diff options
| author | Sunil K Pandey <skpgkp2@gmail.com> | 2021-12-29 08:59:16 -0800 |
|---|---|---|
| committer | Sunil K Pandey <skpgkp2@gmail.com> | 2021-12-29 11:37:49 -0800 |
| commit | 76ddc74e86f7bc36468736dd22c4c29e39cd62d8 (patch) | |
| tree | 41692fa99b3484c5445ef0ea2dfcad0f8990d800 /sysdeps | |
| parent | ef7ea9c1327be11c54462cb539ea8854979a21f3 (diff) | |
| download | glibc-76ddc74e86f7bc36468736dd22c4c29e39cd62d8.zip glibc-76ddc74e86f7bc36468736dd22c4c29e39cd62d8.tar.gz glibc-76ddc74e86f7bc36468736dd22c4c29e39cd62d8.tar.bz2 | |
x86-64: Add vector expm1/expm1f implementation to libmvec
Implement vectorized expm1/expm1f containing SSE, AVX, AVX2 and
AVX512 versions for libmvec as per vector ABI. It also contains
accuracy and ABI tests for vector expm1/expm1f with regenerated ulps.
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
Diffstat (limited to 'sysdeps')
48 files changed, 2713 insertions, 0 deletions
diff --git a/sysdeps/unix/sysv/linux/x86_64/libmvec.abilist b/sysdeps/unix/sysv/linux/x86_64/libmvec.abilist index 4907680..c9d3213 100644 --- a/sysdeps/unix/sysv/linux/x86_64/libmvec.abilist +++ b/sysdeps/unix/sysv/linux/x86_64/libmvec.abilist @@ -52,6 +52,7 @@ GLIBC_2.35 _ZGVbN2v_atan F GLIBC_2.35 _ZGVbN2v_cosh F GLIBC_2.35 _ZGVbN2v_exp10 F GLIBC_2.35 _ZGVbN2v_exp2 F +GLIBC_2.35 _ZGVbN2v_expm1 F GLIBC_2.35 _ZGVbN2vv_hypot F GLIBC_2.35 _ZGVbN4v_acosf F GLIBC_2.35 _ZGVbN4v_asinf F @@ -59,6 +60,7 @@ GLIBC_2.35 _ZGVbN4v_atanf F GLIBC_2.35 _ZGVbN4v_coshf F GLIBC_2.35 _ZGVbN4v_exp10f F GLIBC_2.35 _ZGVbN4v_exp2f F +GLIBC_2.35 _ZGVbN4v_expm1f F GLIBC_2.35 _ZGVbN4vv_hypotf F GLIBC_2.35 _ZGVcN4v_acos F GLIBC_2.35 _ZGVcN4v_asin F @@ -66,6 +68,7 @@ GLIBC_2.35 _ZGVcN4v_atan F GLIBC_2.35 _ZGVcN4v_cosh F GLIBC_2.35 _ZGVcN4v_exp10 F GLIBC_2.35 _ZGVcN4v_exp2 F +GLIBC_2.35 _ZGVcN4v_expm1 F GLIBC_2.35 _ZGVcN4vv_hypot F GLIBC_2.35 _ZGVcN8v_acosf F GLIBC_2.35 _ZGVcN8v_asinf F @@ -73,6 +76,7 @@ GLIBC_2.35 _ZGVcN8v_atanf F GLIBC_2.35 _ZGVcN8v_coshf F GLIBC_2.35 _ZGVcN8v_exp10f F GLIBC_2.35 _ZGVcN8v_exp2f F +GLIBC_2.35 _ZGVcN8v_expm1f F GLIBC_2.35 _ZGVcN8vv_hypotf F GLIBC_2.35 _ZGVdN4v_acos F GLIBC_2.35 _ZGVdN4v_asin F @@ -80,6 +84,7 @@ GLIBC_2.35 _ZGVdN4v_atan F GLIBC_2.35 _ZGVdN4v_cosh F GLIBC_2.35 _ZGVdN4v_exp10 F GLIBC_2.35 _ZGVdN4v_exp2 F +GLIBC_2.35 _ZGVdN4v_expm1 F GLIBC_2.35 _ZGVdN4vv_hypot F GLIBC_2.35 _ZGVdN8v_acosf F GLIBC_2.35 _ZGVdN8v_asinf F @@ -87,6 +92,7 @@ GLIBC_2.35 _ZGVdN8v_atanf F GLIBC_2.35 _ZGVdN8v_coshf F GLIBC_2.35 _ZGVdN8v_exp10f F GLIBC_2.35 _ZGVdN8v_exp2f F +GLIBC_2.35 _ZGVdN8v_expm1f F GLIBC_2.35 _ZGVdN8vv_hypotf F GLIBC_2.35 _ZGVeN16v_acosf F GLIBC_2.35 _ZGVeN16v_asinf F @@ -94,6 +100,7 @@ GLIBC_2.35 _ZGVeN16v_atanf F GLIBC_2.35 _ZGVeN16v_coshf F GLIBC_2.35 _ZGVeN16v_exp10f F GLIBC_2.35 _ZGVeN16v_exp2f F +GLIBC_2.35 _ZGVeN16v_expm1f F GLIBC_2.35 _ZGVeN16vv_hypotf F GLIBC_2.35 _ZGVeN8v_acos F GLIBC_2.35 _ZGVeN8v_asin F @@ -101,4 +108,5 @@ GLIBC_2.35 _ZGVeN8v_atan F GLIBC_2.35 _ZGVeN8v_cosh F GLIBC_2.35 _ZGVeN8v_exp10 F GLIBC_2.35 _ZGVeN8v_exp2 F +GLIBC_2.35 _ZGVeN8v_expm1 F GLIBC_2.35 _ZGVeN8vv_hypot F diff --git a/sysdeps/x86/fpu/bits/math-vector.h b/sysdeps/x86/fpu/bits/math-vector.h index 708e81b..e2f98e1 100644 --- a/sysdeps/x86/fpu/bits/math-vector.h +++ b/sysdeps/x86/fpu/bits/math-vector.h @@ -86,6 +86,10 @@ # define __DECL_SIMD_cosh __DECL_SIMD_x86_64 # undef __DECL_SIMD_coshf # define __DECL_SIMD_coshf __DECL_SIMD_x86_64 +# undef __DECL_SIMD_expm1 +# define __DECL_SIMD_expm1 __DECL_SIMD_x86_64 +# undef __DECL_SIMD_expm1f +# define __DECL_SIMD_expm1f __DECL_SIMD_x86_64 # endif #endif diff --git a/sysdeps/x86/fpu/finclude/math-vector-fortran.h b/sysdeps/x86/fpu/finclude/math-vector-fortran.h index 81d0238..4323305 100644 --- a/sysdeps/x86/fpu/finclude/math-vector-fortran.h +++ b/sysdeps/x86/fpu/finclude/math-vector-fortran.h @@ -42,6 +42,8 @@ !GCC$ builtin (exp10f) attributes simd (notinbranch) if('x86_64') !GCC$ builtin (cosh) attributes simd (notinbranch) if('x86_64') !GCC$ builtin (coshf) attributes simd (notinbranch) if('x86_64') +!GCC$ builtin (expm1) attributes simd (notinbranch) if('x86_64') +!GCC$ builtin (expm1f) attributes simd (notinbranch) if('x86_64') !GCC$ builtin (cos) attributes simd (notinbranch) if('x32') !GCC$ builtin (cosf) attributes simd (notinbranch) if('x32') @@ -69,3 +71,5 @@ !GCC$ builtin (exp10f) attributes simd (notinbranch) if('x32') !GCC$ builtin (cosh) attributes simd (notinbranch) if('x32') !GCC$ builtin (coshf) attributes simd (notinbranch) if('x32') +!GCC$ builtin (expm1) attributes simd (notinbranch) if('x32') +!GCC$ builtin (expm1f) attributes simd (notinbranch) if('x32') diff --git a/sysdeps/x86_64/fpu/Makeconfig b/sysdeps/x86_64/fpu/Makeconfig index 5bc2df1..8de8214 100644 --- a/sysdeps/x86_64/fpu/Makeconfig +++ b/sysdeps/x86_64/fpu/Makeconfig @@ -30,6 +30,7 @@ libmvec-funcs = \ exp \ exp10 \ exp2 \ + expm1 \ hypot \ log \ pow \ diff --git a/sysdeps/x86_64/fpu/Versions b/sysdeps/x86_64/fpu/Versions index 53346d1..58debb2 100644 --- a/sysdeps/x86_64/fpu/Versions +++ b/sysdeps/x86_64/fpu/Versions @@ -20,6 +20,7 @@ libmvec { _ZGVbN2v_cosh; _ZGVcN4v_cosh; _ZGVdN4v_cosh; _ZGVeN8v_cosh; _ZGVbN2v_exp10; _ZGVcN4v_exp10; _ZGVdN4v_exp10; _ZGVeN8v_exp10; _ZGVbN2v_exp2; _ZGVcN4v_exp2; _ZGVdN4v_exp2; _ZGVeN8v_exp2; + _ZGVbN2v_expm1; _ZGVcN4v_expm1; _ZGVdN4v_expm1; _ZGVeN8v_expm1; _ZGVbN2vv_hypot; _ZGVcN4vv_hypot; _ZGVdN4vv_hypot; _ZGVeN8vv_hypot; _ZGVbN4v_acosf; _ZGVcN8v_acosf; _ZGVdN8v_acosf; _ZGVeN16v_acosf; _ZGVbN4v_asinf; _ZGVcN8v_asinf; _ZGVdN8v_asinf; _ZGVeN16v_asinf; @@ -27,6 +28,7 @@ libmvec { _ZGVbN4v_coshf; _ZGVcN8v_coshf; _ZGVdN8v_coshf; _ZGVeN16v_coshf; _ZGVbN4v_exp10f; _ZGVcN8v_exp10f; _ZGVdN8v_exp10f; _ZGVeN16v_exp10f; _ZGVbN4v_exp2f; _ZGVcN8v_exp2f; _ZGVdN8v_exp2f; _ZGVeN16v_exp2f; + _ZGVbN4v_expm1f; _ZGVcN8v_expm1f; _ZGVdN8v_expm1f; _ZGVeN16v_expm1f; _ZGVbN4vv_hypotf; _ZGVcN8vv_hypotf; _ZGVdN8vv_hypotf; _ZGVeN16vv_hypotf; } } diff --git a/sysdeps/x86_64/fpu/libm-test-ulps b/sysdeps/x86_64/fpu/libm-test-ulps index ac70f15..f05ece8 100644 --- a/sysdeps/x86_64/fpu/libm-test-ulps +++ b/sysdeps/x86_64/fpu/libm-test-ulps @@ -1395,6 +1395,26 @@ float: 1 float128: 3 ldouble: 4 +Function: "expm1_vlen16": +float: 