/dports/audio/aften/aften-0.0.8/libaften/x86/ |
H A D | x86_sse_mdct_common.c | 68 XMM0 = _mm_sub_ps(XMM0, XMM1); in mdct_butterfly_8() 74 XMM0 = _mm_shuffle_ps(XMM0, XMM0, _MM_SHUFFLE(3,2,3,2)); in mdct_butterfly_8() 82 XMM0 = _mm_add_ps(XMM0, XMM1); in mdct_butterfly_8() 104 XMM0 = _mm_sub_ps(XMM0, XMM2); in mdct_butterfly_16() 112 XMM0 = _mm_shuffle_ps(XMM0, XMM0, _MM_SHUFFLE(2,3,1,1)); in mdct_butterfly_16() 120 XMM0 = _mm_mul_ps(XMM0, XMM2); in mdct_butterfly_16() 124 XMM0 = _mm_add_ps(XMM0, XMM1); in mdct_butterfly_16() 179 XMM0 = _mm_movehdup_ps(XMM0); in mdct_butterfly_32() 188 XMM0 = _mm_shuffle_ps(XMM0, XMM0, _MM_SHUFFLE(3,3,1,1)); in mdct_butterfly_32() 249 XMM0 = _mm_shuffle_ps(XMM0 , XMM0 , _MM_SHUFFLE(3,3,1,1)); in mdct_butterfly_first() [all …]
|
H A D | x86_sse_mdct_common_init.c | 144 XMM0 = _mm_loadl_pi(XMM0, (__m64*)(S+j+2)); in sse_mdct_ctx_init() 146 XMM0 = _mm_loadh_pi(XMM0, (__m64*)(S+i )); in sse_mdct_ctx_init() 153 XMM0 = _mm_shuffle_ps(XMM0, XMM0, _MM_SHUFFLE(2,3,0,1)); in sse_mdct_ctx_init() 155 XMM0 = _mm_xor_ps(XMM0, PCS_RRNN.v); in sse_mdct_ctx_init() 170 XMM0 = _mm_loadl_pi(XMM0, (__m64*)(S+j+2)); in sse_mdct_ctx_init() 172 XMM0 = _mm_loadh_pi(XMM0, (__m64*)(S+i )); in sse_mdct_ctx_init() 179 XMM0 = _mm_shuffle_ps(XMM0, XMM0, _MM_SHUFFLE(2,3,0,1)); in sse_mdct_ctx_init() 181 XMM0 = _mm_xor_ps(XMM0, PCS_NNRR.v); in sse_mdct_ctx_init() 202 XMM0 = _mm_mul_ps(XMM0, pscalem); in sse_mdct_ctx_init() 207 XMM0 = _mm_shuffle_ps(XMM0, XMM0, _MM_SHUFFLE(0,1,2,3)); in sse_mdct_ctx_init() [all …]
|
/dports/math/igraph/igraph-0.9.5/vendor/plfit/ |
H A D | arithmetic_sse_double.h | 111 XMM0 = _mm_sub_pd(XMM0, XMM4); \ 131 XMM0 = _mm_mul_pd(XMM0, XMM7); \ 152 XMM0 = _mm_sub_pd(XMM0, XMM4); \ 241 XMM0 = _mm_add_pd(XMM0, XMM1); \ 243 XMM0 = _mm_add_pd(XMM0, XMM1); \ 263 XMM0 = _mm_add_pd(XMM0, XMM1); \ 265 XMM0 = _mm_add_pd(XMM0, XMM1); \ 266 XMM0 = _mm_sqrt_pd(XMM0); \ 288 XMM0 = _mm_add_pd(XMM0, XMM1); \ 290 XMM0 = _mm_add_pd(XMM0, XMM1); \ [all …]
|
H A D | arithmetic_sse_float.h | 101 XMM0 = _mm_xor_ps(XMM0, XMM4); \ 121 XMM0 = _mm_mul_ps(XMM0, XMM7); \ 142 XMM0 = _mm_sub_ps(XMM0, XMM4); \ 160 XMM0 = _mm_mul_ps(XMM0, XMM7); \ 228 XMM0 = _mm_add_ps(XMM0, XMM2); \ 231 XMM0 = _mm_add_ps(XMM0, XMM1); \ 247 XMM0 = _mm_add_ps(XMM0, XMM2); \ 250 XMM0 = _mm_add_ps(XMM0, XMM1); \ 252 XMM2 = XMM0; \ 276 XMM0 = _mm_add_ps(XMM0, XMM2); \ [all …]
|
/dports/science/InsightToolkit/ITK-5.0.1/Modules/ThirdParty/libLBFGS/src/ |
