matrixmultiply/
cgemm_kernel.rs

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
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
// Copyright 2016 - 2021 Ulrik Sverdrup "bluss"
//
// Licensed under the Apache License, Version 2.0 <LICENSE-APACHE or
// http://www.apache.org/licenses/LICENSE-2.0> or the MIT license
// <LICENSE-MIT or http://opensource.org/licenses/MIT>, at your
// option. This file may not be copied, modified, or distributed
// except according to those terms.

use crate::kernel::GemmKernel;
use crate::kernel::GemmSelect;
use crate::kernel::{U2, U4, c32, Element, c32_mul as mul};
use crate::archparam;
use crate::cgemm_common::pack_complex;

#[cfg(any(target_arch="x86", target_arch="x86_64"))]
struct KernelAvx2;
#[cfg(any(target_arch="x86", target_arch="x86_64"))]
struct KernelFma;

#[cfg(target_arch = "aarch64")]
#[cfg(has_aarch64_simd)]
struct KernelNeon;

struct KernelFallback;

type T = c32;
type TReal = f32;

/// Detect which implementation to use and select it using the selector's
/// .select(Kernel) method.
///
/// This function is called one or more times during a whole program's
/// execution, it may be called for each gemm kernel invocation or fewer times.
#[inline]
pub(crate) fn detect<G>(selector: G) where G: GemmSelect<T> {
    // dispatch to specific compiled versions
    #[cfg(any(target_arch="x86", target_arch="x86_64"))]
    {
        if is_x86_feature_detected_!("fma") {
            if is_x86_feature_detected_!("avx2") {
                return selector.select(KernelAvx2);
            }
            return selector.select(KernelFma);
        }
    }
    #[cfg(target_arch = "aarch64")]
    #[cfg(has_aarch64_simd)]
    {
        if is_aarch64_feature_detected_!("neon") {
            return selector.select(KernelNeon);
        }
    }
    return selector.select(KernelFallback);
}

#[cfg(any(target_arch="x86", target_arch="x86_64"))]
impl GemmKernel for KernelAvx2 {
    type Elem = T;

    type MRTy = U4;
    type NRTy = U4;

    #[inline(always)]
    fn align_to() -> usize { 32 }

    #[inline(always)]
    fn always_masked() -> bool { KernelFallback::always_masked() }

    #[inline(always)]
    fn nc() -> usize { archparam::C_NC }
    #[inline(always)]
    fn kc() -> usize { archparam::C_KC }
    #[inline(always)]
    fn mc() -> usize { archparam::C_MC }

    pack_methods!{}

    #[inline(always)]
    unsafe fn kernel(
        k: usize,
        alpha: T,
        a: *const T,
        b: *const T,
        beta: T,
        c: *mut T, rsc: isize, csc: isize) {
        kernel_target_avx2(k, alpha, a, b, beta, c, rsc, csc)
    }
}

#[cfg(any(target_arch="x86", target_arch="x86_64"))]
impl GemmKernel for KernelFma {
    type Elem = T;

    type MRTy = U4;
    type NRTy = U4;

    #[inline(always)]
    fn align_to() -> usize { 16 }

    #[inline(always)]
    fn always_masked() -> bool { KernelFallback::always_masked() }

    #[inline(always)]
    fn nc() -> usize { archparam::C_NC }
    #[inline(always)]
    fn kc() -> usize { archparam::C_KC }
    #[inline(always)]
    fn mc() -> usize { archparam::C_MC }

    pack_methods!{}

    #[inline(always)]
    unsafe fn kernel(
        k: usize,
        alpha: T,
        a: *const T,
        b: *const T,
        beta: T,
        c: *mut T, rsc: isize, csc: isize) {
        kernel_target_fma(k, alpha, a, b, beta, c, rsc, csc)
    }
}

#[cfg(target_arch = "aarch64")]
#[cfg(has_aarch64_simd)]
impl GemmKernel for KernelNeon {
    type Elem = T;

    type MRTy = U4;
    type NRTy = U2;

    #[inline(always)]
    fn align_to() -> usize { 16 }

    #[inline(always)]
    fn always_masked() -> bool { KernelFallback::always_masked() }

    #[inline(always)]
    fn nc() -> usize { archparam::C_NC }
    #[inline(always)]
    fn kc() -> usize { archparam::C_KC }
    #[inline(always)]
    fn mc() -> usize { archparam::C_MC }

    pack_methods!{}

    #[inline(always)]
    unsafe fn kernel(
        k: usize,
        alpha: T,
        a: *const T,
        b: *const T,
        beta: T,
        c: *mut T, rsc: isize, csc: isize) {
        kernel_target_neon(k, alpha, a, b, beta, c, rsc, csc)
    }
}

impl GemmKernel for KernelFallback {
    type Elem = T;

    type MRTy = U4;
    type NRTy = U2;