1 + +Function: "expm1_vlen2": +double: 1 + +Function: "expm1_vlen4": +double: 1 +float: 1 + +Function: "expm1_vlen4_avx2": +double: 1 + +Function: "expm1_vlen8": +double: 1 +float: 1 + +Function: "expm1_vlen8_avx2": +float: 1 + Function: "gamma": double: 4 float: 7 diff --git a/sysdeps/x86_64/fpu/multiarch/svml_d_expm12_core-sse2.S b/sysdeps/x86_64/fpu/multiarch/svml_d_expm12_core-sse2.S new file mode 100644 index 0000000..e8cb6fa --- /dev/null +++ b/sysdeps/x86_64/fpu/multiarch/svml_d_expm12_core-sse2.S @@ -0,0 +1,20 @@ +/* SSE2 version of vectorized expm1, vector length is 2. + Copyright (C) 2021 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + <https://www.gnu.org/licenses/>. */ + +#define _ZGVbN2v_expm1 _ZGVbN2v_expm1_sse2 +#include "../svml_d_expm12_core.S" diff --git a/sysdeps/x86_64/fpu/multiarch/svml_d_expm12_core.c b/sysdeps/x86_64/fpu/multiarch/svml_d_expm12_core.c new file mode 100644 index 0000000..9c794e9 --- /dev/null +++ b/sysdeps/x86_64/fpu/multiarch/svml_d_expm12_core.c @@ -0,0 +1,27 @@ +/* Multiple versions of vectorized expm1, vector length is 2. + Copyright (C) 2021 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + <https://www.gnu.org/licenses/>. */ + +#define SYMBOL_NAME _ZGVbN2v_expm1 +#include "ifunc-mathvec-sse4_1.h" + +libc_ifunc_redirected (REDIRECT_NAME, SYMBOL_NAME, IFUNC_SELECTOR ()); + +#ifdef SHARED +__hidden_ver1 (_ZGVbN2v_expm1, __GI__ZGVbN2v_expm1, __redirect__ZGVbN2v_expm1) + __attribute__ ((visibility ("hidden"))); +#endif diff --git a/sysdeps/x86_64/fpu/multiarch/svml_d_expm12_core_sse4.S b/sysdeps/x86_64/fpu/multiarch/svml_d_expm12_core_sse4.S new file mode 100644 index 0000000..db763e3 --- /dev/null +++ b/sysdeps/x86_64/fpu/multiarch/svml_d_expm12_core_sse4.S @@ -0,0 +1,421 @@ +/* Function expm1 vectorized with SSE4. + Copyright (C) 2021 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + https://www.gnu.org/licenses/. */ + +/* + * ALGORITHM DESCRIPTION: + * + * N = (int)(x*2^k/log(2.0)), R = x - N*log(2)/2^k + * exp(x) = 2^(N/2^k) * poly(R) is computed in high-low parts + * expm1(x) = exp(x)-1 is then obtained via multi-precision computation + * + * + */ + +/* Offsets for data table __svml_dexpm1_data_internal + */ +#define Expm1_HA_table 0 +#define poly_coeff 2048 +#define Log2e 2112 +#define L2H 2128 +#define L2L 2144 +#define ExpAddConst 2160 +#define IndexMask 2176 +#define ExpMask 2192 +#define MOne 2208 +#define AbsMask 2224 +#define Threshold 2240 +#define L2 2256 + +#include <sysdep.h> + + .text + .section .text.sse4,"ax",@progbits +ENTRY(_ZGVbN2v_expm1_sse4) + pushq %rbp + cfi_def_cfa_offset(16) + movq %rsp, %rbp + cfi_def_cfa(6, 16) + cfi_offset(6, -16) + andq $-32, %rsp + subq $64, %rsp + movaps %xmm0, %xmm2 + movups Log2e+__svml_dexpm1_data_internal(%rip), %xmm7 + lea __svml_dexpm1_data_internal(%rip), %rsi + mulpd %xmm0, %xmm7 + movups .FLT_10(%rip), %xmm3 + addpd %xmm3, %xmm7 + subpd %xmm3, %xmm7 + +/* argument reduction */ + movups L2H+__svml_dexpm1_data_internal(%rip), %xmm4 + mulpd %xmm7, %xmm4 + movups L2L+__svml_dexpm1_data_internal(%rip), %xmm5 + mulpd %xmm7, %xmm5 + subpd %xmm4, %xmm2 + subpd %xmm5, %xmm2 + +/* polynomial */ + movups poly_coeff+__svml_dexpm1_data_internal(%rip), %xmm12 + movaps %xmm2, %xmm14 + mulpd %xmm2, %xmm12 + mulpd %xmm2, %xmm14 + addpd poly_coeff+16+__svml_dexpm1_data_internal(%rip), %xmm12 + movups ExpAddConst+__svml_dexpm1_data_internal(%rip), %xmm15 + addpd %xmm7, %xmm15 + mulpd %xmm14, %xmm12 + movups poly_coeff+32+__svml_dexpm1_data_internal(%rip), %xmm13 + mulpd %xmm2, %xmm13 + +/* table lookup */ + movdqu IndexMask+__svml_dexpm1_data_internal(%rip), %xmm8 + pand %xmm15, %xmm8 + movups AbsMask+__svml_dexpm1_data_internal(%rip), %xmm1 + pshufd $2, %xmm8, %xmm9 + movaps %xmm1, %xmm6 + movd %xmm8, %eax + andps %xmm0, %xmm6 + movd %xmm9, %ecx + andnps %xmm0, %xmm1 + movdqu ExpMask+__svml_dexpm1_data_internal(%rip), %xmm11 + pand %xmm11, %xmm15 + cmpnlepd Threshold+__svml_dexpm1_data_internal(%rip), %xmm6 + addpd poly_coeff+48+__svml_dexpm1_data_internal(%rip), %xmm13 + movmskpd %xmm6, %edx + psllq $41, %xmm15 + +/* T-1 */ + movups MOne+__svml_dexpm1_data_internal(%rip), %xmm4 + movslq %eax, %rax + movslq %ecx, %rcx + addpd %xmm12, %xmm13 + movups (%rsi,%rax), %xmm3 + movups (%rsi,%rcx), %xmm10 + movaps %xmm3, %xmm6 + unpckhpd %xmm10, %xmm3 + +/* Th1 = (Th-1) + Tl */ + mulpd %xmm15, %xmm3 + mulpd %xmm13, %xmm14 + unpcklpd %xmm10, %xmm6 + orps %xmm15, %xmm6 + addpd %xmm4, %xmm6 + addpd %xmm14, %xmm2 + addpd %xmm3, %xmm6 + +/* T = Th+Tl */ + movaps %xmm6, %xmm5 + subpd %xmm4, %xmm5 + mulpd %xmm5, %xmm2 + addpd %xmm2, %xmm6 + orps %xmm1, %xmm6 + testl %edx, %edx + +/* Go to special inputs processing branch */ + jne L(SPECIAL_VALUES_BRANCH) + # LOE rbx r12 r13 r14 r15 edx xmm0 xmm6 + +/* Restore registers + * and exit the function + */ + +L(EXIT): + movaps %xmm6, %xmm0 + movq %rbp, %rsp + popq %rbp + cfi_def_cfa(7, 8) + cfi_restore(6) + ret + cfi_def_cfa(6, 16) + cfi_offset(6, -16) + +/* Branch to process + * special inputs + */ + +L(SPECIAL_VALUES_BRANCH): + movups %xmm0, 32(%rsp) + movups %xmm6, 48(%rsp) + # LOE rbx r12 r13 r14 r15 edx + + xorl %eax, %eax + movq %r12, 16(%rsp) + /* DW_CFA_expression: r12 (r12) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -48; DW_OP_plus) */ + .cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xd0, 0xff, 0xff, 0xff, 0x22 + movl %eax, %r12d + movq %r13, 8(%rsp) + /* DW_CFA_expression: r13 (r13) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -56; DW_OP_plus) */ + .cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xc8, 0xff, 0xff, 0xff, 0x22 + movl %edx, %r13d + movq %r14, (%rsp) + /* DW_CFA_expression: r14 (r14) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -64; DW_OP_plus) */ + .cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x22 + # LOE rbx r15 r12d r13d + +/* Range mask + * bits check + */ + +L(RANGEMASK_CHECK): + btl %r12d, %r13d + +/* Call scalar math function */ + jc L(SCALAR_MATH_CALL) + # LOE rbx r15 r12d r13d + +/* Special inputs + * processing loop + */ + +L(SPECIAL_VALUES_LOOP): + incl %r12d + cmpl $2, %r12d + +/* Check bits in range mask */ + jl L(RANGEMASK_CHECK) + # LOE rbx r15 r12d r13d + + movq 16(%rsp), %r12 + cfi_restore(12) + movq 8(%rsp), %r13 + cfi_restore(13) + movq (%rsp), %r14 + cfi_restore(14) + movups 48(%rsp), %xmm6 + +/* Go to exit */ + jmp L(EXIT) + /* DW_CFA_expression: r12 (r12) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -48; DW_OP_plus) */ + .cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xd0, 0xff, 0xff, 0xff, 0x22 + /* DW_CFA_expression: r13 (r13) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -56; DW_OP_plus) */ + .cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xc8, 0xff, 0xff, 0xff, 0x22 + /* DW_CFA_expression: r14 (r14) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -64; DW_OP_plus) */ + .cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x22 + # LOE rbx r12 r13 r14 r15 xmm6 + +/* Scalar math fucntion call + * to process special input + */ + +L(SCALAR_MATH_CALL): + movl %r12d, %r14d + movsd 32(%rsp,%r14,8), %xmm0 + call expm1@PLT + # LOE rbx r14 r15 r12d r13d xmm0 + + movsd %xmm0, 48(%rsp,%r14,8) + +/* Process special inputs in loop */ + jmp L(SPECIAL_VALUES_LOOP) + # LOE rbx r15 r12d r13d +END(_ZGVbN2v_expm1_sse4) + + .section .rodata, "a" + .align 16 + +#ifdef __svml_dexpm1_data_internal_typedef +typedef unsigned int VUINT32; +typedef struct { + __declspec(align(16)) VUINT32 Expm1_HA_table[(1<<8)][2]; + __declspec(align(16)) VUINT32 poly_coeff[4][2][2]; + __declspec(align(16)) VUINT32 Log2e[2][2]; + __declspec(align(16)) VUINT32 L2H[2][2]; + __declspec(align(16)) VUINT32 L2L[2][2]; + __declspec(align(16)) VUINT32 ExpAddConst[2][2]; + __declspec(align(16)) VUINT32 IndexMask[2][2]; + __declspec(align(16)) VUINT32 ExpMask[2][2]; + __declspec(align(16)) VUINT32 MOne[2][2]; + __declspec(align(16)) VUINT32 AbsMask[2][2]; + __declspec(align(16)) VUINT32 Threshold[2][2]; + __declspec(align(16)) VUINT32 L2[2][2]; +} __svml_dexpm1_data_internal; +#endif +__svml_dexpm1_data_internal: + /* Expm1_HA_table */ + .quad 0x0000000000000000, 0x0000000000000000 + .quad 0x0000163da8000000, 0x3e3fb33356d84a67 + .quad 0x00002c9a40000000, 0xbe3887f9f1190835 + .quad 0x00004315e8000000, 0x3e1b9fe12f5ce3e7 + .quad 0x000059b0d0000000, 0x3e48ac2ba1d73e2a + .quad 0x0000706b28000000, 0x3e3ddf6ddc6dc404 + .quad 0x0000874518000000, 0x3e1d66f20230d7c9 + .quad 0x00009e3ec8000000, 0x3e46379c1a290f03 + .quad 0x0000b55870000000, 0xbe4833b784eb3a37 + .quad 0x0000cc9228000000, 0x3e4b923fba03db83 + .quad 0x0000e3ec30000000, 0x3e469e8d10103a17 + .quad 0x0000fb66b0000000, 0xbdb2ce50dcdf6e22 + .quad 0x00011301d0000000, 0x3df25b50a4ebbf1b + .quad 0x00012abdc0000000, 0x3e1b0c72fee4aeb5 + .quad 0x0001429ab0000000, 0xbe356d2204cbefe7 + .quad 0x00015a98c8000000, 0x3e24b1ca24901aae + .quad 0x000172b840000000, 0xbe4c15742919041c + .quad 0x00018af938000000, 0x3e2191bd3777ee17 + .quad 0x0001a35be8000000, 0x3e4b7e5ba9e5b4c8 + .quad 0x0001bbe088000000, 0xbe4fdd19632a70c7 + .quad 0x0001d48730000000, 0x3e368b9aa7805b80 + .quad 0x0001ed5020000000, 0x3e47e6c8e5c40d00 + .quad 0x0002063b88000000, 0x3e18a3358ee3bac1 + .quad 0x00021f4990000000, 0x3e37ddc962552fd3 + .quad 0x0002387a70000000, 0xbe38a9dc7993e052 + .quad 0x000251ce50000000, 0xbe135670329f5521 + .quad 0x00026b4568000000, 0xbe40ec1916d42cc6 + .quad 0x000284dfe0000000, 0x3e3f5638096cf15d + .quad 0x00029e9df8000000, 0xbe470108f69ed175 + .quad 0x0002b87fd0000000, 0x3e2b5b31ffbbd48d + .quad 0x0002d285a8000000, 0xbe31bfcf4bff6e2b + .quad 0x0002ecafa8000000, 0x3e33e2f5611ca0f4 + .quad 0x000306fe08000000, 0x3e418db8a96f46ad + .quad 0x0003217100000000, 0xbe4d993e76563187 + .quad 0x00033c08b0000000, 0x3e4320b7fa64e431 + .quad 0x000356c560000000, 0xbe1b5803cdae772e + .quad 0x000371a738000000, 0xbe28aac6ab1d7560 + .quad 0x00038cae70000000, 0xbe47d13cd3d2b1a8 + .quad 0x0003a7db38000000, 0xbe48d30048af21b7 + .quad 0x0003c32dc0000000, 0x3e489d47242000f9 + .quad 0x0003dea650000000, 0xbe4f6e5eee525f6f + .quad 0x0003fa4508000000, 0xbe4a9bff22fa047f + .quad 0x0004160a20000000, 0x3e3f72e29f84325c + .quad 0x000431f5d8000000, 0x3e350a896dc70444 + .quad 0x00044e0860000000, 0x3e18624b40c4dbd0 + .quad 0x00046a41f0000000, 0xbe4717fd446d7686 + .quad 0x000486a2b8000000, 0xbe41f6197f61f2e2 + .quad 0x0004a32af0000000, 0x3e2afa7bcce5b17a + .quad 0x0004bfdad8000000, 0xbe464eaec715e343 + .quad 0x0004dcb298000000, 0x3e3fddd0d63b36ef + .quad 0x0004f9b278000000, 0xbe362d35952cc275 + .quad 0x000516daa0000000, 0x3e467b320e0897a9 + .quad 0x0005342b58000000, 0xbe362b07e20f57c4 + .quad 0x000551a4c8000000, 0x3e42ec9076297631 + .quad 0x00056f4738000000, 0xbe34ad8259913500 + .quad 0x00058d12d8000000, 0xbe4b41c016d6a1ea + .quad 0x0005ab07e0000000, 0xbe45bd5eb539b67f + .quad 0x0005c92688000000, 0x3e42ca35b80e258e + .quad 0x0005e76f18000000, 0xbe4296f5bc8b20da + .quad 0x000605e1b8000000, 0x3e376dc08b076f59 + .quad 0x0006247eb0000000, 0x3e0d2ac258f87d03 + .quad 0x0006434638000000, 0xbe4999e701c483c7 + .quad 0x0006623880000000, 0x3e42a91124893ecf + .quad 0x00068155d8000000, 0xbe4d9ab467bf1d47 + .quad 0x0006a09e68000000, 0xbe380c4336f74d05 + .quad 0x0006c01278000000, 0xbe47a12a08944ab3 + .quad 0x0006dfb240000000, 0xbe4cd72e886ef8ea + .quad 0x0006ff7df8000000, 0x3e3519483cf87e1b + .quad 0x00071f75e8000000, 0x3e2d8bee7ba46e1e + .quad 0x00073f9a48000000, 0x3e24b02e77ab934a + .quad 0x00075feb58000000, 0xbe3bd98374091656 + .quad 0x0007806950000000, 0xbe00d1604f328fec + .quad 0x0007a11470000000, 0x3e4f580c36bea881 + .quad 0x0007c1ed00000000, 0x3e330c1327c49334 + .quad 0x0007e2f338000000, 0xbe330b19defa2fd4 + .quad 0x0008042758000000, 0xbe4e0f2f724f90cc + .quad 0x0008258998000000, 0x3e34cce128acf88b + .quad 0x0008471a48000000, 0xbe3dc385331ad094 + .quad 0x000868d998000000, 0x3e4a2497640720ed + .quad 0x00088ac7d8000000, 0x3e38a669966530bd + .quad 0x0008ace540000000, 0x3e415506dadd3e2b + .quad 0x0008cf3218000000, 0xbe34abb7410d55e3 + .quad 0x0008f1ae98000000, 0x3e31577362b98274 + .quad 0x0009145b08000000, 0x3e4c8ffe2c4530da + .quad 0x00093737b0000000, 0x3e29b8bc9e8a0388 + .quad 0x00095a44c8000000, 0x3e4e4290774da41b + .quad 0x00097d82a0000000, 0xbe00d8d83a30b6f8 + .quad 0x0009a0f170000000, 0x3e2940f737462137 + .quad 0x0009c49180000000, 0x3e451f8480e3e236 + .quad 0x0009e86318000000, 0x3e3e323231824ca8 + .quad 0x000a0c6678000000, 0x3e4aef2b2594d6d4 + .quad 