fastnn 0.3.3

A robust deep learning library in Rust with CUDA kernels
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
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262

//! Raw `extern "C"` declarations for `cuda/kernels.cu`. Only [`super::kernels`] calls these.
//!
//! Every function returns 0 on success.

#![allow(dead_code)]

use std::ffi::c_int;

unsafe extern "C" {
    pub fn fastnn_cuda_init(device_id: c_int) -> c_int;
    pub fn fastnn_cuda_device_count() -> c_int;
    pub fn fastnn_cuda_synchronize() -> c_int;
    pub fn fastnn_cuda_get_memory_info(free: *mut usize, total: *mut usize);

    pub fn fastnn_cuda_malloc(ptr: *mut *mut f32, bytes: usize) -> c_int;
    pub fn fastnn_cuda_free(ptr: *mut f32) -> c_int;
    pub fn fastnn_cuda_memcpy_h2d(dst: *mut f32, src: *const f32, bytes: usize) -> c_int;
    pub fn fastnn_cuda_memcpy_d2h(dst: *mut f32, src: *const f32, bytes: usize) -> c_int;
    pub fn fastnn_cuda_memcpy_d2d(dst: *mut f32, src: *const f32, bytes: usize) -> c_int;
    pub fn fastnn_cuda_memset(ptr: *mut f32, value: c_int, bytes: usize) -> c_int;

    pub fn fastnn_cuda_add(a: *const f32, b: *const f32, out: *mut f32, n: usize) -> c_int;
    pub fn fastnn_cuda_sub(a: *const f32, b: *const f32, out: *mut f32, n: usize) -> c_int;
    pub fn fastnn_cuda_mul(a: *const f32, b: *const f32, out: *mut f32, n: usize) -> c_int;
    pub fn fastnn_cuda_div(a: *const f32, b: *const f32, out: *mut f32, n: usize) -> c_int;

    pub fn fastnn_cuda_add_scalar(a: *const f32, s: f32, out: *mut f32, n: usize) -> c_int;
    pub fn fastnn_cuda_mul_scalar(a: *const f32, s: f32, out: *mut f32, n: usize) -> c_int;
    pub fn fastnn_cuda_pow_scalar(a: *const f32, s: f32, out: *mut f32, n: usize) -> c_int;

    pub fn fastnn_cuda_sqrt(a: *const f32, out: *mut f32, n: usize) -> c_int;
    pub fn fastnn_cuda_abs(a: *const f32, out: *mut f32, n: usize) -> c_int;
    pub fn fastnn_cuda_neg(a: *const f32, out: *mut f32, n: usize) -> c_int;
    pub fn fastnn_cuda_exp(a: *const f32, out: *mut f32, n: usize) -> c_int;
    pub fn fastnn_cuda_log(a: *const f32, out: *mut f32, n: usize) -> c_int;
    pub fn fastnn_cuda_clamp(a: *const f32, lo: f32, hi: f32, out: *mut f32, n: usize) -> c_int;

    pub fn fastnn_cuda_relu(x: *const f32, out: *mut f32, n: usize) -> c_int;
    pub fn fastnn_cuda_relu_backward(
        grad: *const f32,
        x: *const f32,
        out: *mut f32,
        n: usize,
    ) -> c_int;
    pub fn fastnn_cuda_sigmoid(x: *const f32, out: *mut f32, n: usize) -> c_int;
    pub fn fastnn_cuda_sigmoid_backward(
        grad: *const f32,
        y: *const f32,
        out: *mut f32,
        n: usize,
    ) -> c_int;
    pub fn fastnn_cuda_tanh_forward(x: *const f32, out: *mut f32, n: usize) -> c_int;
    pub fn fastnn_cuda_tanh_backward(
        grad: *const f32,
        y: *const f32,
        out: *mut f32,
        n: usize,
    ) -> c_int;
    pub fn fastnn_cuda_gelu(x: *const f32, out: *mut f32, n: usize) -> c_int;
    pub fn fastnn_cuda_gelu_backward(
        grad: *const f32,
        x: *const f32,
        out: *mut f32,
        n: usize,
    ) -> c_int;
    pub fn fastnn_cuda_silu(x: *const f32, out: *mut f32, n: usize) -> c_int;
    pub fn fastnn_cuda_silu_backward(
        grad: *const f32,
        x: *const f32,
        out: *mut f32,
        n: usize,
    ) -> c_int;
    pub fn fastnn_cuda_leaky_relu(x: *const f32, slope: f32, out: *mut f32, n: usize) -> c_int;

    pub fn fastnn_cuda_softmax(x: *const f32, out: *mut f32, rows: c_int, cols: c_int) -> c_int;
    pub fn fastnn_cuda_softmax_backward(
        grad: *const f32,
        y: *const f32,
        out: *mut f32,
        rows: c_int,
        cols: c_int,
    ) -> c_int;
    pub fn fastnn_cuda_log_softmax(x: *const f32, out: *mut f32, rows: c_int, cols: c_int)
    -> c_int;

