0usize..(512 / (32 / cols)) {
            let inner = split * 512 + offset * (32 / cols) + lane / cols;
            if inner < k as usize && col < n as usize {
                acc = fma(a[row * as0 + inner * as1], b[inner * bs0 + col], acc);
            }
        }
        if comptime!(cols <= 16) { acc += plane_shuffle_xor(acc, 16u32); }
        if comptime!(cols <= 8) { acc += plane_shuffle_xor(acc, 8u32); }
        if comptime!(cols <= 4) { acc += plane_shuffle_xor(acc, 4u32); }
        if comptime!(cols <= 2) { acc += plane_shuffle_xor(acc, 2u32); }
        if comptime!(cols == 1) { acc += plane_shuffle_xor(acc, 1u32); }
        if lane < cols && col < n as usize {
            partial[(split * m as usize + row) * n as usize + col] = acc;
        }
    }
}

#[ruda(launch_unchecked)]
fn rows_transposed(a: &Tensor<f32>, b: &Tensor<f32>, partial: &mut Array<f32>,
    #[comptime] m: usize, #[comptime] splits: usize, #[comptime] as0: usize, #[comptime] as1: usize, #[comptime] bs0: usize) {
    let thread = UNIT_POS as usize;
    let split = RUDA_POS_X as usize % splits;
    let tile_row = RUDA_POS_X as usize / splits * 16;
    let local_row = thread / 32 * 4 + thread % 4;
    let row = tile_row + local_row;
    let parity = thread / 4 % 2;
    let col = thread % 32 / 8 * 4;
    let stride = comptime!(if as0 == 1 { 16usize } else { 17usize });
    let mut left = SharedMemory::<f32>::new_aligned(32usize * stride, 32usize);
    let mut right = SharedMemory::<f32>::new_aligned(32usize * 16usize, 32usize);
    let mut acc = Array::<f32>::new(4usize);
    #[unroll]
    for j in 0usize..4 { acc[j] = 0.0f32; }
    // Keep the original even/odd FMA chains and split reduction order exactly.
    for stage in 0usize..16 {
        #[unroll]
        for i in 0usize..4 {
            let p = thread + i * 128;
            if comptime!(as0 == 1) {
                left[(p / 16) * stride + p % 16] = a[(tile_row + p % 16) * as0 + (split * 512 + stage * 32 + p / 16) * as1];
            } else {
                left[(p % 32) * stride + p / 32] = a[(tile_row + p / 32) * as0 + (split * 512 + stage * 32 + p % 32) * as1];
            }
            right[p] = b[(split * 512 + stage * 32 + p / 16) * bs0 + p % 16];
        }
        sync_ruda();
        for group in 0usize..4 {
            #[unroll]
            for i in 0usize..4 {
                let inner = group * 8 + i * 2 + parity;
                let value = left[inner * stride + local_row];
                #[unroll]
                for j in 0usize..4 { acc[j] = fma(value, right[inner * 16 + col + j], acc[j]); }
            }
        }
        sync_ruda();
    }
    #[unroll]
    for j in 0usize..4 {
        let sum = acc[j] + plane_shuffle_xor(acc[j], 4u32);
        if parity == 0 { partial[(split * m + row) * 16 + col + j] = sum; }
    }
}

#[ruda(launch_unchecked)]
fn rank16(a: &Array<f32>, b: &Array<Vector<f32, Const<4>>>, out: &mut Array<f32>,
    #[comptime] n: usize) {
    let thread = UNIT_POS as usize;
    let row = RUDA_POS_Y as usize * 16;
    let col = RUDA_POS_X as usize * 128 + thread;
    let mut shared = SharedMemory::<f32>::new(256usize);
    #[unroll]
    for i in 0usize..2 { shared[thread + i * 128] = a[row * 16 + thread + i * 128]; }
    sync_ruda();
    let mut acc = Array::<f32>::new(16usize);
    #[unroll]
    for i in 0usize..16 { acc[i] = 0.0f32; }
    #[unroll]
    for group in 0usize..4 {
        let value = b[col * 4 + group];
        #[unroll]
        for k in 0usize..4 {
            #[unroll]
            for i in 0usize..16 { acc[i] = fma(shared[i * 16 + group * 4 + k], value[k], acc[i]); }
        }
    }
    #[unroll]
    for i in 0usize..16 { out[(row + i) * n + col] = acc[i]; }
}

#[ruda(launch)]
fn reduce(partial: &Array<f32>, out: &mut Tensor<f32>, splits: u32) {
    let pos = ABSOLUTE_POS as usize;
    if pos < out.len() {
        let mut sum = 0.0f32;
        for split in 0usize..splits as usize { sum += partial[split * out.len() + pos]; }
        let row = pos / out.shape(1);
        let col = pos % out.shape(1);
        out[row * out.stride(0) + col * out.stride(1)] = sum;
    }
}