ggblas 0.1.2

Matrix multiplicatiosn in simple pure rust multithreadded blas-like interface. Exploits intrinsics on available targets.
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

trait Cpu<const ARR: usize> {
    type Unit;
    type Array;
    const STEP: usize;
    const EPR: usize;

    fn n() -> usize;
    unsafe fn zero() -> Self::Unit;
    unsafe fn zero_array() -> Self::Array;
    unsafe fn load(mem_addr: *const f32) -> Self::Unit;
    unsafe fn vec_fma(a: Self::Unit, b: Self::Unit, c: Self::Unit) -> Self::Unit;
    unsafe fn vec_reduce(x: Self::Array, y: *mut f32);
    unsafe fn from_f32(v: f32) -> Self::Unit;
    unsafe fn vec_store(mem_addr: *mut f32, a: Self::Unit);
}

#[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
#[cfg(target_feature = "avx")]
pub mod avx;
#[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
#[cfg(target_feature = "avx")]
pub use avx::CurrentCpu;

#[cfg(target_arch = "wasm32")]
#[cfg(target_feature = "simd128")]
pub mod simd128;
#[cfg(target_arch = "wasm32")]
#[cfg(target_feature = "simd128")]
pub use simd128::CurrentCpu;

#[cfg(any(target_arch = "arm", target_arch = "aarch64"))]
#[cfg(target_feature = "neon")]
pub mod neon;
#[cfg(any(target_arch = "arm", target_arch = "aarch64"))]
#[cfg(target_feature = "neon")]
pub use neon::CurrentCpu;

#[cfg(any(
    target_feature = "neon",
    target_feature = "avx",
    target_feature = "simd128"
))]
#[inline(never)]
/// # Safety
/// This requires the user to check that `k` is actually valid  for all pointers
pub unsafe fn vec_dot_f32(a_row: *const f32, b_row: *const f32, c: *mut f32, k: usize) {
    let np = k & !(CurrentCpu::STEP - 1);

    let mut sum = CurrentCpu::zero_array();
    let mut ax = CurrentCpu::zero_array();
    let mut ay = CurrentCpu::zero_array();

    for i in (0..np).step_by(CurrentCpu::STEP) {
        for j in 0..CurrentCpu::n() {
            ax[j] = CurrentCpu::load(a_row.add(i + j * CurrentCpu::EPR));
            ay[j] = CurrentCpu::load(b_row.add(i + j * CurrentCpu::EPR));

            sum[j] = CurrentCpu::vec_fma(sum[j], ax[j], ay[j]);
        }
    }

    CurrentCpu::vec_reduce(sum, c);

    // leftovers
    for i in np..k {
        *c += *a_row.add(i) * (*b_row.add(i));
    }
}

#[cfg(any(
    target_feature = "neon",
    target_feature = "avx",
    target_feature = "simd128"
))]
/// # Safety
/// This requires the user to check that `n` is actually valid  for all pointers
pub unsafe fn vec_mad_f32(b_row: *const f32, c_row: *mut f32, v: f32, n: usize) {
    let np = n & !(CurrentCpu::STEP - 1);

    let vx = CurrentCpu::from_f32(v);
    let mut ax = CurrentCpu::zero_array();
    let mut ay = CurrentCpu::zero_array();

    for i in (0..np).step_by(CurrentCpu::STEP) {
        for j in 0..CurrentCpu::n() {
            ax[j] = CurrentCpu::load(b_row.add(i + j * CurrentCpu::EPR));
            ay[j] = CurrentCpu::load(c_row.add(i + j * CurrentCpu::EPR));
            ay[j] = CurrentCpu::vec_fma(ay[j], ax[j], vx);
            CurrentCpu::vec_store(c_row.add(i + j * CurrentCpu::EPR), ay[j]);
        }
    }

    // leftovers
    for i in np..n {
        *c_row.add(i) += *b_row.add(i) * v;
    }
}

#[cfg(not(any(
    target_feature = "neon",
    target_feature = "avx",
    target_feature = "simd128"
)))]
#[inline(never)]
pub unsafe fn vec_dot_f32(a_row: *const f32, b_row: *const f32, c: *mut f32, k: usize) {
    // leftovers
    for i in 0..k {
        *c += *a_row.add(i) * (*b_row.add(i));
    }
}