    pub fn fastnn_cuda_layer_norm_forward(
        x: *const f32,
        gamma: *const f32,
        beta: *const f32,
        out: *mut f32,
        mean: *mut f32,
        inv_std: *mut f32,
        rows: c_int,
        cols: c_int,
        eps: f32,
    ) -> c_int;
    pub fn fastnn_cuda_layer_norm_backward(
        grad: *const f32,
        x: *const f32,
        gamma: *const f32,
        mean: *const f32,
        inv_std: *const f32,
        grad_x: *mut f32,
        grad_gamma: *mut f32,
        grad_beta: *mut f32,
        rows: c_int,
        cols: c_int,
    ) -> c_int;

    pub fn fastnn_cuda_embedding_forward(
        ids: *const c_int,
        weight: *const f32,
        out: *mut f32,
        n_ids: c_int,
        dim: c_int,
    ) -> c_int;
    pub fn fastnn_cuda_embedding_backward(
        ids: *const c_int,
        grad: *const f32,
        grad_weight: *mut f32,
        n_ids: c_int,
        dim: c_int,
        vocab: c_int,
    ) -> c_int;

    pub fn fastnn_cuda_im2col(
        input: *const f32,
        cols: *mut f32,
        n: c_int,
        c: c_int,
        h: c_int,
        w: c_int,
        kh: c_int,
        kw: c_int,
        sh: c_int,
        sw: c_int,
        ph: c_int,
        pw: c_int,
        dh: c_int,
        dw: c_int,
        out_h: c_int,
        out_w: c_int,
    ) -> c_int;
    pub fn fastnn_cuda_col2im(
        cols: *const f32,
        image: *mut f32,
        n: c_int,
        c: c_int,
        h: c_int,
        w: c_int,
        kh: c_int,
        kw: c_int,
        sh: c_int,
        sw: c_int,
        ph: c_int,
        pw: c_int,
        dh: c_int,
        dw: c_int,
        out_h: c_int,
        out_w: c_int,
    ) -> c_int;

    pub fn fastnn_cuda_matmul(
        a: *const f32,
        b: *const f32,
        c: *mut f32,
        m: c_int,
        n: c_int,
        k: c_int,
        lda: c_int,
        ldb: c_int,
        ldc: c_int,
        alpha: f32,
        beta: f32,
    ) -> c_int;
    pub fn fastnn_cuda_matmul_batched(
        a: *const f32,
        b: *const f32,
        c: *mut f32,
        m: c_int,
        n: c_int,
        k: c_int,
        batch: c_int,
        alpha: f32,
        beta: f32,
    ) -> c_int;
    pub fn fastnn_cuda_matmul_nt(
        a: *const f32,
        b: *const f32,
        c: *mut f32,
        m: c_int,
        n: c_int,
        k: c_int,
    ) -> c_int;
    pub fn fastnn_cuda_matmul_tn(
        a: *const f32,
        b: *const f32,
        c: *mut f32,
        m: c_int,
        n: c_int,
        k: c_int,
    ) -> c_int;
    pub fn fastnn_cuda_matmul_batched_nt(
        a: *const f32,
        b: *const f32,
        c: *mut f32,
        m: c_int,
        n: c_int,
        k: c_int,
        batch: c_int,
    ) -> c_int;
    pub fn fastnn_cuda_matmul_batched_tn(
        a: *const f32,
        b: *const f32,
        c: *mut f32,
        m: c_int,
        n: c_int,
        k: c_int,
        batch: c_int,
    ) -> c_int;

    pub fn fastnn_cuda_transpose(x: *const f32, out: *mut f32, rows: c_int, cols: c_int) -> c_int;
    pub fn fastnn_cuda_transpose_batched(
        x: *const f32,
        out: *mut f32,
        batch: c_int,
        rows: c_int,
        cols: c_int,
    ) -> c_int;
    /// General gather: `out[i] = x[dot(coords(i, out_strides), in_strides[perm])]`.
    /// A stride of 0 in `in_strides` broadcasts that dimension, which is how
    /// `expand` reuses this kernel.
    pub fn fastnn_cuda_permute_nd(
        x: *const f32,
        out: *mut f32,
        out_strides: *const c_int,
        in_strides: *const c_int,
        perm: *const c_int,
        ndim: c_int,
        numel: c_int,
    ) -> c_int;

    pub fn fastnn_cuda_sum(x: *const f32, out: *mut f32, n: usize) -> c_int;
    pub fn fastnn_cuda_mean(x: *const f32, out: *mut f32, n: usize) -> c_int;
    pub fn fastnn_cuda_max(x: *const f32, out: *mut f32, n: usize) -> c_int;
    pub fn fastnn_cuda_min(x: *const f32, out: *mut f32, n: usize) -> c_int;
    pub fn fastnn_cuda_sum_axis(
        x: *const f32,
        out: *mut f32,
        shape: *const c_int,
        ndim: c_int,
        axis: c_int,
        total: c_int,
    ) -> c_int;
}

/// Signature shared by every element-wise binary kernel.
pub type BinaryKernel = unsafe extern "C" fn(*const f32, *const f32, *mut f32, usize) -> c_int;
/// Signature shared by every element-wise unary kernel.
pub type UnaryKernel = unsafe extern "C" fn(*const f32, *mut f32, usize) -> c_int;
/// Signature shared by every tensor-scalar kernel.
pub type ScalarKernel = unsafe extern "C" fn(*const f32, f32, *mut f32, usize) -> c_int;