0x000a309bf0000000, 0xbe4dae966539f470 + .quad 0x000a5503b0000000, 0x3e41f12ae45a1225 + .quad 0x000a799e10000000, 0x3e49859ac3796fd9 + .quad 0x000a9e6b58000000, 0xbe44301205e0a6de + .quad 0x000ac36bc0000000, 0xbe0606431f9234cb + .quad 0x000ae89f98000000, 0x3e35ad3ad5e8734d + .quad 0x000b0e0728000000, 0x3e38db66590842ad + .quad 0x000b33a2b8000000, 0x3e13c57ebdaff43a + .quad 0x000b597290000000, 0xbe40d536338e3bf7 + .quad 0x000b7f76f0000000, 0x3e47daf237553d84 + .quad 0x000ba5b030000000, 0x3e2420c930819679 + .quad 0x000bcc1e90000000, 0x3e12f074891ee83d + .quad 0x000bf2c258000000, 0x3e4eb8f0442046b8 + .quad 0x000c199be0000000, 0xbe43d56b1eeef9a7 + .quad 0x000c40ab60000000, 0xbd87c2c975903ef8 + .quad 0x000c67f130000000, 0xbe3a82eb4b5dec80 + .quad 0x000c8f6d98000000, 0xbe4fc8c257729a1e + .quad 0x000cb720e0000000, 0xbe48837cb757e1a1 + .quad 0x000cdf0b58000000, 0xbe4511e031dd83b5 + .quad 0x000d072d48000000, 0x3e403c4bdc687918 + .quad 0x000d2f8708000000, 0x3deb13e315bc2473 + .quad 0x000d5818e0000000, 0xbe4822dbc6d12fd3 + .quad 0x000d80e318000000, 0xbe3367c68447b063 + .quad 0x000da9e600000000, 0x3e4ed9942b84600d + .quad 0x000dd321f0000000, 0x3e480da3025b4aef + .quad 0x000dfc9730000000, 0x3e4bdcdaf5cb4656 + .quad 0x000e264618000000, 0xbe4852f6baf6c4f0 + .quad 0x000e502ee8000000, 0xbe1d30027630bb40 + .quad 0x000e7a51f8000000, 0x3e4e3a641a5aa459 + .quad 0x000ea4afa0000000, 0x3e452486cc2c7b9d + .quad 0x000ecf4830000000, 0xbe438cc07b927e77 + .quad 0x000efa1bf0000000, 0xbe39ea5d888e02de + .quad 0x000f252b38000000, 0xbe2288ad162f2d20 + .quad 0x000f507658000000, 0x3e4b722a033a7c26 + .quad 0x000f7bfdb0000000, 0xbe431a0f63b7625a + .quad 0x000fa7c180000000, 0x3e39e90d82e90a7e + .quad 0x000fd3c228000000, 0x3e4c7b8f884badd2 + /*== poly_coeff[4] ==*/ + .align 16 + .quad 0x3f81111168877F38, 0x3f81111168877F38 /* coeff5 */ + .quad 0x3fa55555C2A9C0F3, 0x3fa55555C2A9C0F3 /* coeff4 */ + .quad 0x3fc555555555541D, 0x3fc555555555541D /* coeff3 */ + .quad 0x3fdFFFFFFFFFFE5C, 0x3fdFFFFFFFFFFE5C /* coeff2 */ + /*== Log2e ==*/ + .align 16 + .quad 0x40671547652B82FE, 0x40671547652B82FE + /*== L2H ==*/ + .align 16 + .quad 0x3f762e42fef80000, 0x3f762e42fef80000 + /*== L2L ==*/ + .align 16 + .quad 0x3d41cf79abc9e3b4, 0x3d41cf79abc9e3b4 + /*== ExpAddConst ==*/ + .align 16 + .quad 0x42f80000001ff800, 0x42f80000001ff800 + /*== IndexMask ==*/ + .align 16 + .quad 0x00000000000007f0, 0x00000000000007f0 + /*== ExpMask ==*/ + .align 16 + .quad 0x00000000003ff800, 0x00000000003ff800 + /*== MOne ==*/ + .align 16 + .quad 0xbff0000000000000, 0xbff0000000000000 + /*== AbsMask ==*/ + .align 16 + .quad 0x7fffffffffffffff, 0x7fffffffffffffff + /*== Threshold ==*/ + .align 16 + .quad 0x40861DA04CBAFE43, 0x40861DA04CBAFE43 + /*== L2 ==*/ + .align 16 + .quad 0x3f762e42fefa39ef, 0x3f762e42fefa39ef + .align 16 + .type __svml_dexpm1_data_internal,@object + .size __svml_dexpm1_data_internal,.-__svml_dexpm1_data_internal + .align 16 + +.FLT_10: + .long 0x00000000,0x43380000,0x00000000,0x43380000 + .type .FLT_10,@object + .size .FLT_10,16 diff --git a/sysdeps/x86_64/fpu/multiarch/svml_d_expm14_core-sse.S b/sysdeps/x86_64/fpu/multiarch/svml_d_expm14_core-sse.S new file mode 100644 index 0000000..e701670 --- /dev/null +++ b/sysdeps/x86_64/fpu/multiarch/svml_d_expm14_core-sse.S @@ -0,0 +1,20 @@ +/* SSE version of vectorized expm1, vector length is 4. + Copyright (C) 2021 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + <https://www.gnu.org/licenses/>. */ + +#define _ZGVdN4v_expm1 _ZGVdN4v_expm1_sse_wrapper +#include "../svml_d_expm14_core.S" diff --git a/sysdeps/x86_64/fpu/multiarch/svml_d_expm14_core.c b/sysdeps/x86_64/fpu/multiarch/svml_d_expm14_core.c new file mode 100644 index 0000000..4215d7d --- /dev/null +++ b/sysdeps/x86_64/fpu/multiarch/svml_d_expm14_core.c @@ -0,0 +1,27 @@ +/* Multiple versions of vectorized expm1, vector length is 4. + Copyright (C) 2021 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + <https://www.gnu.org/licenses/>. */ + +#define SYMBOL_NAME _ZGVdN4v_expm1 +#include "ifunc-mathvec-avx2.h" + +libc_ifunc_redirected (REDIRECT_NAME, SYMBOL_NAME, IFUNC_SELECTOR ()); + +#ifdef SHARED +__hidden_ver1 (_ZGVdN4v_expm1, __GI__ZGVdN4v_expm1, __redirect__ZGVdN4v_expm1) + __attribute__ ((visibility ("hidden"))); +#endif diff --git a/sysdeps/x86_64/fpu/multiarch/svml_d_expm14_core_avx2.S b/sysdeps/x86_64/fpu/multiarch/svml_d_expm14_core_avx2.S new file mode 100644 index 0000000..c34f73a --- /dev/null +++ b/sysdeps/x86_64/fpu/multiarch/svml_d_expm14_core_avx2.S @@ -0,0 +1,408 @@ +/* Function expm1 vectorized with AVX2. + Copyright (C) 2021 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + https://www.gnu.org/licenses/. */ + +/* + * ALGORITHM DESCRIPTION: + * + * N = (int)(x*2^k/log(2.0)), R = x - N*log(2)/2^k + * exp(x) = 2^(N/2^k) * poly(R) is computed in high-low parts + * expm1(x) = exp(x)-1 is then obtained via multi-precision computation + * + * + */ + +/* Offsets for data table __svml_dexpm1_data_internal + */ +#define Expm1_HA_table 0 +#define poly_coeff 2048 +#define Log2e 2176 +#define L2H 2208 +#define L2L 2240 +#define ExpAddConst 2272 +#define IndexMask 2304 +#define ExpMask 2336 +#define MOne 2368 +#define AbsMask 2400 +#define Threshold 2432 +#define L2 2464 + +#include <sysdep.h> + + .text + .section .text.avx2,"ax",@progbits +ENTRY(_ZGVdN4v_expm1_avx2) + pushq %rbp + cfi_def_cfa_offset(16) + movq %rsp, %rbp + cfi_def_cfa(6, 16) + cfi_offset(6, -16) + andq $-32, %rsp + subq $96, %rsp + lea __svml_dexpm1_data_internal(%rip), %r8 + vmovapd %ymm0, %ymm3 + vmulpd Log2e+__svml_dexpm1_data_internal(%rip), %ymm3, %ymm4 + +/* argument reduction */ + vmovupd L2H+__svml_dexpm1_data_internal(%rip), %ymm2 + vmovupd AbsMask+__svml_dexpm1_data_internal(%rip), %ymm5 + vroundpd $0, %ymm4, %ymm8 + vaddpd ExpAddConst+__svml_dexpm1_data_internal(%rip), %ymm8, %ymm0 + vfnmadd213pd %ymm3, %ymm8, %ymm2 + +/* table lookup */ + vandps IndexMask+__svml_dexpm1_data_internal(%rip), %ymm0, %ymm9 + vandpd %ymm5, %ymm3, %ymm6 + vcmpnle_uqpd Threshold+__svml_dexpm1_data_internal(%rip), %ymm6, %ymm7 + vfnmadd231pd L2L+__svml_dexpm1_data_internal(%rip), %ymm8, %ymm2 + vandnpd %ymm3, %ymm5, %ymm1 + vmovmskpd %ymm7, %eax + vmovupd poly_coeff+64+__svml_dexpm1_data_internal(%rip), %ymm7 + vmulpd %ymm2, %ymm2, %ymm8 + vfmadd213pd poly_coeff+96+__svml_dexpm1_data_internal(%rip), %ymm2, %ymm7 + vandps ExpMask+__svml_dexpm1_data_internal(%rip), %ymm0, %ymm0 + vextractf128 $1, %ymm9, %xmm10 + vmovd %xmm9, %edx + vmovd %xmm10, %esi + vpextrd $2, %xmm9, %ecx + vpextrd $2, %xmm10, %edi + movslq %edx, %rdx + movslq %ecx, %rcx + movslq %esi, %rsi + movslq %edi, %rdi + vmovupd (%r8,%rdx), %xmm13 + vmovupd (%r8,%rcx), %xmm14 + vmovupd (%r8,%rsi), %xmm4 + vmovupd (%r8,%rdi), %xmm5 + vunpcklpd %xmm14, %xmm13, %xmm11 + vunpcklpd %xmm5, %xmm4, %xmm12 + vpsllq $41, %ymm0, %ymm10 + vunpckhpd %xmm14, %xmm13, %xmm15 + vunpckhpd %xmm5, %xmm4, %xmm13 + vinsertf128 $1, %xmm12, %ymm11, %ymm6 + +/* polynomial */ + vmovupd poly_coeff+__svml_dexpm1_data_internal(%rip), %ymm12 + +/* T-1 */ + vmovupd MOne+__svml_dexpm1_data_internal(%rip), %ymm11 + vfmadd213pd poly_coeff+32+__svml_dexpm1_data_internal(%rip), %ymm2, %ymm12 + vfmadd213pd %ymm7, %ymm8, %ymm12 + vorpd %ymm10, %ymm6, %ymm9 + vfmadd213pd %ymm2, %ymm8, %ymm12 + vaddpd %ymm11, %ymm9, %ymm2 + vinsertf128 $1, %xmm13, %ymm15, %ymm14 + +/* Th1 = (Th-1) + Tl */ + vfmadd213pd %ymm2, %ymm10, %ymm14 + +/* T = Th+Tl */ + vsubpd %ymm11, %ymm14, %ymm0 + vfmadd213pd %ymm14, %ymm12, %ymm0 + vorpd %ymm1, %ymm0, %ymm0 + testl %eax, %eax + +/* Go to special inputs processing branch */ + jne L(SPECIAL_VALUES_BRANCH) + # LOE rbx r12 r13 r14 r15 eax ymm0 ymm3 + +/* Restore registers + * and exit the function + */ + +L(EXIT): + movq %rbp, %rsp + popq %rbp + cfi_def_cfa(7, 8) + cfi_restore(6) + ret + cfi_def_cfa(6, 16) + cfi_offset(6, -16) + +/* Branch to process + * special inputs + */ + +L(SPECIAL_VALUES_BRANCH): + vmovupd %ymm3, 32(%rsp) + vmovupd %ymm0, 64(%rsp) + # LOE rbx r12 r13 r14 r15 eax ymm0 + + xorl %edx, %edx + # LOE rbx r12 r13 r14 r15 eax edx + + vzeroupper + movq %r12, 16(%rsp) + /* DW_CFA_expression: r12 (r12) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -80; DW_OP_plus) */ + .cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xb0, 0xff, 0xff, 0xff, 0x22 + movl %edx, %r12d + movq %r13, 8(%rsp) + /* DW_CFA_expression: r13 (r13) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -88; DW_OP_plus) */ + .cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xa8, 0xff, 0xff, 0xff, 0x22 + movl %eax, %r13d + movq %r14, (%rsp) + /* DW_CFA_expression: r14 (r14) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -96; DW_OP_plus) */ + .cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xa0, 0xff, 0xff, 0xff, 0x22 + # LOE rbx r15 r12d r13d + +/* Range mask + * bits check + */ + +L(RANGEMASK_CHECK): + btl %r12d, %r13d + +/* Call scalar math function */ + jc L(SCALAR_MATH_CALL) + # LOE rbx r15 r12d r13d + +/* Special inputs + * processing loop + */ + +L(SPECIAL_VALUES_LOOP): + incl %r12d + cmpl $4, %r12d + +/* Check bits in range mask */ + jl L(RANGEMASK_CHECK) + # LOE rbx r15 r12d r13d + + movq 16(%rsp), %r12 + cfi_restore(12) + movq 8(%rsp), %r13 + cfi_restore(13) + movq (%rsp), %r14 + cfi_restore(14) + vmovupd 64(%rsp), %ymm0 + +/* Go to exit */ + jmp L(EXIT) + /* DW_CFA_expression: r12 (r12) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -80; DW_OP_plus) */ + .cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xb0, 0xff, 0xff, 0xff, 0x22 + /* DW_CFA_expression: r13 (r13) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -88; DW_OP_plus) */ + .cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xa8, 0xff, 0xff, 0xff, 0x22 + /* DW_CFA_expression: r14 (r14) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -96; DW_OP_plus) */ + .cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xa0, 0xff, 0xff, 0xff, 0x22 + # LOE rbx r12 r13 r14 r15 ymm0 + +/* Scalar math fucntion call + * to process special input + */ + +L(SCALAR_MATH_CALL): + movl %r12d, %r14d + movsd 32(%rsp,%r14,8), %xmm0 + call expm1@PLT + # LOE rbx r14 r15 r12d r13d xmm0 + + movsd %xmm0, 64(%rsp,%r14,8) + +/* Process special inputs in loop */ + jmp L(SPECIAL_VALUES_LOOP) + # LOE rbx r15 r12d r13d +END(_ZGVdN4v_expm1_avx2) + + .section .rodata, "a" + .align 32 + +#ifdef __svml_dexpm1_data_internal_typedef +typedef unsigned int VUINT32; +typedef struct { + __declspec(align(32)) VUINT32 Expm1_HA_table[(1<<8)][2]; + __declspec(align(32)) VUINT32 poly_coeff[4][4][2]; + __declspec(align(32)) VUINT32 Log2e[4][2]; + __declspec(align(32)) VUINT32 L2H[4][2]; + __declspec(align(32)) VUINT32 L2L[4][2]; + __declspec(align(32)) VUINT32 ExpAddConst[4][2]; + __declspec(align(32)) VUINT32 IndexMask[4][2]; + __declspec(align(32)) VUINT32 ExpMask[4][2]; + __declspec(align(32)) VUINT32 MOne[4][2]; + __declspec(align(32)) VUINT32 AbsMask[4][2]; + __declspec(align(32)) VUINT32 Threshold[4][2]; + __declspec(align(32)) VUINT32 L2[4][2]; +} __svml_dexpm1_data_internal; +#endif +__svml_dexpm1_data_internal: + /* Expm1_HA_table */ + .quad 0x0000000000000000, 0x0000000000000000 + .quad 0x0000163da8000000, 0x3e3fb33356d84a67 + .quad 0x00002c9a40000000, 0xbe3887f9f1190835 + .quad 0x00004315e8000000, 0x3e1b9fe12f5ce3e7 + .quad 0x000059b0d0000000, 0x3e48ac2ba1d73e2a + .quad 0x0000706b28000000, 0x3e3ddf6ddc6dc404 + .quad 0x0000874518000000, 0x3e1d66f20230d7c9 + .quad 0x00009e3ec8000000, 0x3e46379c1a290f03 + .quad 0x0000b55870000000, 0xbe4833b784eb3a37 + .quad 0x0000cc9228000000, 0x3e4b923fba03db83 + .quad 0x0000e3ec30000000, 0x3e469e8d10103a17 + .quad 0x0000fb66b0000000, 0xbdb2ce50dcdf6e22 + .quad 0x00011301d0000000, 0x3df25b50a4ebbf1b + .quad 0x00012abdc0000000, 0x3e1b0c72fee4aeb5 + .quad 0x0001429ab0000000, 0xbe356d2204cbefe7 + .quad 0x00015a98c8000000, 0x3e24b1ca24901aae + .quad 0x000172b840000000, 0xbe4c15742919041c + .quad 0x00018af938000000, 0x3e2191bd3777ee17 + .quad 0x0001a35be8000000, 0x3e4b7e5ba9e5b4c8 + .quad 0x0001bbe088000000, 0xbe4fdd19632a70c7 + .quad 0x0001d48730000000, 0x3e368b9aa7805b80 + .quad 0x0001ed5020000000, 0x3e47e6c8e5c40d00 + .quad 0x0002063b88000000, 0x3e18a3358ee3bac1 + .quad 0x00021f4990000000, 0x3e37ddc962552fd3 + .quad 0x0002387a70000000, 0xbe38a9dc7993e052 + .quad 0x000251ce50000000, 0xbe135670329f5521 + .quad 0x00026b4568000000, 0xbe40ec1916d42cc6 + .quad 0x000284dfe0000000, 0x3e3f5638096cf15d + .quad 0x00029e9df8000000, 0xbe470108f69ed175 + .quad 0x0002b87fd0000000, 0x3e2b5b31ffbbd48d + .quad 0x0002d285a8000000, 0xbe31bfcf4bff6e2b + .quad 0x0002ecafa8000000, 0x3e33e2f5611ca0f4 + .quad 0x000306fe08000000, 0x3e418db8a96f46ad + .quad 0x0003217100000000, 0xbe4d993e76563187 + .quad 0x00033c08b0000000, 0x3e4320b7fa64e431 + .quad 0x000356c560000000, 0xbe1b5803cdae772e + .quad 0x000371a738000000, 0xbe28aac6ab1d7560 + .quad 0x00038cae70000000, 0xbe47d13cd3d2b1a8 + .quad 0x0003a7db38000000, 0xbe48d30048af21b7 + .quad 0x0003c32dc0000000, 0x3e489d47242000f9 + .quad 0x0003dea650000000, 0xbe4f6e5eee525f6f + .quad 0x0003fa4508000000, 0xbe4a9bff22fa047f + .quad 0x0004160a20000000, 0x3e3f72e29f84325c + .quad 0x000431f5d8000000, 0x3e350a896dc70444 + .quad 0x00044e0860000000, 0x3e18624b40c4dbd0 + .quad 0x00046a41f0000000, 0xbe4717fd446d7686 + .quad 0x000486a2b8000000, 0xbe41f6197f61f2e2 + .quad 0x0004a32af0000000, 0x3e2afa7bcce5b17a + .quad 0x0004bfdad8000000, 0xbe464eaec715e343 + .quad 0x0004dcb298000000, 0x3e3fddd0d63b36ef + .quad 0x0004f9b278000000, 0xbe362d35952cc275 + .quad 0x000516daa0000000, 0x3e467b320e0897a9 + .quad 0x0005342b58000000, 0xbe362b07e20f57c4 + .quad 0x000551a4c8000000, 0x3e42ec9076297631 + .quad 0x00056f4738000000, 0xbe34ad8259913500 + .quad 0x00058d12d8000000, 0xbe4b41c016d6a1ea + .quad 0x0005ab07e0000000, 0xbe45bd5eb539b67f + .quad 0x0005c92688000000, 0x3e42ca35b80e258e + .quad 0x0005e76f18000000, 0xbe4296f5bc8b20da + .quad 0x000605e1b8000000, 0x3e376dc08b076f59 + .quad 0x0006247eb0000000, 0x3e0d2ac258f87d03 + .quad 0x0006434638000000, 0xbe4999e701c483c7 + .quad 0x0006623880000000, 0x3e42a91124893ecf + .quad 0x00068155d8000000, 0xbe4d9ab467bf1d47 + .quad 0x0006a09e68000000, 0xbe380c4336f74d05 + .quad 0x0006c01278000000, 0xbe47a12a08944ab3 + .quad 0x0006dfb240000000, 0xbe4cd72e886ef8ea + .quad 0x0006ff7df8000000, 0x3e3519483cf87e1b + .quad 0x00071f75e8000000, 0x3e2d8bee7ba46e1e + .quad 0x00073f9a48000000, 0x3e24b02e77ab934a + .quad 0x00075feb58000000, 0xbe3bd98374091656 + .quad 0x0007806950000000, 0xbe00d1604f328fec + .quad 0x0007a11470000000, 0x3e4f580c36bea881 + .quad 0x0007c1ed00000000, 0x3e330c1327c49334 + .quad 0x0007e2f338000000, 0xbe330b19defa2fd4 + .quad 0x0008042758000000, 0xbe4e0f2f724f90cc + .quad 0x0008258998000000, 0x3e34cce128acf88b + .quad 0x0008471a48000000, 0xbe3dc385331ad094 + .quad 0x000868d998000000, 0x3e4a2497640720ed + .quad 0x00088ac7d8000000, 0x3e38a669966530bd + .quad 0x0008ace540000000, 0x3e415506dadd3e2b + .quad 0x0008cf3218000000, 0xbe34abb7410d55e3 + .quad 0x0008f1ae98000000, 0x3e31577362b98274 + .quad 0x0009145b08000000, 0x3e4c8ffe2c4530da + .quad 0x00093737b0000000, 0x3e29b8bc9e8a0388 + .quad 0x00095a44c8000000, 0x3e4e4290774da41b + .quad 0x00097d82a0000000, 0xbe00d8d83a30b6f8 + .quad 0x0009a0f170000000, 0x3e2940f737462137 + .quad 0x0009c49180000000, 0x3e451f8480e3e236 + .quad 0x0009e86318000000, 0x3e3e323231824ca8 + .quad 0x000a0c6678000000, 0x3e4aef2b2594d6d4 + .quad 0x000a309bf0000000, 0xbe4dae966539f470 + .quad 0x000a5503b0000000, 0x3e41f12ae45a1225 + .quad 0x000a799e10000000, 0x3e49859ac3796fd9 + .quad 0x000a9e6b58000000, 0xbe44301205e0a6de + .quad 0x000ac36bc0000000, 0xbe0606431f9234cb + .quad 0x000ae89f98000000, 0x3e35ad3ad5e8734d + .quad 0x000b0e0728000000, 0x3e38db66590842ad + .quad 0x000b33a2b8000000, 0x3e13c57ebdaff43a + .quad 0x000b597290000000, 0xbe40d536338e3bf7 + .quad 0x000b7f76f0000000, 0x3e47daf237553d84 + .quad 0x000ba5b030000000, 0x3e2420c930819679 + .quad 0x000bcc1e90000000, 0x3e12f074891ee83d + .quad 0x000bf2c258000000, 0x3e4eb8f0442046b8 + .quad 0x000c199be0000000, 0xbe43d56b1eeef9a7 + .quad 0x000c40ab60000000, 0xbd87c2c975903ef8 + .quad 0x000c67f130000000, 0xbe3a82eb4b5dec80 + .quad 0x000c8f6d98000000, 0xbe4fc8c257729a1e + .quad 0x000cb720e0000000, 0xbe48837cb757e1a1 + .quad 0x000cdf0b58000000, 0xbe4511e031dd83b5 + .quad 0x000d072d48000000, 0x3e403c4bdc687918 + .quad 0x000d2f8708000000, 0x3deb13e315bc2473 + .quad 0x000d5818e0000000, 0xbe4822dbc6d12fd3 + .quad 0x000d80e318000000, 0xbe3367c68447b063 + .quad 0x000da9e600000000, 0x3e4ed9942b84600d + .quad 0x000dd321f0000000, 0x3e480da3025b4aef + .quad 0x000dfc9730000000, 0x3e4bdcdaf5cb4656 + .quad 0x000e264618000000, 0xbe4852f6baf6c4f0 + .quad 0x000e502ee8000000, 0xbe1d30027630bb40 + .quad 0x000e7a51f8000000, 0x3e4e3a641a5aa459 + .quad 0x000ea4afa0000000, 0x3e452486cc2c7b9d + .quad 0x000ecf4830000000, 0xbe438cc07b927e77 + .quad 0x000efa1bf0000000, 0xbe39ea5d888e02de + .quad 0x000f252b38000000, 0xbe2288ad162f2d20 + .quad 0x000f507658000000, 0x3e4b722a033a7c26 + .quad 0x000f7bfdb0000000, 0xbe431a0f63b7625a + .quad 0x000fa7c180000000, 0x3e39e90d82e90a7e + .quad 0x000fd3c228000000, 0x3e4c7b8f884badd2 + /*== poly_coeff[4] ==*/ + .align 32 + .quad 0x3f81111168877F38, 0x3f81111168877F38, 0x3f81111168877F38, 0x3f81111168877F38 /* coeff5 */ + .quad 0x3fa55555C2A9C0F3, 0x3fa55555C2A9C0F3, 0x3fa55555C2A9C0F3, 0x3fa55555C2A9C0F3 /* coeff4 */ + .quad 0x3fc555555555541D, 0x3fc555555555541D, 0x3fc555555555541D, 0x3fc555555555541D /* coeff3 */ + .quad 0x3fdFFFFFFFFFFE5C, 0x3fdFFFFFFFFFFE5C, 0x3fdFFFFFFFFFFE5C, 0x3fdFFFFFFFFFFE5C /* coeff2 */ + /*== Log2e ==*/ + .align 32 + .quad 0x40671547652B82FE, 0x40671547652B82FE, 0x40671547652B82FE, 0x40671547652B82FE + /*== L2H ==*/ + .align 32 + .quad 0x3f762e42fef80000, 0x3f762e42fef80000, 0x3f762e42fef80000, 0x3f762e42fef80000 + /*== L2L ==*/ + .align 32 + .quad 0x3d41cf79abc9e3b4, 0x3d41cf79abc9e3b4, 0x3d41cf79abc9e3b4, 0x3d41cf79abc9e3b4 + /*== ExpAddConst ==*/ + .align 32 + .quad 0x42f80000001ff800, 0x42f80000001ff800, 0x42f80000001ff800, 0x42f80000001ff800 + /*== IndexMask ==*/ + .align 32 + .quad 0x00000000000007f0, 0x00000000000007f0, 0x00000000000007f0, 0x00000000000007f0 + /*== ExpMask ==*/ + .align 32 + .quad 0x00000000003ff800, 0x00000000003ff800, 0x00000000003ff800, 0x00000000003ff800 + /*== MOne ==*/ + .align 32 + .quad 0xbff0000000000000, 0xbff0000000000000, 0xbff0000000000000, 0xbff0000000000000 + /*== AbsMask ==*/ + .align 32 + .quad 0x7fffffffffffffff, 0x7fffffffffffffff, 0x7fffffffffffffff, 0x7fffffffffffffff + /*== Threshold ==*/ + .align 32 + .quad 0x40861DA04CBAFE43, 0x40861DA04CBAFE43, 0x40861DA04CBAFE43, 0x40861DA04CBAFE43 + /*== L2 ==*/ + .align 32 + .quad 0x3f762e42fefa39ef, 0x3f762e42fefa39ef, 0x3f762e42fefa39ef, 0x3f762e42fefa39ef + .align 32 + .type __svml_dexpm1_data_internal,@object + .size __svml_dexpm1_data_internal,.