H A D | arithmetic_sse_double.h | 111 XMM0 = _mm_sub_pd(XMM0, XMM4); \ 131 XMM0 = _mm_mul_pd(XMM0, XMM7); \ 152 XMM0 = _mm_sub_pd(XMM0, XMM4); \ 241 XMM0 = _mm_add_pd(XMM0, XMM1); \ 243 XMM0 = _mm_add_pd(XMM0, XMM1); \ 263 XMM0 = _mm_add_pd(XMM0, XMM1); \ 265 XMM0 = _mm_add_pd(XMM0, XMM1); \ 266 XMM0 = _mm_sqrt_pd(XMM0); \ 288 XMM0 = _mm_add_pd(XMM0, XMM1); \ 290 XMM0 = _mm_add_pd(XMM0, XMM1); \ [all …]
|
H A D | arithmetic_sse_float.h | 108 XMM0 = _mm_xor_ps(XMM0, XMM4); \ 128 XMM0 = _mm_mul_ps(XMM0, XMM7); \ 149 XMM0 = _mm_sub_ps(XMM0, XMM4); \ 167 XMM0 = _mm_mul_ps(XMM0, XMM7); \ 235 XMM0 = _mm_add_ps(XMM0, XMM2); \ 238 XMM0 = _mm_add_ps(XMM0, XMM1); \ 254 XMM0 = _mm_add_ps(XMM0, XMM2); \ 257 XMM0 = _mm_add_ps(XMM0, XMM1); \ 259 XMM2 = XMM0; \ 283 XMM0 = _mm_add_ps(XMM0, XMM2); \ [all …]
|
/dports/math/R-cran-igraph/igraph/src/plfit/ |
H A D | arithmetic_sse_double.h | 111 XMM0 = _mm_sub_pd(XMM0, XMM4); \ 131 XMM0 = _mm_mul_pd(XMM0, XMM7); \ 152 XMM0 = _mm_sub_pd(XMM0, XMM4); \ 241 XMM0 = _mm_add_pd(XMM0, XMM1); \ 243 XMM0 = _mm_add_pd(XMM0, XMM1); \ 263 XMM0 = _mm_add_pd(XMM0, XMM1); \ 265 XMM0 = _mm_add_pd(XMM0, XMM1); \ 266 XMM0 = _mm_sqrt_pd(XMM0); \ 288 XMM0 = _mm_add_pd(XMM0, XMM1); \ 290 XMM0 = _mm_add_pd(XMM0, XMM1); \ [all …]
|
H A D | arithmetic_sse_float.h | 101 XMM0 = _mm_xor_ps(XMM0, XMM4); \ 121 XMM0 = _mm_mul_ps(XMM0, XMM7); \ 142 XMM0 = _mm_sub_ps(XMM0, XMM4); \ 160 XMM0 = _mm_mul_ps(XMM0, XMM7); \ 228 XMM0 = _mm_add_ps(XMM0, XMM2); \ 231 XMM0 = _mm_add_ps(XMM0, XMM1); \ 247 XMM0 = _mm_add_ps(XMM0, XMM2); \ 250 XMM0 = _mm_add_ps(XMM0, XMM1); \ 252 XMM2 = XMM0; \ 276 XMM0 = _mm_add_ps(XMM0, XMM2); \ [all …]
|
/dports/emulators/dolphin-emu/dolphin-3152428/Source/Core/Core/PowerPC/Jit64Common/ |
H A D | Jit64AsmCommon.cpp | 197 XORPD(XMM0, R(XMM0)); in GenFrsqrte() 502 CVTTPS2DQ(XMM0, R(XMM0)); in GenQuantizedStore() 507 PACKSSDW(XMM0, R(XMM0)); in GenQuantizedStore() 508 PACKUSWB(XMM0, R(XMM0)); in GenQuantizedStore() 512 PACKSSDW(XMM0, R(XMM0)); in GenQuantizedStore() 513 PACKSSWB(XMM0, R(XMM0)); in GenQuantizedStore() 533 PACKSSDW(XMM0, R(XMM0)); in GenQuantizedStore() 655 PMOVZXBD(XMM0, R(XMM0)); in GenQuantizedLoad() 668 PMOVSXBD(XMM0, R(XMM0)); in GenQuantizedLoad() 682 PMOVZXWD(XMM0, R(XMM0)); in GenQuantizedLoad() [all …]
|
/dports/science/PETSc/petsc-3.14.1/src/mat/impls/baij/seq/ |
