| 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 | #[rustc_const_unstable (feature = "stdarch_const_x86" , issue = "149298" )] |
| 18 | pub const fn _mm_addsub_ps(a: __m128, b: __m128) -> __m128 { |
| 19 | unsafe { |
| 20 | let a: Simd = a.as_f32x4(); |
| 21 | let b: Simd = b.as_f32x4(); |
| 22 | let add: Simd = simd_add(x:a, y:b); |
| 23 | let sub: Simd = simd_sub(lhs:a, rhs:b); |
| 24 | simd_shuffle!(add, sub, [4, 1, 6, 3]) |
| 25 | } |
| 26 | } |
| 27 | |
| 28 | /// Alternatively add and subtract packed double-precision (64-bit) |
| 29 | /// floating-point elements in `a` to/from packed elements in `b`. |
| 30 | /// |
| 31 | /// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_addsub_pd) |
| 32 | #[inline ] |
| 33 | #[target_feature (enable = "sse3" )] |
| 34 | #[cfg_attr (test, assert_instr(addsubpd))] |
| 35 | #[stable (feature = "simd_x86" , since = "1.27.0" )] |
| 36 | #[rustc_const_unstable (feature = "stdarch_const_x86" , issue = "149298" )] |
| 37 | pub const fn _mm_addsub_pd(a: __m128d, b: __m128d) -> __m128d { |
| 38 | unsafe { |
| 39 | let a: Simd = a.as_f64x2(); |
| 40 | let b: Simd = b.as_f64x2(); |
| 41 | let add: Simd = simd_add(x:a, y:b); |
| 42 | let sub: Simd = simd_sub(lhs:a, rhs:b); |
| 43 | simd_shuffle!(add, sub, [2, 1]) |
| 44 | } |
| 45 | } |
| 46 | |
| 47 | /// Horizontally adds adjacent pairs of double-precision (64-bit) |
| 48 | /// floating-point elements in `a` and `b`, and pack the results. |
| 49 | /// |
| 50 | /// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_hadd_pd) |
| 51 | #[inline ] |
| 52 | #[target_feature (enable = "sse3" )] |
| 53 | #[cfg_attr (test, assert_instr(haddpd))] |
| 54 | #[stable (feature = "simd_x86" , since = "1.27.0" )] |
| 55 | #[rustc_const_unstable (feature = "stdarch_const_x86" , issue = "149298" )] |
| 56 | pub const fn _mm_hadd_pd(a: __m128d, b: __m128d) -> __m128d { |
| 57 | unsafe { |
| 58 | let even: __m128d = simd_shuffle!(a, b, [0, 2]); |
| 59 | let odd: __m128d = simd_shuffle!(a, b, [1, 3]); |
| 60 | simd_add(x:even, y:odd) |
| 61 | } |
| 62 | } |
| 63 | |
| 64 | /// Horizontally adds adjacent pairs of single-precision (32-bit) |
| 65 | /// floating-point elements in `a` and `b`, and pack the results. |
| 66 | /// |
| 67 | /// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_hadd_ps) |
| 68 | #[inline ] |
| 69 | #[target_feature (enable = "sse3" )] |
| 70 | #[cfg_attr (test, assert_instr(haddps))] |
| 71 | #[stable (feature = "simd_x86" , since = "1.27.0" )] |
| 72 | #[rustc_const_unstable (feature = "stdarch_const_x86" , issue = "149298" )] |
| 73 | pub const fn _mm_hadd_ps(a: __m128, b: __m128) -> __m128 { |
| 74 | unsafe { |
| 75 | let even: __m128 = simd_shuffle!(a, b, [0, 2, 4, 6]); |
