1 | //! Streaming SIMD Extensions 3 (SSE3) |
2 | |
3 | use crate::core_arch::{simd::*, x86::*}; |
4 | use crate::intrinsics::simd::*; |
5 | |
6 | #[cfg (test)] |
7 | use stdarch_test::assert_instr; |
8 | |
9 | /// Alternatively add and subtract packed single-precision (32-bit) |
10 | /// floating-point elements in `a` to/from packed elements in `b`. |
11 | /// |
12 | /// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_addsub_ps) |
13 | #[inline ] |
14 | #[target_feature (enable = "sse3" )] |
15 | #[cfg_attr (test, assert_instr(addsubps))] |
16 | #[stable (feature = "simd_x86" , since = "1.27.0" )] |
17 | pub fn _mm_addsub_ps(a: __m128, b: __m128) -> __m128 { |
18 | unsafe { |
19 | let a: f32x4 = a.as_f32x4(); |
20 | let b: f32x4 = b.as_f32x4(); |
21 | let add: f32x4 = simd_add(x:a, y:b); |
22 | let sub: f32x4 = simd_sub(lhs:a, rhs:b); |
23 | simd_shuffle!(add, sub, [4, 1, 6, 3]) |
24 | } |
25 | } |
26 | |
27 | /// Alternatively add and subtract packed double-precision (64-bit) |
28 | /// floating-point elements in `a` to/from packed elements in `b`. |
29 | /// |
30 | /// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_addsub_pd) |
31 | #[inline ] |
32 | #[target_feature (enable = "sse3" )] |
33 | #[cfg_attr (test, assert_instr(addsubpd))] |
34 | #[stable (feature = "simd_x86" , since = "1.27.0" )] |
35 | pub fn _mm_addsub_pd(a: __m128d, b: __m128d) -> __m128d { |
36 | unsafe { |
37 | let a: f64x2 = a.as_f64x2(); |
38 | let b: f64x2 = b.as_f64x2(); |
39 | let add: f64x2 = simd_add(x:a, y:b); |
40 | let sub: f64x2 = simd_sub(lhs:a, rhs:b); |
41 | simd_shuffle!(add, sub, [2, 1]) |
42 | } |
43 | } |
44 | |
45 | /// Horizontally adds adjacent pairs of double-precision (64-bit) |
46 | /// floating-point elements in `a` and `b`, and pack the results. |
47 | /// |
48 | /// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_hadd_pd) |
49 | #[inline ] |
50 | #[target_feature (enable = "sse3" )] |
51 | #[cfg_attr (test, assert_instr(haddpd))] |
52 | #[stable (feature = "simd_x86" , since = "1.27.0" )] |
53 | pub fn _mm_hadd_pd(a: __m128d, b: __m128d) -> __m128d { |
54 | unsafe { haddpd(a, b) } |
55 | } |
56 | |
57 | /// Horizontally adds adjacent pairs of single-precision (32-bit) |
58 | /// floating-point elements in `a` and `b`, and pack the results. |
59 | /// |
60 | /// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_hadd_ps) |
61 | #[inline ] |
62 | #[target_feature (enable = "sse3" )] |
63 | #[cfg_attr (test, assert_instr(haddps))] |
64 | #[stable (feature = "simd_x86" , since = "1.27.0" )] |
65 | pub fn _mm_hadd_ps(a: __m128, b: __m128) -> __m128 { |
66 | unsafe { haddps(a, b) } |
67 | } |
68 | |
69 | /// Horizontally subtract adjacent pairs of double-precision (64-bit) |
70 | /// floating-point elements in `a` and `b`, and pack the results. |
71 | /// |
72 | /// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_hsub_pd) |
73 | #[inline ] |
74 | #[target_feature (enable = "sse3" )] |
75 | #[cfg_attr (test, assert_instr(hsubpd))] |
76 | #[stable (feature = "simd_x86" , since = "1.27.0" )] |
77 | pub fn _mm_hsub_pd(a: __m128d, b: __m128d) -> __m128d { |
78 | unsafe { hsubpd(a, b) } |
79 | } |
80 | |
81 | /// Horizontally adds adjacent pairs of single-precision (32-bit) |
82 | /// floating-point elements in `a` and `b`, and pack the results. |
83 | /// |