H A D | dgefa4.c | 209 SSE_SUB_PS(XMM1,XMM0) in PetscKernel_A_gets_inverse_A_4_SSE() 222 SSE_SHUFFLE(XMM0,XMM0,0xB1) in PetscKernel_A_gets_inverse_A_4_SSE() 231 SSE_SHUFFLE(XMM0,XMM0,0x4E) in PetscKernel_A_gets_inverse_A_4_SSE() 238 SSE_SUB_PS(XMM0,XMM7) in PetscKernel_A_gets_inverse_A_4_SSE() 240 SSE_SHUFFLE(XMM0,XMM0,0x4E) in PetscKernel_A_gets_inverse_A_4_SSE() 255 SSE_ADD_PS(XMM0,XMM2) in PetscKernel_A_gets_inverse_A_4_SSE() 264 SSE_SUB_PS(XMM0,XMM1) in PetscKernel_A_gets_inverse_A_4_SSE() 281 SSE_ADD_PS(XMM0,XMM7) in PetscKernel_A_gets_inverse_A_4_SSE() 291 SSE_SUB_PS(XMM7,XMM0) in PetscKernel_A_gets_inverse_A_4_SSE() 307 SSE_SUB_PS(XMM0,XMM2) in PetscKernel_A_gets_inverse_A_4_SSE() [all …]
|
/dports/emulators/ppsspp-qt5/ppsspp-1.12.3/Core/MIPS/x86/ |
H A D | CompVFPU.cpp | 535 XORPS(XMM0, R(XMM0)); in Comp_VVectorInit() 583 XORPS(XMM0, R(XMM0)); in Comp_VIdt() 1321 XORPS(XMM0, R(XMM0)); in Comp_Vcmp() 1736 PXOR(XMM0, R(XMM0)); in Comp_Vx2i() 1748 PUNPCKLBW(XMM0, R(XMM0)); in Comp_Vx2i() 1750 PUNPCKLWD(XMM0, R(XMM0)); in Comp_Vx2i() 1760 PXOR(XMM0, R(XMM0)); in Comp_Vx2i() 1942 SHUFPS(XMM0, R(XMM0), _MM_SHUFFLE(0,0,0,0)); in Comp_Vcst() 1989 XORPS(XMM0, R(XMM0)); in Comp_Vsgn() 2622 XORPS(XMM0, R(XMM0)); in Comp_VMatrixInit() [all …]
|
/dports/emulators/ppsspp/ppsspp-1.12.3/Core/MIPS/x86/ |
H A D | CompVFPU.cpp | 535 XORPS(XMM0, R(XMM0)); in Comp_VVectorInit() 583 XORPS(XMM0, R(XMM0)); in Comp_VIdt() 1321 XORPS(XMM0, R(XMM0)); in Comp_Vcmp() 1736 PXOR(XMM0, R(XMM0)); in Comp_Vx2i() 1748 PUNPCKLBW(XMM0, R(XMM0)); in Comp_Vx2i() 1750 PUNPCKLWD(XMM0, R(XMM0)); in Comp_Vx2i() 1760 PXOR(XMM0, R(XMM0)); in Comp_Vx2i() 1942 SHUFPS(XMM0, R(XMM0), _MM_SHUFFLE(0,0,0,0)); in Comp_Vcst() 1989 XORPS(XMM0, R(XMM0)); in Comp_Vsgn() 2622 XORPS(XMM0, R(XMM0)); in Comp_VMatrixInit() [all …]
|
/dports/emulators/libretro-ppsspp/ppsspp-1.12.3/Core/MIPS/x86/ |
H A D | CompVFPU.cpp | 535 XORPS(XMM0, R(XMM0)); in Comp_VVectorInit() 583 XORPS(XMM0, R(XMM0)); in Comp_VIdt() 1321 XORPS(XMM0, R(XMM0)); in Comp_Vcmp() 1736 PXOR(XMM0, R(XMM0)); in Comp_Vx2i() 1748 PUNPCKLBW(XMM0, R(XMM0)); in Comp_Vx2i() 1750 PUNPCKLWD(XMM0, R(XMM0)); in Comp_Vx2i() 1760 PXOR(XMM0, R(XMM0)); in Comp_Vx2i() 1942 SHUFPS(XMM0, R(XMM0), _MM_SHUFFLE(0,0,0,0)); in Comp_Vcst() 1989 XORPS(XMM0, R(XMM0)); in Comp_Vsgn() 2622 XORPS(XMM0, R(XMM0)); in Comp_VMatrixInit() [all …]
|
/dports/math/openblas/OpenBLAS-0.3.18/kernel/x86_64/ |
