| /dports/math/blasfeo/blasfeo-0.1.2/kernel/avx/ |
| H A D | kernel_dgecp_lib4.c | 385 c_0 = _mm_mul_pd( _mm256_castpd256_pd128( alpha_0 ), c_0 ); in kernel_dgecpsc_8_0_lib4()
388 c_0 = _mm_mul_pd( _mm256_castpd256_pd128( alpha_0 ), c_0 ); in kernel_dgecpsc_8_0_lib4()
395 c_0 = _mm_mul_pd( _mm256_castpd256_pd128( alpha_0 ), c_0 ); in kernel_dgecpsc_8_0_lib4()
402 c_0 = _mm_mul_pd( _mm256_castpd256_pd128( alpha_0 ), c_0 ); in kernel_dgecpsc_8_0_lib4()
413 c_0 = _mm_mul_pd( _mm256_castpd256_pd128( alpha_0 ), c_0 ); in kernel_dgecpsc_8_0_lib4()
416 c_0 = _mm_mul_pd( _mm256_castpd256_pd128( alpha_0 ), c_0 ); in kernel_dgecpsc_8_0_lib4()
420 c_0 = _mm_mul_pd( _mm256_castpd256_pd128( alpha_0 ), c_0 ); in kernel_dgecpsc_8_0_lib4()
3266 _mm_store_sd( &A[5], _mm256_castpd256_pd128( a0 ) ); in kernel_dtrset_4_lib4()
3267 _mm_store_pd( &A[6], _mm256_castpd256_pd128( a0 ) ); in kernel_dtrset_4_lib4()
3269 _mm_store_pd( &A[10], _mm256_castpd256_pd128( a0 ) ); in kernel_dtrset_4_lib4()
[all …]
|
| H A D | kernel_dgetrf_pivot_lib.c | 188 inv = _mm_div_pd( _mm256_castpd256_pd128( ones ), inv ); in kernel_dgetrf_pivot_8_lib()
392 inv = _mm_div_pd( _mm256_castpd256_pd128( ones ), inv ); in kernel_dgetrf_pivot_8_lib()
581 inv = _mm_div_pd( _mm256_castpd256_pd128( ones ), inv ); in kernel_dgetrf_pivot_8_lib()
754 inv = _mm_div_pd( _mm256_castpd256_pd128( ones ), inv ); in kernel_dgetrf_pivot_8_lib()
912 inv = _mm_div_pd( _mm256_castpd256_pd128( ones ), inv ); in kernel_dgetrf_pivot_8_lib()
1053 inv = _mm_div_pd( _mm256_castpd256_pd128( ones ), inv ); in kernel_dgetrf_pivot_8_lib()
1178 inv = _mm_div_pd( _mm256_castpd256_pd128( ones ), inv ); in kernel_dgetrf_pivot_8_lib()
1287 inv = _mm_div_pd( _mm256_castpd256_pd128( ones ), inv ); in kernel_dgetrf_pivot_8_lib()
1543 inv = _mm_div_pd( _mm256_castpd256_pd128( ones ), inv ); in kernel_dgetrf_pivot_4_lib()
1683 inv = _mm_div_pd( _mm256_castpd256_pd128( ones ), inv ); in kernel_dgetrf_pivot_4_lib()
[all …]
|
| /dports/math/blis/blis-0.8.1/kernels/zen/3/sup/broken/ |
| H A D | bli_gemmsup_rv_zen_asm_z3x4n.c | 261 _mm_storeu_pd((double *)(tC ), _mm256_castpd256_pd128(ymm4)); in bli_zgemmsup_rv_zen_asm_3x4n()
262 _mm_storeu_pd((double *)(tC+1), _mm256_castpd256_pd128(ymm8)); in bli_zgemmsup_rv_zen_asm_3x4n()
272 _mm_storeu_pd((double *)(tC ), _mm256_castpd256_pd128(ymm5)); in bli_zgemmsup_rv_zen_asm_3x4n()
273 _mm_storeu_pd((double *)(tC+1), _mm256_castpd256_pd128(ymm9)); in bli_zgemmsup_rv_zen_asm_3x4n()
643 _mm_storeu_pd((double *)(tC), _mm256_castpd256_pd128(ymm4)); in bli_zgemmsup_rv_zen_asm_2x4n()
672 _mm_storeu_pd((double *)(tC), _mm256_castpd256_pd128(ymm5)); in bli_zgemmsup_rv_zen_asm_2x4n()
