redstone-ml 0.0.0

High-performance Machine Learning, Auto-Differentiation and Tensor Algebra crate for Rust
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
263
264
265
266
267

#![allow(unused_mut)]
#![allow(unused_variables)]

use crate::dtype::{IntegerDataType, NumericDataType};
use std::hint::assert_unchecked;


pub(super) fn get_sum_of_products_function<const N: usize, T: SumOfProductsType>(strides: &[usize; N])
                                                                                 -> unsafe fn(ptrs: &[*mut T; N], stride: &[usize; N], count: usize) {
    if N == 2 { // 1 operand + 1 output
        if strides[0] == 1 && strides[1] == 0 {
            // return <T as EinsumDataType>::operand_strides_1_out_stride_0; // can be implemented later if needed
        }
    }

    if N == 3 { // 2 operands + 1 output
        let mut code = if strides[0] == 0 { 0 } else if strides[0] == 1 { 4 } else { 8 };
        code += if strides[1] == 0 { 0 } else if strides[1] == 1 { 2 } else { 8 };
        code += if strides[2] == 0 { 0 } else if strides[2] == 1 { 1 } else { 8 };

        match code {
            2 => { return <T as SumOfProductsType>::sum_of_products_in_strides_0_1_out_stride_0; }
            3 => { return <T as SumOfProductsType>::sum_of_products_in_strides_0_1_out_stride_1; }
            4 => { return <T as SumOfProductsType>::sum_of_products_in_strides_1_0_out_stride_0; }
            5 => { return <T as SumOfProductsType>::sum_of_products_in_strides_1_0_out_stride_1; }
            6 => { return <T as SumOfProductsType>::sum_of_products_in_strides_1_1_out_stride_0; }
            7 => { return <T as SumOfProductsType>::sum_of_products_in_strides_1_1_out_stride_1; }
            _ => {}
        }
    }

    if strides[N - 1] == 0 {
        return <T as SumOfProductsType>::sum_of_products_out_stride_0;
    }

    <T as SumOfProductsType>::sum_of_products_generic
}

// called when the number of operands cannot be provided as a const generic
pub(super) fn get_sum_of_products_function_generic_nops<T: SumOfProductsType>(strides: &[usize])
                                                                              -> unsafe fn(ptrs: &[*mut T], stride: &[usize], count: usize) {
    let nops = strides.len() - 1;

    if strides[nops] == 0 {
        return match nops {
            3 => { <T as SumOfProductsType>::sum_of_products_in_strides_n_n_n_out_stride_0 },
            _ => { <T as SumOfProductsType>::sum_of_products_out_stride_0_ }
        }
    }

    <T as SumOfProductsType>::sum_of_products_generic_
}


pub(crate) trait SumOfProductsType: NumericDataType {
    #[inline(always)]
    unsafe fn sum_of_products_generic_(ptrs: &[*mut Self], strides: &[usize], count: usize) {
        let nops = ptrs.len();
        assert_unchecked(count > 0);
        assert_unchecked(nops > 0);

        let dst = ptrs[nops - 1];
        let ptrs = &ptrs[0..nops - 1];

        let mut k = count;
        while k != 0 {
            k -= 1;

            let dst = dst.add(k * strides[nops - 1]);
            *dst += ptrs.iter().zip(strides.iter())
                        .map(|(ptr, stride)| *ptr.add(k * stride))
                        .product();
        }
    }

    #[inline(always)]
    unsafe fn sum_of_products_out_stride_0_(ptrs: &[*mut Self], strides: &[usize], count: usize) {
        assert_unchecked(count > 0);
        assert_unchecked(strides[strides.len() - 1] == 0);

        let nops = ptrs.len() - 1;
        let dst = ptrs[nops];
        let ptrs = &ptrs[..nops];

        let mut sum = Self::default();

        let mut k = count;
        while k != 0 {
            k -= 1;

            sum += ptrs.iter().zip(strides.iter())
                       .map(|(ptr, stride)| *ptr.add(k * stride))
                       .product();
        }

        *dst += sum;
    }

    #[inline(always)]
    unsafe fn sum_of_products_muladd<const N: usize>(ptrs: &[*mut Self; N], strides: &[usize; N], count: usize) {
        assert_unchecked(count > 0);
        assert_unchecked(N == 3);

        let mut dst = ptrs[N - 1];

        let scalar = *ptrs[0];
        let mut data = ptrs[1];

        for _ in 0..count {
            *dst = scalar.mul_add(*data, *dst);
            dst = dst.add(1);
            data = data.add(1);
        }
    }

    #[inline(always)]
    unsafe fn sum_of_scaled_array<const N: usize>(ptrs: &[*mut Self; N], strides: &[usize; N], count: usize) {
        assert_unchecked(count > 0);

        let mut dst = ptrs[N - 1];
        let scalar = *ptrs[0];
        let mut data = ptrs[1];

        let mut sum = Self::zero();
        for i in 0..count {
            sum += *data.add(i);
        }

        *dst = scalar.mul_add(sum, *dst);
    }

    #[inline(always)]
    unsafe fn sum_of_products_in_strides_1_1_out_stride_1<const N: usize>(ptrs: &[*mut Self; N], strides: &[usize; N], count: usize) {
        assert_unchecked(strides[0] == 1);
        assert_unchecked(strides[1] == 1);
        assert_unchecked(strides[2] == 1);
        assert_unchecked(N == 3);
        assert_unchecked(count > 0);

        let mut data0 = ptrs[0];
        let mut data1 = ptrs[1];
        let mut dst = ptrs[2];

        for i in 0..count {
            *dst += (*data0) * (*data1);

            dst = dst.add(1);
            data0 = data0.add(1);
            data1 = data1.add(1);
        }