-__svml_dexpm1_data_internal diff --git a/sysdeps/x86_64/fpu/multiarch/svml_d_expm18_core-avx2.S b/sysdeps/x86_64/fpu/multiarch/svml_d_expm18_core-avx2.S new file mode 100644 index 0000000..3b75d1d --- /dev/null +++ b/sysdeps/x86_64/fpu/multiarch/svml_d_expm18_core-avx2.S @@ -0,0 +1,20 @@ +/* AVX2 version of vectorized expm1, vector length is 8. + Copyright (C) 2021 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + <https://www.gnu.org/licenses/>. */ + +#define _ZGVeN8v_expm1 _ZGVeN8v_expm1_avx2_wrapper +#include "../svml_d_expm18_core.S" diff --git a/sysdeps/x86_64/fpu/multiarch/svml_d_expm18_core.c b/sysdeps/x86_64/fpu/multiarch/svml_d_expm18_core.c new file mode 100644 index 0000000..860edf6 --- /dev/null +++ b/sysdeps/x86_64/fpu/multiarch/svml_d_expm18_core.c @@ -0,0 +1,27 @@ +/* Multiple versions of vectorized expm1, vector length is 8. + Copyright (C) 2021 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + <https://www.gnu.org/licenses/>. */ + +#define SYMBOL_NAME _ZGVeN8v_expm1 +#include "ifunc-mathvec-avx512-skx.h" + +libc_ifunc_redirected (REDIRECT_NAME, SYMBOL_NAME, IFUNC_SELECTOR ()); + +#ifdef SHARED +__hidden_ver1 (_ZGVeN8v_expm1, __GI__ZGVeN8v_expm1, __redirect__ZGVeN8v_expm1) + __attribute__ ((visibility ("hidden"))); +#endif diff --git a/sysdeps/x86_64/fpu/multiarch/svml_d_expm18_core_avx512.S b/sysdeps/x86_64/fpu/multiarch/svml_d_expm18_core_avx512.S new file mode 100644 index 0000000..64cee91 --- /dev/null +++ b/sysdeps/x86_64/fpu/multiarch/svml_d_expm18_core_avx512.S @@ -0,0 +1,334 @@ +/* Function expm1 vectorized with AVX-512. + Copyright (C) 2021 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + https://www.gnu.org/licenses/. */ + +/* + * ALGORITHM DESCRIPTION: + * After computing exp(x) in high-low parts, an accurate computation is performed to obtain exp(x)-1 + * Typical exp() implementation, except that: + * - tables are small (16 elements), allowing for fast gathers + * - all arguments processed in the main path + * - final VSCALEF assists branch-free design (correct overflow/underflow and special case responses) + * - a VAND is used to ensure the reduced argument |R|<2, even for large inputs + * - RZ mode used to avoid oveflow to +/-Inf for x*log2(e); helps with special case handling + * + * + */ + +/* Offsets for data table __svml_dexpm1_data_internal_avx512 + */ +#define Exp_tbl_H 0 +#define Exp_tbl_L 128 +#define L2E 256 +#define Shifter 320 +#define Threshold 384 +#define SgnMask 448 +#define L2H 512 +#define L2L 576 +#define ZThres 640 +#define EMask 704 +#define poly_coeff7 768 +#define poly_coeff6 832 +#define poly_coeff5 896 +#define poly_coeff4 960 +#define poly_coeff3 1024 +#define poly_coeff2 1088 +#define One 1152 + +#include <sysdep.h> + + .text + .section .text.evex512,"ax",@progbits +ENTRY(_ZGVeN8v_expm1_skx) + pushq %rbp + cfi_def_cfa_offset(16) + movq %rsp, %rbp + cfi_def_cfa(6, 16) + cfi_offset(6, -16) + andq $-64, %rsp + subq $192, %rsp + vmovups L2E+__svml_dexpm1_data_internal_avx512(%rip), %zmm6 + vmovups Shifter+__svml_dexpm1_data_internal_avx512(%rip), %zmm4 + vmovups L2H+__svml_dexpm1_data_internal_avx512(%rip), %zmm11 + vmovups L2L+__svml_dexpm1_data_internal_avx512(%rip), %zmm5 + vmovups Threshold+__svml_dexpm1_data_internal_avx512(%rip), %zmm3 + vmovups poly_coeff5+__svml_dexpm1_data_internal_avx512(%rip), %zmm13 + vmovups poly_coeff4+__svml_dexpm1_data_internal_avx512(%rip), %zmm15 + +/* polynomial */ + vmovups poly_coeff7+__svml_dexpm1_data_internal_avx512(%rip), %zmm12 + +/* set Z0=max(Z0, -128.0) */ + vmovups ZThres+__svml_dexpm1_data_internal_avx512(%rip), %zmm8 + vmovups poly_coeff3+__svml_dexpm1_data_internal_avx512(%rip), %zmm14 + vmovups __svml_dexpm1_data_internal_avx512(%rip), %zmm9 + vmovaps %zmm0, %zmm2 + +/* 2^(52-4)*1.5 + x * log2(e) */ + vfmadd213pd {rn-sae}, %zmm4, %zmm2, %zmm6 + vmovups Exp_tbl_L+__svml_dexpm1_data_internal_avx512(%rip), %zmm0 + vcmppd $21, {sae}, %zmm3, %zmm2, %k0 + +/* Z0 ~ x*log2(e), rounded to 4 fractional bits */ + vsubpd {rn-sae}, %zmm4, %zmm6, %zmm7 + vpermt2pd Exp_tbl_H+64+__svml_dexpm1_data_internal_avx512(%rip), %zmm6, %zmm9 + vpermt2pd Exp_tbl_L+64+__svml_dexpm1_data_internal_avx512(%rip), %zmm6, %zmm0 + vandpd SgnMask+__svml_dexpm1_data_internal_avx512(%rip), %zmm2, %zmm1 + +/* R = x - Z0*log(2) */ + vfnmadd213pd {rn-sae}, %zmm2, %zmm7, %zmm11 + vmaxpd {sae}, %zmm8, %zmm7, %zmm10 + vfnmadd231pd {rn-sae}, %zmm7, %zmm5, %zmm11 + kmovw %k0, %edx + +/* ensure |R|<2 even for special cases */ + vandpd EMask+__svml_dexpm1_data_internal_avx512(%rip), %zmm11, %zmm3 + vmovups poly_coeff6+__svml_dexpm1_data_internal_avx512(%rip), %zmm11 + +/* scale Th */ + vscalefpd {rn-sae}, %zmm10, %zmm9, %zmm4 + vfmadd231pd {rn-sae}, %zmm3, %zmm13, %zmm15 + vfmadd231pd {rn-sae}, %zmm3, %zmm12, %zmm11 + vmovups poly_coeff2+__svml_dexpm1_data_internal_avx512(%rip), %zmm12 + vmulpd {rn-sae}, %zmm3, %zmm3, %zmm13 + vfmadd231pd {rn-sae}, %zmm3, %zmm14, %zmm12 + vfmadd213pd {rn-sae}, %zmm15, %zmm13, %zmm11 + vfmadd213pd {rn-sae}, %zmm12, %zmm13, %zmm11 + +/* Tlr + R+ R*Poly */ + vfmadd213pd {rn-sae}, %zmm0, %zmm13, %zmm11 + +/* Th - 1 */ + vmovups One+__svml_dexpm1_data_internal_avx512(%rip), %zmm0 + vaddpd {rn-sae}, %zmm3, %zmm11, %zmm14 + vsubpd {rn-sae}, %zmm0, %zmm4, %zmm15 + +/* (Th-1)+Th*(Tlr + R+ R*Poly) */ + vfmadd213pd {rn-sae}, %zmm15, %zmm14, %zmm4 + vorpd %zmm1, %zmm4, %zmm0 + testl %edx, %edx + +/* Go to special inputs processing branch */ + jne L(SPECIAL_VALUES_BRANCH) + # LOE rbx r12 r13 r14 r15 edx zmm0 zmm2 + +/* Restore registers + * and exit the function + */ + +L(EXIT): + movq %rbp, %rsp + popq %rbp + cfi_def_cfa(7, 8) + cfi_restore(6) + ret + cfi_def_cfa(6, 16) + cfi_offset(6, -16) + +/* Branch to process + * special inputs + */ + +L(SPECIAL_VALUES_BRANCH): + vmovups %zmm2, 64(%rsp) + vmovups %zmm0, 128(%rsp) + # LOE rbx r12 r13 r14 r15 edx zmm0 + + xorl %eax, %eax + # LOE rbx r12 r13 r14 r15 eax edx + + vzeroupper + movq %r12, 