    #[inline(always)]
    fn align_to() -> usize { 0 }

    #[inline(always)]
    fn always_masked() -> bool { true }

    #[inline(always)]
    fn nc() -> usize { archparam::C_NC }
    #[inline(always)]
    fn kc() -> usize { archparam::C_KC }
    #[inline(always)]
    fn mc() -> usize { archparam::C_MC }

    pack_methods!{}

    #[inline(always)]
    unsafe fn kernel(
        k: usize,
        alpha: T,
        a: *const T,
        b: *const T,
        beta: T,
        c: *mut T, rsc: isize, csc: isize) {
        kernel_fallback_impl(k, alpha, a, b, beta, c, rsc, csc)
    }
}

// Kernel AVX2
#[cfg(any(target_arch="x86", target_arch="x86_64"))]
macro_rules! loop_m { ($i:ident, $e:expr) => { loop4!($i, $e) }; }
#[cfg(any(target_arch="x86", target_arch="x86_64"))]
macro_rules! loop_n { ($j:ident, $e:expr) => { loop4!($j, $e) }; }

#[cfg(any(target_arch="x86", target_arch="x86_64"))]
kernel_fallback_impl_complex! {
    // instantiate separately
    [inline target_feature(enable="avx2") target_feature(enable="fma")] [fma_yes]
    kernel_target_avx2, T, TReal, KernelAvx2::MR, KernelAvx2::NR, 4
}


// Kernel Fma
#[cfg(any(target_arch="x86", target_arch="x86_64"))]
macro_rules! loop_m { ($i:ident, $e:expr) => { loop4!($i, $e) }; }
#[cfg(any(target_arch="x86", target_arch="x86_64"))]
macro_rules! loop_n { ($j:ident, $e:expr) => { loop4!($j, $e) }; }

#[cfg(any(target_arch="x86", target_arch="x86_64"))]
kernel_fallback_impl_complex! {
    // instantiate separately
    [inline target_feature(enable="fma")] [fma_no]
    kernel_target_fma, T, TReal, KernelFma::MR, KernelFma::NR, 2
}

// Kernel neon

#[cfg(target_arch = "aarch64")]
#[cfg(has_aarch64_simd)]
macro_rules! loop_m { ($i:ident, $e:expr) => { loop4!($i, $e) }; }
#[cfg(target_arch = "aarch64")]
#[cfg(has_aarch64_simd)]
macro_rules! loop_n { ($j:ident, $e:expr) => { loop2!($j, $e) }; }

#[cfg(target_arch = "aarch64")]
#[cfg(has_aarch64_simd)]
kernel_fallback_impl_complex! {
    [inline target_feature(enable="neon")] [fma_yes]
    kernel_target_neon, T, TReal, KernelNeon::MR, KernelNeon::NR, 1
}

// Kernel fallback

macro_rules! loop_m { ($i:ident, $e:expr) => { loop4!($i, $e) }; }
macro_rules! loop_n { ($j:ident, $e:expr) => { loop2!($j, $e) }; }

kernel_fallback_impl_complex! {
    [inline(always)] [fma_no]
    kernel_fallback_impl, T, TReal, KernelFallback::MR, KernelFallback::NR, 1
}

#[inline(always)]
unsafe fn at(ptr: *const TReal, i: usize) -> TReal {
    *ptr.add(i)
}

#[cfg(test)]
mod tests {
    use super::*;
    use crate::kernel::test::test_complex_packed_kernel;

    #[test]
    fn test_kernel_fallback_impl() {
        test_complex_packed_kernel::<KernelFallback, _, TReal>("kernel");
    }

    #[cfg(target_arch = "aarch64")]
    #[cfg(has_aarch64_simd)]
    mod test_kernel_aarch64 {
        use super::test_complex_packed_kernel;
        use super::super::*;
        #[cfg(feature = "std")]
        use std::println;
        macro_rules! test_arch_kernels {
            ($($feature_name:tt, $name:ident, $kernel_ty:ty),*) => {
                $(
                #[test]
                fn $name() {
                    if is_aarch64_feature_detected_!($feature_name) {
                        test_complex_packed_kernel::<$kernel_ty, _, TReal>(stringify!($name));
                    } else {
                        #[cfg(feature = "std")]
                        println!("Skipping, host does not have feature: {:?}", $feature_name);
                    }
                }
                )*
            }
        }

        test_arch_kernels! {
            "neon", neon, KernelNeon
        }
    }

    #[cfg(any(target_arch="x86", target_arch="x86_64"))]
    mod test_arch_kernels {
        use super::test_complex_packed_kernel;
        use super::super::*;
        #[cfg(feature = "std")]
        use std::println;
        macro_rules! test_arch_kernels_x86 {
            ($($feature_name:tt, $name:ident, $kernel_ty:ty),*) => {
                $(
                #[test]
                fn $name() {
                    if is_x86_feature_detected_!($feature_name) {
                        test_complex_packed_kernel::<$kernel_ty, _, TReal>(stringify!($name));
                    } else {
                        #[cfg(feature = "std")]
                        println!("Skipping, host does not have feature: {:?}", $feature_name);
                    }
                }
                )*
            }
        }

        test_arch_kernels_x86! {
            "fma", fma, KernelFma,
            "avx2", avx2, KernelAvx2
        }
    }
}