        // TODO specialised SIMD implementation
    }

    #[inline(always)]
    unsafe fn sum_of_products_generic<const N: usize>(ptrs: &[*mut Self; N], strides: &[usize; N], count: usize) {
        Self::sum_of_products_generic_(ptrs, strides, count);
    }

    #[inline(always)]
    unsafe fn sum_of_products_in_strides_n_n_n_out_stride_0(ptrs: &[*mut Self], strides: &[usize], count: usize) {
        assert_unchecked(ptrs.len() == 4);
        assert_unchecked(strides.len() == 4);
        Self::sum_of_products_out_stride_0_(ptrs, strides, count);
    }

    #[inline(always)]
    unsafe fn sum_of_products_in_strides_n_n_out_stride_0(ptrs: &[*mut Self], strides: &[usize], count: usize) {
        assert_unchecked(ptrs.len() == 3);
        assert_unchecked(strides.len() == 3);
        Self::strided_dot_product(ptrs[0], strides[0], ptrs[1], strides[1], ptrs[2], count);
    }

    #[inline(always)]
    unsafe fn sum_of_products_out_stride_0<const N: usize>(ptrs: &[*mut Self; N], strides: &[usize; N], count: usize) {
        Self::sum_of_products_out_stride_0_(ptrs, strides, count);
    }

    #[inline(always)]
    unsafe fn sum_of_products_in_strides_0_1_out_stride_0<const N: usize>(ptrs: &[*mut Self; N], strides: &[usize; N], count: usize) {
        assert_unchecked(N == 3);
        assert_unchecked(strides[0] == 0);
        assert_unchecked(strides[1] == 1);
        assert_unchecked(strides[2] == 0);
        Self::sum_of_scaled_array(ptrs, strides, count);
    }

    #[inline(always)]
    unsafe fn sum_of_products_in_strides_0_1_out_stride_1<const N: usize>(ptrs: &[*mut Self; N], strides: &[usize; N], count: usize) {
        assert_unchecked(N == 3);
        assert_unchecked(strides[0] == 0);
        assert_unchecked(strides[1] == 1);
        assert_unchecked(strides[2] == 1);
        Self::sum_of_products_muladd(ptrs, strides, count);
    }

    #[inline(always)]
    unsafe fn sum_of_products_in_strides_1_0_out_stride_0<const N: usize>(ptrs: &[*mut Self; N], strides: &[usize; N], count: usize) {
        assert_unchecked(strides[0] == 1);
        assert_unchecked(strides[1] == 0);
        assert_unchecked(strides[2] == 0);
        assert_unchecked(N == 3);

        let mut ptrs = *ptrs;
        ptrs.swap(0, 1);

        Self::sum_of_scaled_array(&ptrs, strides, count);
    }

    #[inline(always)]
    unsafe fn sum_of_products_in_strides_1_0_out_stride_1<const N: usize>(ptrs: &[*mut Self; N], strides: &[usize; N], count: usize) {
        assert_unchecked(strides[0] == 1);
        assert_unchecked(strides[1] == 0);
        assert_unchecked(strides[2] == 1);
        assert_unchecked(N == 3);

        let mut ptrs = *ptrs;
        ptrs.swap(0, 1);

        Self::sum_of_products_muladd(&ptrs, strides, count);
    }

    #[inline(always)]
    unsafe fn sum_of_products_in_strides_1_1_out_stride_0<const N: usize>(ptrs: &[*mut Self; N], strides: &[usize; N], count: usize) {
        assert_unchecked(strides[0] == 1);
        assert_unchecked(strides[1] == 1);
        assert_unchecked(strides[2] == 0);
        assert_unchecked(N == 3);
        assert_unchecked(count > 0);

        Self::dot_product(ptrs[0], ptrs[1], ptrs[2], count);
    }
}

impl<T: IntegerDataType> SumOfProductsType for T {}

macro_rules! simd_sum_of_products_kernels {
    ($ptrs:ident, $strides:ident, $count:ident, $dst:ident, $($func_name:ident, { $($body:tt)* };)+) => {
        $(
            #[cfg(neon_simd)]
            unsafe fn $func_name<const N: usize>($ptrs: &[*mut Self; N], $strides: &[usize; N], mut $count: usize) {
                use crate::ops::simd_sum_of_products::SIMDSumOfProducts;

                assert_unchecked($count > 0);
                $($body)*
            }
        )+
    }
}

macro_rules! accelerated_sum_of_products {
    ($ptrs:ident, $strides:ident, $count:ident, $dst:ident, $($func_name:ident, { $($body:tt)* },)+) => {
        impl SumOfProductsType for f32 {
            simd_sum_of_products_kernels!($ptrs, $strides, $count, $dst, $($func_name, { $($body)* };)+);
        }

        impl SumOfProductsType for f64 {
            simd_sum_of_products_kernels!($ptrs, $strides, $count, $dst, $($func_name, { $($body)* };)+);
        }
    };
}

accelerated_sum_of_products!(ptrs, strides, count, dst,
    sum_of_products_muladd, { Self::simd_sum_of_products_muladd(*ptrs[0], ptrs[1], ptrs[2], count); },

    sum_of_scaled_array, { Self::simd_sum_of_scaled_array(*ptrs[0], ptrs[1], ptrs[2], count); },
);