| 76 | let odd: __m128 = simd_shuffle!(a, b, [1, 3, 5, 7]); |
| 77 | simd_add(x:even, y:odd) |
| 78 | } |
| 79 | } |
| 80 | |
| 81 | /// Horizontally subtract adjacent pairs of double-precision (64-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_pd) |
| 85 | #[inline ] |
| 86 | #[target_feature (enable = "sse3" )] |
| 87 | #[cfg_attr (test, assert_instr(hsubpd))] |
| 88 | #[stable (feature = "simd_x86" , since = "1.27.0" )] |
| 89 | #[rustc_const_unstable (feature = "stdarch_const_x86" , issue = "149298" )] |
| 90 | pub const fn _mm_hsub_pd(a: __m128d, b: __m128d) -> __m128d { |
| 91 | unsafe { |
| 92 | let even: __m128d = simd_shuffle!(a, b, [0, 2]); |
| 93 | let odd: __m128d = simd_shuffle!(a, b, [1, 3]); |
| 94 | simd_sub(lhs:even, rhs:odd) |
| 95 | } |
| 96 | } |
| 97 | |
| 98 | /// Horizontally adds adjacent pairs of single-precision (32-bit) |
| 99 | /// floating-point elements in `a` and `b`, and pack the results. |
| 100 | /// |
| 101 | /// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_hsub_ps) |
| 102 | #[inline ] |
| 103 | #[target_feature (enable = "sse3" )] |
| 104 | #[cfg_attr (test, assert_instr(hsubps))] |
| 105 | #[stable (feature = "simd_x86" , since = "1.27.0" )] |
| 106 | #[rustc_const_unstable (feature = "stdarch_const_x86" , issue = "149298" )] |
| 107 | pub const fn _mm_hsub_ps(a: __m128, b: __m128) -> __m128 { |
| 108 | unsafe { |
| 109 | let even: __m128 = simd_shuffle!(a, b, [0, 2, 4, 6]); |
| 110 | let odd: __m128 = simd_shuffle!(a, b, [1, 3, 5, 7]); |
| 111 | simd_sub(lhs:even, rhs:odd) |
| 112 | } |
| 113 | } |
| 114 | |
| 115 | /// Loads 128-bits of integer data from unaligned memory. |
| 116 | /// This intrinsic may perform better than `_mm_loadu_si128` |
| 117 | /// when the data crosses a cache line boundary. |
| 118 | /// |
| 119 | /// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_lddqu_si128) |
| 120 | #[inline ] |
| 121 | #[target_feature (enable = "sse3" )] |
| 122 | #[cfg_attr (test, assert_instr(lddqu))] |
| 123 | #[stable (feature = "simd_x86" , since = "1.27.0" )] |
| 124 | pub unsafe fn _mm_lddqu_si128(mem_addr: *const __m128i) -> __m128i { |
| 125 | transmute(src:lddqu(mem_addr as *const _)) |
| 126 | } |
| 127 | |
| 128 | /// Duplicate the low double-precision (64-bit) floating-point element |
| 129 | /// from `a`. |
| 130 | /// |
| 131 | /// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_movedup_pd) |
| 132 | #[inline ] |
| 133 | #[target_feature (enable = "sse3" )] |
| 134 | #[cfg_attr (test, assert_instr(movddup))] |
| 135 | #[stable (feature = "simd_x86" , since = "1.27.0" )] |
| 136 | #[rustc_const_unstable (feature = "stdarch_const_x86" , issue = "149298" )] |
| 137 | pub const fn _mm_movedup_pd(a: __m128d) -> __m128d { |
| 138 | unsafe { simd_shuffle!(a, a, [0, 0]) } |
