1 | // Copyright 2020 Yevhenii Reizner |
2 | // |
3 | // Use of this source code is governed by a BSD-style license that can be |
4 | // found in the LICENSE file. |
5 | |
6 | // Based on https://github.com/Lokathor/wide (Zlib) |
7 | |
8 | use bytemuck::cast; |
9 | |
10 | use super::f32x4; |
11 | |
12 | cfg_if::cfg_if! { |
13 | if #[cfg(all(feature = "simd" , target_feature = "sse2" ))] { |
14 | #[cfg (target_arch = "x86" )] |
15 | use core::arch::x86::*; |
16 | #[cfg (target_arch = "x86_64" )] |
17 | use core::arch::x86_64::*; |
18 | |
19 | #[derive (Clone, Copy, Debug)] |
20 | #[repr (C, align(16))] |
21 | pub struct i32x4(pub __m128i); |
22 | } else if #[cfg(all(feature = "simd" , target_feature = "simd128" ))] { |
23 | use core::arch::wasm32::*; |
24 | |
25 | #[derive (Clone, Copy, Debug)] |
26 | #[repr (C, align(16))] |
27 | pub struct i32x4(pub v128); |
28 | } else if #[cfg(all(feature = "simd" , target_arch = "aarch64" , target_feature = "neon" ))] { |
29 | use core::arch::aarch64::*; |
30 | |
31 | #[derive (Clone, Copy, Debug)] |
32 | #[repr (C, align(16))] |
33 | pub struct i32x4(pub int32x4_t); |
34 | } else { |
35 | #[derive (Clone, Copy, Debug)] |
36 | #[repr (C, align(16))] |
37 | pub struct i32x4([i32; 4]); |
38 | } |
39 | } |
40 | |
41 | unsafe impl bytemuck::Zeroable for i32x4 {} |
42 | unsafe impl bytemuck::Pod for i32x4 {} |
43 | |
44 | impl Default for i32x4 { |
45 | fn default() -> Self { |
46 | Self::splat(0) |
47 | } |
48 | } |
49 | |
50 | impl i32x4 { |
51 | pub fn splat(n: i32) -> Self { |
52 | cast([n, n, n, n]) |
53 | } |
54 | |
55 | pub fn blend(self, t: Self, f: Self) -> Self { |
56 | cfg_if::cfg_if! { |
57 | if #[cfg(all(feature = "simd" , target_feature = "sse4.1" ))] { |
58 | Self(unsafe { _mm_blendv_epi8(f.0, t.0, self.0) }) |
59 | } else if #[cfg(all(feature = "simd" , target_feature = "simd128" ))] { |
60 | Self(v128_bitselect(t.0, f.0, self.0)) |
61 | } else if #[cfg(all(feature = "simd" , target_arch = "aarch64" , target_feature = "neon" ))] { |
62 | Self(unsafe { vbslq_s32(cast(self.0), t.0, f.0) }) |
63 | } else { |
64 | super::generic_bit_blend(self, t, f) |
65 | } |
66 | } |
67 | } |
68 | |
69 | pub fn cmp_eq(self, rhs: Self) -> Self { |
70 | cfg_if::cfg_if! { |
71 | if #[cfg(all(feature = "simd" , target_feature = "sse2" ))] { |
72 | cast(Self(cast(unsafe { _mm_cmpeq_epi32(self.0, rhs.0) }))) |
73 | } else if #[cfg(all(feature = "simd" , target_feature = "simd128" ))] { |
74 | Self(i32x4_eq(self.0, rhs.0)) |
75 | } else if #[cfg(all(feature = "simd" , target_arch = "aarch64" , target_feature = "neon" ))] { |
76 | Self(unsafe { cast(vceqq_s32(self.0, rhs.0)) }) |
77 | } else { |
78 | Self([ |
79 | if self.0[0] == rhs.0[0] { -1 } else { 0 }, |
80 | if self.0[1] == rhs.0[1] { -1 } else { 0 }, |
81 | if self.0[2] == rhs.0[2] { -1 } else { 0 }, |
82 | if self.0[3] == rhs.0[3] { -1 } else { 0 }, |
83 | ]) |
84 | } |
85 | } |
86 | } |
87 | |
88 | pub fn cmp_gt(self, rhs: Self) -> Self { |
89 | cfg_if::cfg_if! { |
90 | if #[cfg(all(feature = "simd" , target_feature = "sse2" ))] { |
91 | cast(Self(cast(unsafe { _mm_cmpgt_epi32(self.0, rhs.0) }))) |
92 | } else if #[cfg(all(feature = "simd" , target_feature = "simd128" ))] { |