H A D | dgemm_kernel_6x4_piledriver.S | 240 VXOR1282(XMM0,XMM0,XMM0); 471 VXOR1282(XMM0,XMM0,XMM0); 647 VXOR1282(XMM0,XMM0,XMM0); 772 VXOR1281(XMM0,XMM0,XMM0); 918 VXOR1282(XMM0,XMM0,XMM0); 1069 VXOR1282(XMM0,XMM0,XMM0); 1187 VXOR1282(XMM0,XMM0,XMM0); 1276 VXOR1281(XMM0,XMM0,XMM0); 1381 VXOR1282(XMM0,XMM0,XMM0); 1492 VXOR1282(XMM0,XMM0,XMM0); [all …]
|
/dports/graphics/libosmesa-gallium/mesa-21.3.6/src/mesa/x86/ |
H A D | sse_xform3.S | 151 MOVLPS ( S(0), XMM0 ) 152 MOVLPS ( XMM0, D(0) ) 153 MOVSS ( S(2), XMM0 ) 154 MOVSS ( XMM0, D(2) ) 202 XORPS( XMM0, XMM0 ) /* clean the working register */ 220 MOVSS ( S(2), XMM0 ) /* sz */ 288 MOVSS ( S(2), XMM0 ) /* oz */ 293 MOVSS ( S(2), XMM0 ) /* oz */ 415 MOVLPS( S(0), XMM0 ) /* oy | ox */ 420 MOVSS( S(2), XMM0 ) [all …]
|
/dports/lang/clover/mesa-21.3.6/src/mesa/x86/ |
H A D | sse_xform3.S | 151 MOVLPS ( S(0), XMM0 ) 152 MOVLPS ( XMM0, D(0) ) 153 MOVSS ( S(2), XMM0 ) 154 MOVSS ( XMM0, D(2) ) 202 XORPS( XMM0, XMM0 ) /* clean the working register */ 220 MOVSS ( S(2), XMM0 ) /* sz */ 288 MOVSS ( S(2), XMM0 ) /* oz */ 293 MOVSS ( S(2), XMM0 ) /* oz */ 415 MOVLPS( S(0), XMM0 ) /* oy | ox */ 420 MOVSS( S(2), XMM0 ) [all …]
|
/dports/graphics/libosmesa/mesa-21.3.6/src/mesa/x86/ |
H A D | sse_xform3.S | 151 MOVLPS ( S(0), XMM0 ) 152 MOVLPS ( XMM0, D(0) ) 153 MOVSS ( S(2), XMM0 ) 154 MOVSS ( XMM0, D(2) ) 202 XORPS( XMM0, XMM0 ) /* clean the working register */ 220 MOVSS ( S(2), XMM0 ) /* sz */ 288 MOVSS ( S(2), XMM0 ) /* oz */ 293 MOVSS ( S(2), XMM0 ) /* oz */ 415 MOVLPS( S(0), XMM0 ) /* oy | ox */ 420 MOVSS( S(2), XMM0 ) [all …]
|
/dports/graphics/mesa-libs/mesa-21.3.6/src/mesa/x86/ |
H A D | sse_xform3.S | 151 MOVLPS ( S(0), XMM0 ) 152 MOVLPS ( XMM0, D(0) ) 153 MOVSS ( S(2), XMM0 ) 154 MOVSS ( XMM0, D(2) ) 202 XORPS( XMM0, XMM0 ) /* clean the working register */ 220 MOVSS ( S(2), XMM0 ) /* sz */ 288 MOVSS ( S(2), XMM0 ) /* oz */ 293 MOVSS ( S(2), XMM0 ) /* oz */ 415 MOVLPS( S(0), XMM0 ) /* oy | ox */ 420 MOVSS( S(2), XMM0 ) [all …]
|
/dports/graphics/mesa-dri-gallium/mesa-21.3.6/src/mesa/x86/ |
H A D | sse_xform3.S | 151 MOVLPS ( S(0), XMM0 ) 152 MOVLPS ( XMM0, D(0) ) 153 MOVSS ( S(2), XMM0 ) 154 MOVSS ( XMM0, D(2) ) 202 XORPS( XMM0, XMM0 ) /* clean the working register */ 220 MOVSS ( S(2), XMM0 ) /* sz */ 288 MOVSS ( S(2), XMM0 ) /* oz */ 293 MOVSS ( S(2), XMM0 ) /* oz */ 415 MOVLPS( S(0), XMM0 ) /* oy | ox */ 420 MOVSS( S(2), XMM0 ) [all …]
|
/dports/graphics/mesa-gallium-va/mesa-21.3.6/src/mesa/x86/ |