| 139 | } |
| 140 | |
| 141 | /// Loads a double-precision (64-bit) floating-point element from memory |
| 142 | /// into both elements of return vector. |
| 143 | /// |
| 144 | /// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_loaddup_pd) |
| 145 | #[inline ] |
| 146 | #[target_feature (enable = "sse3" )] |
| 147 | #[cfg_attr (test, assert_instr(movddup))] |
| 148 | #[stable (feature = "simd_x86" , since = "1.27.0" )] |
| 149 | #[rustc_const_unstable (feature = "stdarch_const_x86" , issue = "149298" )] |
| 150 | pub const unsafe fn _mm_loaddup_pd(mem_addr: *const f64) -> __m128d { |
| 151 | _mm_load1_pd(mem_addr) |
| 152 | } |
| 153 | |
| 154 | /// Duplicate odd-indexed single-precision (32-bit) floating-point elements |
| 155 | /// from `a`. |
| 156 | /// |
| 157 | /// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_movehdup_ps) |
| 158 | #[inline ] |
| 159 | #[target_feature (enable = "sse3" )] |
| 160 | #[cfg_attr (test, assert_instr(movshdup))] |
| 161 | #[stable (feature = "simd_x86" , since = "1.27.0" )] |
| 162 | #[rustc_const_unstable (feature = "stdarch_const_x86" , issue = "149298" )] |
| 163 | pub const fn _mm_movehdup_ps(a: __m128) -> __m128 { |
| 164 | unsafe { simd_shuffle!(a, a, [1, 1, 3, 3]) } |
| 165 | } |
| 166 | |
| 167 | /// Duplicate even-indexed single-precision (32-bit) floating-point elements |
| 168 | /// from `a`. |
| 169 | /// |
| 170 | /// [Intel's documentation](https://www.intel.com/content/www/us/en/docs/intrinsics-guide/index.html#text=_mm_moveldup_ps) |
| 171 | #[inline ] |
| 172 | #[target_feature (enable = "sse3" )] |
| 173 | #[cfg_attr (test, assert_instr(movsldup))] |
| 174 | #[stable (feature = "simd_x86" , since = "1.27.0" )] |
| 175 | #[rustc_const_unstable (feature = "stdarch_const_x86" , issue = "149298" )] |
| 176 | pub const fn _mm_moveldup_ps(a: __m128) -> __m128 { |
| 177 | unsafe { simd_shuffle!(a, a, [0, 0, 2, 2]) } |
| 178 | } |
| 179 | |
| 180 | #[allow (improper_ctypes)] |
| 181 | unsafe extern "C" { |
| 182 | #[link_name = "llvm.x86.sse3.ldu.dq" ] |
| 183 | unsafefn lddqu(mem_addr: *const i8) -> i8x16; |
| 184 | } |
| 185 | |
| 186 | #[cfg (test)] |
| 187 | mod tests { |
| 188 | use crate::core_arch::assert_eq_const as assert_eq; |
| 189 | use stdarch_test::simd_test; |
| 190 | |
| 191 | use crate::core_arch::x86::*; |
| 192 | |
| 193 | #[simd_test(enable = "sse3" )] |
| 194 | const fn test_mm_addsub_ps() { |
| 195 | let a = _mm_setr_ps(-1.0, 5.0, 0.0, -10.0); |
| 196 | let b = _mm_setr_ps(-100.0, 20.0, 0.0, -5.0); |
| 197 | let r = _mm_addsub_ps(a, b); |
| 198 | assert_eq_m128(r, _mm_setr_ps(99.0, 25.0, 0.0, -15.0)); |
| 199 | } |
| 200 | |
| 201 | #[simd_test(enable = "sse3" )] |
| 202 | const fn test_mm_addsub_pd() { |