93 | Self(i32x4_gt(self.0, rhs.0)) |
94 | } else if #[cfg(all(feature = "simd" , target_arch = "aarch64" , target_feature = "neon" ))] { |
95 | Self(unsafe { cast(vcgtq_s32(self.0, rhs.0)) }) |
96 | } else { |
97 | Self([ |
98 | if self.0[0] > rhs.0[0] { -1 } else { 0 }, |
99 | if self.0[1] > rhs.0[1] { -1 } else { 0 }, |
100 | if self.0[2] > rhs.0[2] { -1 } else { 0 }, |
101 | if self.0[3] > rhs.0[3] { -1 } else { 0 }, |
102 | ]) |
103 | } |
104 | } |
105 | } |
106 | |
107 | pub fn cmp_lt(self, rhs: Self) -> Self { |
108 | cfg_if::cfg_if! { |
109 | if #[cfg(all(feature = "simd" , target_feature = "sse2" ))] { |
110 | cast(Self(cast(unsafe { _mm_cmplt_epi32(self.0, rhs.0) }))) |
111 | } else if #[cfg(all(feature = "simd" , target_feature = "simd128" ))] { |
112 | Self(i32x4_lt(self.0, rhs.0)) |
113 | } else if #[cfg(all(feature = "simd" , target_arch = "aarch64" , target_feature = "neon" ))] { |
114 | Self(unsafe { cast(vcltq_s32(self.0, rhs.0)) }) |
115 | } else { |
116 | Self([ |
117 | if self.0[0] < rhs.0[0] { -1 } else { 0 }, |
118 | if self.0[1] < rhs.0[1] { -1 } else { 0 }, |
119 | if self.0[2] < rhs.0[2] { -1 } else { 0 }, |
120 | if self.0[3] < rhs.0[3] { -1 } else { 0 }, |
121 | ]) |
122 | } |
123 | } |
124 | } |
125 | |
126 | pub fn to_f32x4(self) -> f32x4 { |
127 | cfg_if::cfg_if! { |
128 | if #[cfg(all(feature = "simd" , target_feature = "sse2" ))] { |
129 | cast(Self(cast(unsafe { _mm_cvtepi32_ps(self.0) }))) |
130 | } else if #[cfg(all(feature = "simd" , target_feature = "simd128" ))] { |
131 | cast(Self(f32x4_convert_i32x4(self.0))) |
132 | } else if #[cfg(all(feature = "simd" , target_arch = "aarch64" , target_feature = "neon" ))] { |
133 | cast(Self(unsafe { cast(vcvtq_f32_s32(self.0)) })) |
134 | } else { |
135 | let arr: [i32; 4] = cast(self); |
136 | cast([ |
137 | arr[0] as f32, |
138 | arr[1] as f32, |
139 | arr[2] as f32, |
140 | arr[3] as f32, |
141 | ]) |
142 | } |
143 | } |
144 | } |
145 | |
146 | pub fn to_f32x4_bitcast(self) -> f32x4 { |
147 | bytemuck::cast(self) |
148 | } |
149 | } |
150 | |
151 | impl From<[i32; 4]> for i32x4 { |
152 | fn from(v: [i32; 4]) -> Self { |
153 | cast(v) |
154 | } |
155 | } |
156 | |
157 | impl From<i32x4> for [i32; 4] { |
158 | fn from(v: i32x4) -> Self { |
159 | cast(v) |
160 | } |
161 | } |
162 | |
163 | impl core::ops::Add for i32x4 { |
164 | type Output = Self; |
165 | |
166 | fn add(self, rhs: Self) -> Self::Output { |
167 | cfg_if::cfg_if! { |
168 | if #[cfg(all(feature = "simd" , target_feature = "sse2" ))] { |
169 | Self(unsafe { _mm_add_epi32(self.0, rhs.0) }) |
170 | } else if #[cfg(all(feature = "simd" , target_feature = "simd128" ))] { |
171 | Self(i32x4_add(self.0, rhs.0)) |
172 | } else if #[cfg(all(feature = "simd" , target_arch = "aarch64" , target_feature = "neon" ))] { |
173 | Self(unsafe { vaddq_s32(self.0, rhs.0) }) |
174 | } else { |
175 | Self([ |
176 | self.0[0].wrapping_add(rhs.0[0]), |
177 | self.0[1].wrapping_add(rhs.0[1]), |
178 | self.0[2].wrapping_add(rhs.0[2]), |
179 | self.0[3].wrapping_add(rhs.0[3]), |
180 | ]) |
181 | } |
182 | } |
183 | } |
184 | } |
185 | |
186 | impl core::ops::BitAnd for i32x4 { |
187 | type Output = Self; |