878 _mm_storeu_pd((double *)(tC), _mm256_castpd256_pd128(ymm4)); in bli_zgemmsup_rv_zen_asm_1x4n()
885 _mm_storeu_pd((double *)(tC), _mm256_castpd256_pd128(ymm5)); in bli_zgemmsup_rv_zen_asm_1x4n()
903 _mm_storeu_pd((double *)(tC), _mm256_castpd256_pd128(ymm4)); in bli_zgemmsup_rv_zen_asm_1x4n()
1110 _mm_storeu_pd((double *)(tC), _mm256_castpd256_pd128(ymm4)); in bli_zgemmsup_rv_zen_asm_3x2()
[all …]
|
| /dports/science/gromacs/gromacs-2021.4/src/gromacs/simd/impl_x86_avx_256/ |
| H A D | impl_x86_avx_256_util_double.h | 167 _mm_storeu_pd(base + align * offset[0], _mm256_castpd256_pd128(t0)); in transposeScatterStoreU()
168 _mm_storeu_pd(base + align * offset[1], _mm256_castpd256_pd128(t1)); in transposeScatterStoreU()
212 tA = _mm_add_pd(tA, _mm256_castpd256_pd128(t0)); in transposeScatterIncrU()
213 tB = _mm_add_pd(tB, _mm256_castpd256_pd128(v2.simdInternal_)); in transposeScatterIncrU()
219 tA = _mm_add_pd(tA, _mm256_castpd256_pd128(t1)); in transposeScatterIncrU()
220 tB = _mm_add_pd(tB, _mm256_castpd256_pd128(v2.simdInternal_)); in transposeScatterIncrU()
274 tA = _mm_sub_pd(tA, _mm256_castpd256_pd128(t0)); in transposeScatterDecrU()
275 tB = _mm_sub_pd(tB, _mm256_castpd256_pd128(v2.simdInternal_)); in transposeScatterDecrU()
281 tA = _mm_sub_pd(tA, _mm256_castpd256_pd128(t1)); in transposeScatterDecrU()
282 tB = _mm_sub_pd(tB, _mm256_castpd256_pd128(v2.simdInternal_)); in transposeScatterDecrU()
[all …]
|
| H A D | impl_x86_avx_256_simd4_double.h | 207 tmp1 = _mm256_castpd256_pd128(a.simdInternal_); in dotProduct()
283 a0 = _mm256_castpd256_pd128(a.simdInternal_); in reduce()
|
| /dports/multimedia/gstreamer1-plugins-rust/gst-plugins-rs-d0466b3eee114207f851b37cae0015c0e718f021/cargo-crates/rustfft-5.1.1/src/avx/ |
| H A D | avx64_utils.rs | 31 _mm256_castpd256_pd128(rows1[0]),
135 _mm256_castpd256_pd128(rows1[0]),
137 _mm256_castpd256_pd128(rows2[0]),
139 _mm256_castpd256_pd128(rows3[0]),
141 _mm256_castpd256_pd128(rows4[0]),
|
| /dports/devel/libsimdpp/libsimdpp-2.1-400-g9dac213d/simdpp/detail/shuffle/ |
| H A D | sse_float64_4x2.h | 151 __m128d a1 = _mm256_castpd256_pd128(a.native());
152 __m128d b1 = _mm256_castpd256_pd128(b.native());
171 __m128d a1 = _mm256_castpd256_pd128(a.native());
172 __m128d b1 = _mm256_castpd256_pd128(b.native());
|
| /dports/devel/inastemp/inastemp-0.2.4/Src/AVX/ |
| H A D | InaVecAVXDouble.hpp | 251 const __m128d rest = _mm256_castpd256_pd128(vec); in horizontalSum()
259 const __m128d rest = _mm256_castpd256_pd128(vec); in horizontalMul()
305 __m128d vallower = _mm256_castpd256_pd128(x); in exp()
341 __m128d vallower = _mm256_castpd256_pd128(x); in expLowAcc()