84 | /// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_hsub_ps) |
85 | #[inline ] |
86 | #[target_feature (enable = "sse3" )] |
87 | #[cfg_attr (test, assert_instr(hsubps))] |
88 | #[stable (feature = "simd_x86" , since = "1.27.0" )] |
89 | pub fn _mm_hsub_ps(a: __m128, b: __m128) -> __m128 { |
90 | unsafe { hsubps(a, b) } |
91 | } |
92 | |
93 | /// Loads 128-bits of integer data from unaligned memory. |
94 | /// This intrinsic may perform better than `_mm_loadu_si128` |
95 | /// when the data crosses a cache line boundary. |
96 | /// |
97 | /// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_lddqu_si128) |
98 | #[inline ] |
99 | #[target_feature (enable = "sse3" )] |
100 | #[cfg_attr (test, assert_instr(lddqu))] |
101 | #[stable (feature = "simd_x86" , since = "1.27.0" )] |
102 | pub unsafe fn _mm_lddqu_si128(mem_addr: *const __m128i) -> __m128i { |
103 | transmute(src:lddqu(mem_addr as *const _)) |
104 | } |
105 | |
106 | /// Duplicate the low double-precision (64-bit) floating-point element |
107 | /// from `a`. |
108 | /// |
109 | /// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_movedup_pd) |
110 | #[inline ] |
111 | #[target_feature (enable = "sse3" )] |
112 | #[cfg_attr (test, assert_instr(movddup))] |
113 | #[stable (feature = "simd_x86" , since = "1.27.0" )] |
114 | pub fn _mm_movedup_pd(a: __m128d) -> __m128d { |
115 | unsafe { simd_shuffle!(a, a, [0, 0]) } |
116 | } |
117 | |
118 | /// Loads a double-precision (64-bit) floating-point element from memory |
119 | /// into both elements of return vector. |
120 | /// |
121 | /// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_loaddup_pd) |
122 | #[inline ] |
123 | #[target_feature (enable = "sse3" )] |
124 | #[cfg_attr (test, assert_instr(movddup))] |
125 | #[stable (feature = "simd_x86" , since = "1.27.0" )] |
126 | pub unsafe fn _mm_loaddup_pd(mem_addr: *const f64) -> __m128d { |
127 | _mm_load1_pd(mem_addr) |
128 | } |
129 | |
130 | /// Duplicate odd-indexed single-precision (32-bit) floating-point elements |
131 | /// from `a`. |
132 | /// |
133 | /// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_movehdup_ps) |
134 | #[inline ] |
135 | #[target_feature (enable = "sse3" )] |
136 | #[cfg_attr (test, assert_instr(movshdup))] |
137 | #[stable (feature = "simd_x86" , since = "1.27.0" )] |
138 | pub fn _mm_movehdup_ps(a: __m128) -> __m128 { |
139 | unsafe { simd_shuffle!(a, a, [1, 1, 3, 3]) } |
140 | } |
141 | |
142 | /// Duplicate even-indexed single-precision (32-bit) floating-point elements |
143 | /// from `a`. |
144 | /// |
145 | /// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_moveldup_ps) |
146 | #[inline ] |
147 | #[target_feature (enable = "sse3" )] |
148 | #[cfg_attr (test, assert_instr(movsldup))] |
149 | #[stable (feature = "simd_x86" , since = "1.27.0" )] |
150 | pub fn _mm_moveldup_ps(a: __m128) -> __m128 { |
151 | unsafe { simd_shuffle!(a, a, [0, 0, 2, 2]) } |
152 | } |
153 | |
154 | #[allow (improper_ctypes)] |
155 | unsafe extern "C" { |
156 | #[link_name = "llvm.x86.sse3.hadd.pd" ] |
157 | unsafefn haddpd(a: __m128d, b: __m128d) -> __m128d; |
158 | #[link_name = "llvm.x86.sse3.hadd.ps" ] |
159 | unsafefn haddps(a: __m128, b: __m128) -> __m128; |
160 | #[link_name = "llvm.x86.sse3.hsub.pd" ] |
161 | unsafefn hsubpd(a: __m128d, b: __m128d) -> __m128d; |