| 203 | let a = _mm_setr_pd(-1.0, 5.0); |
| 204 | let b = _mm_setr_pd(-100.0, 20.0); |
| 205 | let r = _mm_addsub_pd(a, b); |
| 206 | assert_eq_m128d(r, _mm_setr_pd(99.0, 25.0)); |
| 207 | } |
| 208 | |
| 209 | #[simd_test(enable = "sse3" )] |
| 210 | const fn test_mm_hadd_pd() { |
| 211 | let a = _mm_setr_pd(-1.0, 5.0); |
| 212 | let b = _mm_setr_pd(-100.0, 20.0); |
| 213 | let r = _mm_hadd_pd(a, b); |
| 214 | assert_eq_m128d(r, _mm_setr_pd(4.0, -80.0)); |
| 215 | } |
| 216 | |
| 217 | #[simd_test(enable = "sse3" )] |
| 218 | const fn test_mm_hadd_ps() { |
| 219 | let a = _mm_setr_ps(-1.0, 5.0, 0.0, -10.0); |
| 220 | let b = _mm_setr_ps(-100.0, 20.0, 0.0, -5.0); |
| 221 | let r = _mm_hadd_ps(a, b); |
| 222 | assert_eq_m128(r, _mm_setr_ps(4.0, -10.0, -80.0, -5.0)); |
| 223 | } |
| 224 | |
| 225 | #[simd_test(enable = "sse3" )] |
| 226 | const fn test_mm_hsub_pd() { |
| 227 | let a = _mm_setr_pd(-1.0, 5.0); |
| 228 | let b = _mm_setr_pd(-100.0, 20.0); |
| 229 | let r = _mm_hsub_pd(a, b); |
| 230 | assert_eq_m128d(r, _mm_setr_pd(-6.0, -120.0)); |
| 231 | } |
| 232 | |
| 233 | #[simd_test(enable = "sse3" )] |
| 234 | const fn test_mm_hsub_ps() { |
| 235 | let a = _mm_setr_ps(-1.0, 5.0, 0.0, -10.0); |
| 236 | let b = _mm_setr_ps(-100.0, 20.0, 0.0, -5.0); |
| 237 | let r = _mm_hsub_ps(a, b); |
| 238 | assert_eq_m128(r, _mm_setr_ps(-6.0, 10.0, -120.0, 5.0)); |
| 239 | } |
| 240 | |
| 241 | #[simd_test(enable = "sse3" )] |
| 242 | fn test_mm_lddqu_si128() { |
| 243 | #[rustfmt::skip] |
| 244 | let a = _mm_setr_epi8( |
| 245 | 1, 2, 3, 4, |
| 246 | 5, 6, 7, 8, |
| 247 | 9, 10, 11, 12, |
| 248 | 13, 14, 15, 16, |
| 249 | ); |
| 250 | let r = unsafe { _mm_lddqu_si128(&a) }; |
| 251 | assert_eq_m128i(a, r); |
| 252 | } |
| 253 | |
| 254 | #[simd_test(enable = "sse3" )] |
| 255 | const fn test_mm_movedup_pd() { |
| 256 | let a = _mm_setr_pd(-1.0, 5.0); |
| 257 | let r = _mm_movedup_pd(a); |
| 258 | assert_eq_m128d(r, _mm_setr_pd(-1.0, -1.0)); |
| 259 | } |
| 260 | |
| 261 | #[simd_test(enable = "sse3" )] |
| 262 | const fn test_mm_movehdup_ps() { |
| 263 | let a = _mm_setr_ps(-1.0, 5.0, 0.0, -10.0); |
| 264 | let r = _mm_movehdup_ps(a); |
| 265 | assert_eq_m128(r, _mm_setr_ps(5.0, 5.0, -10.0, -10.0)); |
| 266 | } |
| 267 | |
| 268 | #[simd_test(enable = "sse3" )] |
| 269 | const fn test_mm_moveldup_ps() { |
| 270 | let a = _mm_setr_ps(-1.0, 5.0, 0.0, -10.0); |
| 271 | let r = _mm_moveldup_ps(a); |
| 272 | assert_eq_m128(r, _mm_setr_ps(-1.0, -1.0, 0.0, 0.0)); |
| 273 | } |
| 274 | |
| 275 | #[simd_test(enable = "sse3" )] |
| 276 | const fn test_mm_loaddup_pd() { |
| 277 | let d = -5.0; |
| 278 | let r = unsafe { _mm_loaddup_pd(&d) }; |
| 279 | assert_eq_m128d(r, _mm_setr_pd(d, d)); |
| 280 | } |
| 281 | } |
| 282 | |