601 __m128d vallower = _mm256_castpd256_pd128(val_a01_b01_a23_b23); in MultiHorizontalSum()
|
| /dports/audio/libsoxr/soxr-0.1.3-Source/src/ |
| H A D | dev64s.h | 36 _mm_store_sd(a, _mm256_castpd256_pd128(_mm256_hadd_pd(b,b)));} in vStorSum()
|
| /dports/audio/audacity/audacity-Audacity-3.1.3/lib-src/libsoxr/src/ |
| H A D | dev64s.h | 36 _mm_store_sd(a, _mm256_castpd256_pd128(_mm256_hadd_pd(b,b)));} in vStorSum()
|
| /dports/graphics/povray38/povunix-v3.8.0-beta.2-src/platform/x86/avx2fma3/ |
| H A D | avx2fma3noise.cpp | 273 sumr_up = _mm_add_pd(_mm256_castpd256_pd128(sumr),sumr_up); in AVX2FMA3Noise()
472 xy_up = _mm_add_pd(_mm256_castpd256_pd128(xy),xy_up); in AVX2FMA3DNoise()
476 z_up = _mm_add_pd(_mm256_castpd256_pd128(z),z_up); in AVX2FMA3DNoise()
|
| /dports/audio/gogglesmm/gogglesmm-1.2.2/cfox/lib/ |
| H A D | FXMat3d.cpp | 55 #define _mm256_storeu_sd(p,x) _mm_storel_pd(p,_mm256_castpd256_pd128(x))
58 #define _mm256_store_sd(p,x) _mm_storel_pd(p,_mm256_castpd256_pd128(x))
61 …56_div_sd(a,b) _mm256_castpd128_pd256(_mm_div_sd(_mm256_castpd256_pd128(a),_mm256_castpd256_pd1…
62 …56_mul_sd(a,b) _mm256_castpd128_pd256(_mm_mul_sd(_mm256_castpd256_pd128(a),_mm256_castpd256_pd1…
63 …56_add_sd(a,b) _mm256_castpd128_pd256(_mm_add_sd(_mm256_castpd256_pd128(a),_mm256_castpd256_pd1…
64 …56_sub_sd(a,b) _mm256_castpd128_pd256(_mm_sub_sd(_mm256_castpd256_pd128(a),_mm256_castpd256_pd1…
|
| /dports/x11-toolkits/fox17/fox-1.7.77/lib/ |
| H A D | FXMat3d.cpp | 55 #define _mm256_storeu_sd(p,x) _mm_storel_pd(p,_mm256_castpd256_pd128(x))
58 #define _mm256_store_sd(p,x) _mm_storel_pd(p,_mm256_castpd256_pd128(x))
61 …56_div_sd(a,b) _mm256_castpd128_pd256(_mm_div_sd(_mm256_castpd256_pd128(a),_mm256_castpd256_pd1…
62 …56_mul_sd(a,b) _mm256_castpd128_pd256(_mm_mul_sd(_mm256_castpd256_pd128(a),_mm256_castpd256_pd1…
63 …56_add_sd(a,b) _mm256_castpd128_pd256(_mm_add_sd(_mm256_castpd256_pd128(a),_mm256_castpd256_pd1…
64 …56_sub_sd(a,b) _mm256_castpd128_pd256(_mm_sub_sd(_mm256_castpd256_pd128(a),_mm256_castpd256_pd1…
|
| /dports/graphics/povray38/povunix-v3.8.0-beta.2-src/platform/x86/avx/ |
| H A D | avxnoise.cpp | 286 sumr_up = _mm_add_pd(_mm256_castpd256_pd128(sumr),sumr_up); in AVXNoise()
485 xy_up = _mm_add_pd(_mm256_castpd256_pd128(xy),xy_up); in AVXDNoise()
489 z_up = _mm_add_pd(_mm256_castpd256_pd128(z),z_up); in AVXDNoise()
|
| /dports/science/gromacs/gromacs-2021.4/src/gromacs/simd/impl_x86_avx_128_fma/ |
| H A D | impl_x86_avx_128_fma_simd4_double.h | 200 tmp1 = _mm256_castpd256_pd128(a.simdInternal_); in dotProduct()
276 a0 = _mm256_castpd256_pd128(a.simdInternal_); in reduce()
|
| /dports/science/gromacs/gromacs-2021.4/src/gromacs/simd/impl_x86_avx_512/ |