162 | #[link_name = "llvm.x86.sse3.hsub.ps" ] |
163 | unsafefn hsubps(a: __m128, b: __m128) -> __m128; |
164 | #[link_name = "llvm.x86.sse3.ldu.dq" ] |
165 | unsafefn lddqu(mem_addr: *const i8) -> i8x16; |
166 | } |
167 | |
168 | #[cfg (test)] |
169 | mod tests { |
170 | use stdarch_test::simd_test; |
171 | |
172 | use crate::core_arch::x86::*; |
173 | |
174 | #[simd_test(enable = "sse3" )] |
175 | unsafe fn test_mm_addsub_ps() { |
176 | let a = _mm_setr_ps(-1.0, 5.0, 0.0, -10.0); |
177 | let b = _mm_setr_ps(-100.0, 20.0, 0.0, -5.0); |
178 | let r = _mm_addsub_ps(a, b); |
179 | assert_eq_m128(r, _mm_setr_ps(99.0, 25.0, 0.0, -15.0)); |
180 | } |
181 | |
182 | #[simd_test(enable = "sse3" )] |
183 | unsafe fn test_mm_addsub_pd() { |
184 | let a = _mm_setr_pd(-1.0, 5.0); |
185 | let b = _mm_setr_pd(-100.0, 20.0); |
186 | let r = _mm_addsub_pd(a, b); |
187 | assert_eq_m128d(r, _mm_setr_pd(99.0, 25.0)); |
188 | } |
189 | |
190 | #[simd_test(enable = "sse3" )] |
191 | unsafe fn test_mm_hadd_pd() { |
192 | let a = _mm_setr_pd(-1.0, 5.0); |
193 | let b = _mm_setr_pd(-100.0, 20.0); |
194 | let r = _mm_hadd_pd(a, b); |
195 | assert_eq_m128d(r, _mm_setr_pd(4.0, -80.0)); |
196 | } |
197 | |
198 | #[simd_test(enable = "sse3" )] |
199 | unsafe fn test_mm_hadd_ps() { |
200 | let a = _mm_setr_ps(-1.0, 5.0, 0.0, -10.0); |
201 | let b = _mm_setr_ps(-100.0, 20.0, 0.0, -5.0); |
202 | let r = _mm_hadd_ps(a, b); |
203 | assert_eq_m128(r, _mm_setr_ps(4.0, -10.0, -80.0, -5.0)); |
204 | } |
205 | |
206 | #[simd_test(enable = "sse3" )] |
207 | unsafe fn test_mm_hsub_pd() { |
208 | let a = _mm_setr_pd(-1.0, 5.0); |
209 | let b = _mm_setr_pd(-100.0, 20.0); |
210 | let r = _mm_hsub_pd(a, b); |
211 | assert_eq_m128d(r, _mm_setr_pd(-6.0, -120.0)); |
212 | } |
213 | |
214 | #[simd_test(enable = "sse3" )] |
215 | unsafe fn test_mm_hsub_ps() { |
216 | let a = _mm_setr_ps(-1.0, 5.0, 0.0, -10.0); |
217 | let b = _mm_setr_ps(-100.0, 20.0, 0.0, -5.0); |
218 | let r = _mm_hsub_ps(a, b); |
219 | assert_eq_m128(r, _mm_setr_ps(-6.0, 10.0, -120.0, 5.0)); |
220 | } |
221 | |
222 | #[simd_test(enable = "sse3" )] |
223 | unsafe fn test_mm_lddqu_si128() { |
224 | #[rustfmt::skip] |
225 | let a = _mm_setr_epi8( |
226 | 1, 2, 3, 4, |
227 | 5, 6, 7, 8, |
228 | 9, 10, 11, 12, |
229 | 13, 14, 15, 16, |
230 | ); |
231 | let r = _mm_lddqu_si128(&a); |
232 | assert_eq_m128i(a, r); |
233 | } |
234 | |
235 | #[simd_test(enable = "sse3" )] |
236 | unsafe fn test_mm_movedup_pd() { |
237 | let a = _mm_setr_pd(-1.0, 5.0); |
238 | let r = _mm_movedup_pd(a); |
239 | assert_eq_m128d(r, _mm_setr_pd(-1.0, -1.0)); |
240 | } |
241 | |
242 | #[simd_test(enable = "sse3" )] |
243 | unsafe fn test_mm_movehdup_ps() { |
244 | let a = _mm_setr_ps(-1.0, 5.0, 0.0, -10.0); |
245 | let r = _mm_movehdup_ps(a); |
246 | assert_eq_m128(r, _mm_setr_ps(5.0, 5.0, -10.0, -10.0)); |
247 | } |
248 | |
249 | #[simd_test(enable = "sse3" )] |
250 | unsafe fn test_mm_moveldup_ps() { |
251 | let a = _mm_setr_ps(-1.0, 5.0, 0.0, -10.0); |
252 | let r = _mm_moveldup_ps(a); |
253 | assert_eq_m128(r, _mm_setr_ps(-1.0, -1.0, 0.0, 0.0)); |
254 | } |
255 | |
256 | #[simd_test(enable = "sse3" )] |
257 | unsafe fn test_mm_loaddup_pd() { |
258 | let d = -5.0; |
259 | let r = _mm_loaddup_pd(&d); |
260 | assert_eq_m128d(r, _mm_setr_pd(d, d)); |
261 | } |
262 | } |
263 | |