#[cfg(not(any(
    target_feature = "neon",
    target_feature = "avx",
    target_feature = "simd128"
)))]
pub unsafe fn vec_mad_f32(a_row: *const f32, c_row: *mut f32, v: f32, n: usize) {
    for i in 0..n {
        *c_row.add(i) += *a_row.add(i) * v;
    }
}

#[cfg(feature = "f16")]
#[cfg(target_arch = "wasm32")]
#[cfg(target_feature = "simd128")]
pub mod simd128_f16;

#[cfg(feature = "f16")]
#[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
#[cfg(target_feature = "avx")]
mod avx_f16;

#[cfg(feature = "f16")]
#[cfg(not(any(target_arch = "arm", target_arch = "aarch64")))]
pub mod f16 {
    #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
    #[cfg(target_feature = "avx")]
    pub use super::avx_f16::CurrentCpuF16;

    #[cfg(target_arch = "wasm32")]
    #[cfg(target_feature = "simd128")]
    pub use super::simd128_f16::CurrentCpuF16;

    use half::f16;

    pub trait CpuF16<const ARR: usize> {
        type Unit;
        type Array;
        const STEP: usize;
        const EPR: usize;

        fn n() -> usize;
        unsafe fn zero() -> Self::Unit;
        unsafe fn zero_array() -> Self::Array;
        unsafe fn load(mem_addr: *const f16) -> Self::Unit;
        unsafe fn vec_fma(a: Self::Unit, b: Self::Unit, c: Self::Unit) -> Self::Unit;
        unsafe fn vec_reduce(x: Self::Array, y: *mut f32);
        unsafe fn from_f32(v: f32) -> Self::Unit;
        unsafe fn vec_store(mem_addr: *mut f16, a: Self::Unit);
    }

    #[cfg(any(
        target_feature = "neon",
        target_feature = "avx",
        target_feature = "simd128"
    ))]
    #[inline(never)]
    pub unsafe fn vec_dot_f16(a_row: *const f16, b_row: *const f16, c: *mut f32, k: usize) {
        let mut sumf = 0.0f32;
        let np = k & !(CurrentCpuF16::STEP - 1);

        let mut sum = CurrentCpuF16::zero_array();
        let mut ax = CurrentCpuF16::zero_array();
        let mut ay = CurrentCpuF16::zero_array();

        for i in (0..np).step_by(CurrentCpuF16::STEP) {
            for j in 0..CurrentCpuF16::n() {
                ax[j] = CurrentCpuF16::load(a_row.add(i + j * CurrentCpuF16::EPR));
                ay[j] = CurrentCpuF16::load(b_row.add(i + j * CurrentCpuF16::EPR));

                sum[j] = CurrentCpuF16::vec_fma(sum[j], ax[j], ay[j]);
            }
        }

        CurrentCpuF16::vec_reduce(sum, &mut sumf);

        // leftovers
        for i in np..k {
            sumf += (*a_row.add(i)).to_f32() * (*b_row.add(i)).to_f32();
        }
        *c = sumf;
    }

    #[cfg(not(any(
        target_feature = "neon",
        target_feature = "avx",
        target_feature = "simd128"
    )))]
    #[inline(never)]
    pub unsafe fn vec_dot_f16(a_row: *const f16, b_row: *const f16, c: *mut f32, k: usize) {
        // leftovers
        let mut sum = 0.0;
        for i in 0..k {
            sum += (*a_row.add(i)).to_f32() * (*b_row.add(i)).to_f32();
        }
        *c = sum;
    }

    pub unsafe fn f32_to_f16(x: *const f32, y: *mut f16, n: usize) {
        let mut i = 0;
        #[cfg(target_feature = "f16c")]
        #[cfg(target_feature = "avx")]
        {
            #[cfg(target_arch = "x86")]
            use core::arch::x86::*;
            #[cfg(target_arch = "x86_64")]
            use core::arch::x86_64::*;

            while i + 7 < n {
                let x_vec: __m256 = _mm256_loadu_ps(x.add(i));
                let y_vec = _mm256_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT);
                _mm_storeu_si128(y.add(i) as *mut __m128i, y_vec);
                i += 8;
            }
            while i + 3 < n {
                let x_vec: __m128 = _mm_loadu_ps(x.add(i));
                let y_vec = _mm_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT);
                _mm_storel_epi64(y.add(i) as *mut __m128i, y_vec);
                i += 4;
            }
        }
        while i < n {
            *y.add(i) = f16::from_f32(*x.add(i));
            i += 1;
        }
    }
}