188 | |
189 | fn bitand(self, rhs: Self) -> Self::Output { |
190 | cfg_if::cfg_if! { |
191 | if #[cfg(all(feature = "simd" , target_feature = "sse2" ))] { |
192 | Self(unsafe { _mm_and_si128(self.0, rhs.0) }) |
193 | } else if #[cfg(all(feature = "simd" , target_feature = "simd128" ))] { |
194 | Self(v128_and(self.0, rhs.0)) |
195 | } else if #[cfg(all(feature = "simd" , target_arch = "aarch64" , target_feature = "neon" ))] { |
196 | Self(unsafe { vandq_s32(self.0, rhs.0) }) |
197 | } else { |
198 | Self([ |
199 | self.0[0] & rhs.0[0], |
200 | self.0[1] & rhs.0[1], |
201 | self.0[2] & rhs.0[2], |
202 | self.0[3] & rhs.0[3], |
203 | ]) |
204 | } |
205 | } |
206 | } |
207 | } |
208 | |
209 | impl core::ops::Mul for i32x4 { |
210 | type Output = Self; |
211 | |
212 | fn mul(self, rhs: Self) -> Self::Output { |
213 | cfg_if::cfg_if! { |
214 | if #[cfg(all(feature = "simd" , target_feature = "sse4.1" ))] { |
215 | Self(unsafe { _mm_mullo_epi32(self.0, rhs.0) }) |
216 | } else if #[cfg(all(feature = "simd" , target_feature = "simd128" ))] { |
217 | Self(i32x4_mul(self.0, rhs.0)) |
218 | } else if #[cfg(all(feature = "simd" , target_arch = "aarch64" , target_feature = "neon" ))] { |
219 | Self(unsafe { vmulq_s32(self.0, rhs.0) }) |
220 | } else { |
221 | // Cast is required, since we have to use scalar multiplication on SSE2. |
222 | let a: [i32; 4] = cast(self); |
223 | let b: [i32; 4] = cast(rhs); |
224 | Self(cast([ |
225 | a[0].wrapping_mul(b[0]), |
226 | a[1].wrapping_mul(b[1]), |
227 | a[2].wrapping_mul(b[2]), |
228 | a[3].wrapping_mul(b[3]), |
229 | ])) |
230 | } |
231 | } |
232 | } |
233 | } |
234 | |
235 | impl core::ops::BitOr for i32x4 { |
236 | type Output = Self; |
237 | |
238 | #[inline ] |
239 | fn bitor(self, rhs: Self) -> Self::Output { |
240 | cfg_if::cfg_if! { |
241 | if #[cfg(all(feature = "simd" , target_feature = "sse2" ))] { |
242 | Self(unsafe { _mm_or_si128(self.0, rhs.0) }) |
243 | } else if #[cfg(all(feature = "simd" , target_feature = "simd128" ))] { |
244 | Self(v128_or(self.0, rhs.0)) |
245 | } else if #[cfg(all(feature = "simd" , target_arch = "aarch64" , target_feature = "neon" ))] { |
246 | Self(unsafe { vorrq_s32(self.0, rhs.0) }) |
247 | } else { |
248 | Self([ |
249 | self.0[0] | rhs.0[0], |
250 | self.0[1] | rhs.0[1], |
251 | self.0[2] | rhs.0[2], |
252 | self.0[3] | rhs.0[3], |
253 | ]) |
254 | } |
255 | } |
256 | } |
257 | } |
258 | |
259 | impl core::ops::BitXor for i32x4 { |
260 | type Output = Self; |
261 | |
262 | #[inline ] |
263 | fn bitxor(self, rhs: Self) -> Self::Output { |
264 | cfg_if::cfg_if! { |
265 | if #[cfg(all(feature = "simd" , target_feature = "sse2" ))] { |
266 | Self(unsafe { _mm_xor_si128(self.0, rhs.0) }) |
267 | } else if #[cfg(all(feature = "simd" , target_feature = "simd128" ))] { |
268 | Self(v128_xor(self.0, rhs.0)) |
269 | } else if #[cfg(all(feature = "simd" , target_arch = "aarch64" , target_feature = "neon" ))] { |
270 | Self(unsafe { veorq_s32(self.0, rhs.0) }) |
271 | } else { |
272 | Self([ |
273 | self.0[0] ^ rhs.0[0], |
274 | self.0[1] ^ rhs.0[1], |
275 | self.0[2] ^ rhs.0[2], |
276 | self.0[3] ^ rhs.0[3], |
277 | ]) |
278 | } |
279 | } |
280 | } |
281 | } |
282 | |