H A D | sse_xform3.S | 151 MOVLPS ( S(0), XMM0 ) 152 MOVLPS ( XMM0, D(0) ) 153 MOVSS ( S(2), XMM0 ) 154 MOVSS ( XMM0, D(2) ) 202 XORPS( XMM0, XMM0 ) /* clean the working register */ 220 MOVSS ( S(2), XMM0 ) /* sz */ 288 MOVSS ( S(2), XMM0 ) /* oz */ 293 MOVSS ( S(2), XMM0 ) /* oz */ 415 MOVLPS( S(0), XMM0 ) /* oy | ox */ 420 MOVSS( S(2), XMM0 ) [all …]
|
/dports/graphics/mesa-dri-classic/mesa-20.2.3/src/mesa/x86/ |
H A D | sse_xform3.S | 151 MOVLPS ( S(0), XMM0 ) 152 MOVLPS ( XMM0, D(0) ) 153 MOVSS ( S(2), XMM0 ) 154 MOVSS ( XMM0, D(2) ) 202 XORPS( XMM0, XMM0 ) /* clean the working register */ 220 MOVSS ( S(2), XMM0 ) /* sz */ 288 MOVSS ( S(2), XMM0 ) /* oz */ 293 MOVSS ( S(2), XMM0 ) /* oz */ 415 MOVLPS( S(0), XMM0 ) /* oy | ox */ 420 MOVSS( S(2), XMM0 ) [all …]
|
/dports/graphics/mesa-gallium-vdpau/mesa-21.3.6/src/mesa/x86/ |
H A D | sse_xform3.S | 151 MOVLPS ( S(0), XMM0 ) 152 MOVLPS ( XMM0, D(0) ) 153 MOVSS ( S(2), XMM0 ) 154 MOVSS ( XMM0, D(2) ) 202 XORPS( XMM0, XMM0 ) /* clean the working register */ 220 MOVSS ( S(2), XMM0 ) /* sz */ 288 MOVSS ( S(2), XMM0 ) /* oz */ 293 MOVSS ( S(2), XMM0 ) /* oz */ 415 MOVLPS( S(0), XMM0 ) /* oy | ox */ 420 MOVSS( S(2), XMM0 ) [all …]
|
/dports/graphics/mesa-gallium-xa/mesa-21.3.6/src/mesa/x86/ |
H A D | sse_xform3.S | 151 MOVLPS ( S(0), XMM0 ) 152 MOVLPS ( XMM0, D(0) ) 153 MOVSS ( S(2), XMM0 ) 154 MOVSS ( XMM0, D(2) ) 202 XORPS( XMM0, XMM0 ) /* clean the working register */ 220 MOVSS ( S(2), XMM0 ) /* sz */ 288 MOVSS ( S(2), XMM0 ) /* oz */ 293 MOVSS ( S(2), XMM0 ) /* oz */ 415 MOVLPS( S(0), XMM0 ) /* oy | ox */ 420 MOVSS( S(2), XMM0 ) [all …]
|
/dports/graphics/mesa-dri/mesa-21.3.6/src/mesa/x86/ |
H A D | sse_xform3.S | 151 MOVLPS ( S(0), XMM0 ) 152 MOVLPS ( XMM0, D(0) ) 153 MOVSS ( S(2), XMM0 ) 154 MOVSS ( XMM0, D(2) ) 202 XORPS( XMM0, XMM0 ) /* clean the working register */ 220 MOVSS ( S(2), XMM0 ) /* sz */ 288 MOVSS ( S(2), XMM0 ) /* oz */ 293 MOVSS ( S(2), XMM0 ) /* oz */ 415 MOVLPS( S(0), XMM0 ) /* oy | ox */ 420 MOVSS( S(2), XMM0 ) [all …]
|
/dports/emulators/dolphin-emu/dolphin-3152428/Source/Core/Core/PowerPC/Jit64/ |
H A D | Jit_FloatingPoint.cpp | 334 MULPD(XMM0, Ra); in fmaddXX() 339 MULSD(XMM0, Ra); in fmaddXX() 431 XORPD(XMM0, R(XMM0)); in fselx() 474 ANDPD(XMM0, Rb); in fselx() 653 MINSD(XMM0, Rb); in fctiwx() 658 CVTPD2DQ(XMM0, R(XMM0)); in fctiwx() 663 CVTTPD2DQ(XMM0, R(XMM0)); in fctiwx() 667 MOVSD(Rd, XMM0); in fctiwx() 700 MOVAPD(XMM0, Rb); in frsqrtex() 702 MOVSD(Rd, XMM0); in frsqrtex() [all …]
|