| H A D | impl_x86_avx_512_simd4_double.h | 200 tmp1 = _mm256_castpd256_pd128(a.simdInternal_); in dotProduct()
284 a0 = _mm256_castpd256_pd128(a.simdInternal_); in reduce()
|
| /dports/math/openblas/OpenBLAS-0.3.18/kernel/simd/ |
| H A D | intrin_avx.h | 54 __m128d lo = _mm256_castpd256_pd128(sum_halves); in v_sum_f64()
|
| /dports/math/py-numpy/numpy-1.20.3/numpy/core/src/common/simd/avx2/ |
| H A D | memory.h | 67 #define npyv_storel_f64(PTR, VEC) _mm_storeu_pd(PTR, _mm256_castpd256_pd128(VEC)) in NPYV_IMPL_AVX2_MEM_INT()
134 __m128d a0 = _mm256_castpd256_pd128(a); in npyv_storen_f64()
278 __m128d a0 = _mm256_castpd256_pd128(_mm256_castsi256_pd(a)); in npyv_storen_till_s64()
|
| /dports/math/blasfeo/blasfeo-0.1.2/kernel/avx2/ |
| H A D | kernel_dgetrf_pivot_lib.c | 188 inv = _mm_div_pd( _mm256_castpd256_pd128( ones ), inv ); in kernel_dgetrf_pivot_12_lib()
423 inv = _mm_div_pd( _mm256_castpd256_pd128( ones ), inv ); in kernel_dgetrf_pivot_12_lib()
646 inv = _mm_div_pd( _mm256_castpd256_pd128( ones ), inv ); in kernel_dgetrf_pivot_12_lib()
856 inv = _mm_div_pd( _mm256_castpd256_pd128( ones ), inv ); in kernel_dgetrf_pivot_12_lib()
1054 inv = _mm_div_pd( _mm256_castpd256_pd128( ones ), inv ); in kernel_dgetrf_pivot_12_lib()
1238 inv = _mm_div_pd( _mm256_castpd256_pd128( ones ), inv ); in kernel_dgetrf_pivot_12_lib()
1409 inv = _mm_div_pd( _mm256_castpd256_pd128( ones ), inv ); in kernel_dgetrf_pivot_12_lib()
1567 inv = _mm_div_pd( _mm256_castpd256_pd128( ones ), inv ); in kernel_dgetrf_pivot_12_lib()
1714 inv = _mm_div_pd( _mm256_castpd256_pd128( ones ), inv ); in kernel_dgetrf_pivot_12_lib()
1846 inv = _mm_div_pd( _mm256_castpd256_pd128( ones ), inv ); in kernel_dgetrf_pivot_12_lib()
[all …]
|
| /dports/lang/gcc10/gcc-10.3.0/gcc/testsuite/gcc.target/i386/ |
| H A D | avx-typecast-1.c | 60 fd.x = _mm256_castpd256_pd128 (c.x); in avx_test()
|
| /dports/lang/gcc11-devel/gcc-11-20211009/gcc/testsuite/gcc.target/i386/ |
| H A D | avx-typecast-1.c | 60 fd.x = _mm256_castpd256_pd128 (c.x); in avx_test()
|
| /dports/lang/gcc11/gcc-11.2.0/gcc/testsuite/gcc.target/i386/ |
| H A D | avx-typecast-1.c | 60 fd.x = _mm256_castpd256_pd128 (c.x); in avx_test()
|
| /dports/devel/avr-gcc/gcc-10.2.0/gcc/testsuite/gcc.target/i386/ |
| H A D | avx-typecast-1.c | 60 fd.x = _mm256_castpd256_pd128 (c.x); in avx_test()
|
| /dports/misc/cxx_atomics_pic/gcc-11.2.0/gcc/testsuite/gcc.target/i386/ |
| H A D | avx-typecast-1.c | 60 fd.x = _mm256_castpd256_pd128 (c.x); in avx_test()
|
| /dports/lang/gcc10-devel/gcc-10-20211008/gcc/testsuite/gcc.target/i386/ |
| H A D | avx-typecast-1.c | 60 fd.x = _mm256_castpd256_pd128 (c.x); in avx_test()
|