16(%rsp) + /* DW_CFA_expression: r12 (r12) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -64; DW_OP_and; DW_OP_const4s: -176; DW_OP_plus) */ + .cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x50, 0xff, 0xff, 0xff, 0x22 + movl %eax, %r12d + movq %r13, 8(%rsp) + /* DW_CFA_expression: r13 (r13) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -64; DW_OP_and; DW_OP_const4s: -184; DW_OP_plus) */ + .cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x48, 0xff, 0xff, 0xff, 0x22 + movl %edx, %r13d + movq %r14, (%rsp) + /* DW_CFA_expression: r14 (r14) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -64; DW_OP_and; DW_OP_const4s: -192; DW_OP_plus) */ + .cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x40, 0xff, 0xff, 0xff, 0x22 + # LOE rbx r15 r12d r13d + +/* Range mask + * bits check + */ + +L(RANGEMASK_CHECK): + btl %r12d, %r13d + +/* Call scalar math function */ + jc L(SCALAR_MATH_CALL) + # LOE rbx r15 r12d r13d + +/* Special inputs + * processing loop + */ + +L(SPECIAL_VALUES_LOOP): + incl %r12d + cmpl $8, %r12d + +/* Check bits in range mask */ + jl L(RANGEMASK_CHECK) + # LOE rbx r15 r12d r13d + + movq 16(%rsp), %r12 + cfi_restore(12) + movq 8(%rsp), %r13 + cfi_restore(13) + movq (%rsp), %r14 + cfi_restore(14) + vmovups 128(%rsp), %zmm0 + +/* Go to exit */ + jmp L(EXIT) + /* DW_CFA_expression: r12 (r12) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -64; DW_OP_and; DW_OP_const4s: -176; DW_OP_plus) */ + .cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x50, 0xff, 0xff, 0xff, 0x22 + /* DW_CFA_expression: r13 (r13) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -64; DW_OP_and; DW_OP_const4s: -184; DW_OP_plus) */ + .cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x48, 0xff, 0xff, 0xff, 0x22 + /* DW_CFA_expression: r14 (r14) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -64; DW_OP_and; DW_OP_const4s: -192; DW_OP_plus) */ + .cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x40, 0xff, 0xff, 0xff, 0x22 + # LOE rbx r12 r13 r14 r15 zmm0 + +/* Scalar math fucntion call + * to process special input + */ + +L(SCALAR_MATH_CALL): + movl %r12d, %r14d + movsd 64(%rsp,%r14,8), %xmm0 + call expm1@PLT + # LOE rbx r14 r15 r12d r13d xmm0 + + movsd %xmm0, 128(%rsp,%r14,8) + +/* Process special inputs in loop */ + jmp L(SPECIAL_VALUES_LOOP) + # LOE rbx r15 r12d r13d +END(_ZGVeN8v_expm1_skx) + + .section .rodata, "a" + .align 64 + +#ifdef __svml_dexpm1_data_internal_avx512_typedef +typedef unsigned int VUINT32; +typedef struct { + __declspec(align(64)) VUINT32 Exp_tbl_H[16][2]; + __declspec(align(64)) VUINT32 Exp_tbl_L[16][2]; + __declspec(align(64)) VUINT32 L2E[8][2]; + __declspec(align(64)) VUINT32 Shifter[8][2]; + __declspec(align(64)) VUINT32 Threshold[8][2]; + __declspec(align(64)) VUINT32 SgnMask[8][2]; + __declspec(align(64)) VUINT32 L2H[8][2]; + __declspec(align(64)) VUINT32 L2L[8][2]; + __declspec(align(64)) VUINT32 ZThres[8][2]; + __declspec(align(64)) VUINT32 EMask[8][2]; + __declspec(align(64)) VUINT32 poly_coeff7[8][2]; + __declspec(align(64)) VUINT32 poly_coeff6[8][2]; + __declspec(align(64)) VUINT32 poly_coeff5[8][2]; + __declspec(align(64)) VUINT32 poly_coeff4[8][2]; + __declspec(align(64)) VUINT32 poly_coeff3[8][2]; + __declspec(align(64)) VUINT32 poly_coeff2[8][2]; + __declspec(align(64)) VUINT32 One[8][2]; + } __svml_dexpm1_data_internal_avx512; +#endif +__svml_dexpm1_data_internal_avx512: + /*== Exp_tbl_H ==*/ + .quad 0x3ff0000000000000 + .quad 0x3ff0b5586cf9890f + .quad 0x3ff172b83c7d517b + .quad 0x3ff2387a6e756238 + .quad 0x3ff306fe0a31b715 + .quad 0x3ff3dea64c123422 + .quad 0x3ff4bfdad5362a27 + .quad 0x3ff5ab07dd485429 + .quad 0x3ff6a09e667f3bcd + .quad 0x3ff7a11473eb0187 + .quad 0x3ff8ace5422aa0db + .quad 0x3ff9c49182a3f090 + .quad 0x3ffae89f995ad3ad + .quad 0x3ffc199bdd85529c + .quad 0x3ffd5818dcfba487 + .quad 0x3ffea4afa2a490da + /*== Exp_tbl_L ==*/ + .align 64 + .quad 0x0000000000000000 + .quad 0x3c979aa65d837b6d + .quad 0xbc801b15eaa59348 + .quad 0x3c968efde3a8a894 + .quad 0x3c834d754db0abb6 + .quad 0x3c859f48a72a4c6d + .quad 0x3c7690cebb7aafb0 + .quad 0x3c9063e1e21c5409 + .quad 0xbc93b3efbf5e2228 + .quad 0xbc7b32dcb94da51d + .quad 0x3c8db72fc1f0eab4 + .quad 0x3c71affc2b91ce27 + .quad 0x3c8c1a7792cb3387 + .quad 0x3c736eae30af0cb3 + .quad 0x3c74a385a63d07a7 + .quad 0xbc8ff7128fd391f0 + /*== log2(e) ==*/ + .align 64 + .quad 0x3ff71547652B82FE, 0x3ff71547652B82FE, 0x3ff71547652B82FE, 0x3ff71547652B82FE, 0x3ff71547652B82FE, 0x3ff71547652B82FE, 0x3ff71547652B82FE, 0x3ff71547652B82FE + /*== Shifter=2^(52-4)*1.5 ==*/ + .align 64 + .quad 0x42f8000000003ff0, 0x42f8000000003ff0, 0x42f8000000003ff0, 0x42f8000000003ff0, 0x42f8000000003ff0, 0x42f8000000003ff0, 0x42f8000000003ff0, 0x42f8000000003ff0 + /*== Threshold ==*/ + .align 64 + .quad 0x40861DA04CBAFE44, 0x40861DA04CBAFE44, 0x40861DA04CBAFE44, 0x40861DA04CBAFE44, 0x40861DA04CBAFE44, 0x40861DA04CBAFE44, 0x40861DA04CBAFE44, 0x40861DA04CBAFE44 + /*== Sgn ==*/ + .align 64 + .quad 0x8000000000000000, 0x8000000000000000, 0x8000000000000000, 0x8000000000000000, 0x8000000000000000, 0x8000000000000000, 0x8000000000000000, 0x8000000000000000 + /*== L2H = log(2)_high ==*/ + .align 64 + .quad 0x3fe62e42fefa39ef, 0x3fe62e42fefa39ef, 0x3fe62e42fefa39ef, 0x3fe62e42fefa39ef, 0x3fe62e42fefa39ef, 0x3fe62e42fefa39ef, 0x3fe62e42fefa39ef, 0x3fe62e42fefa39ef + /*== L2L = log(2)_low ==*/ + .align 64 + .quad 0x3c7abc9e3b39803f, 0x3c7abc9e3b39803f, 0x3c7abc9e3b39803f, 0x3c7abc9e3b39803f, 0x3c7abc9e3b39803f, 0x3c7abc9e3b39803f, 0x3c7abc9e3b39803f, 0x3c7abc9e3b39803f + /*== ZThres ==*/ + .align 64 + .quad 0xc060000000000000, 0xc060000000000000, 0xc060000000000000, 0xc060000000000000, 0xc060000000000000, 0xc060000000000000, 0xc060000000000000, 0xc060000000000000 + /*== EMask ==*/ + .align 64 + .quad 0xbfffffffffffffff, 0xbfffffffffffffff, 0xbfffffffffffffff, 0xbfffffffffffffff, 0xbfffffffffffffff, 0xbfffffffffffffff, 0xbfffffffffffffff, 0xbfffffffffffffff + /*== poly_coeff7 ==*/ + .align 64 + .quad 0x3f2a020410303d8a, 0x3f2a020410303d8a, 0x3f2a020410303d8a, 0x3f2a020410303d8a, 0x3f2a020410303d8a, 0x3f2a020410303d8a, 0x3f2a020410303d8a, 0x3f2a020410303d8a + /*== poly_coeff6 ==*/ + .align 64 + .quad 0x3f56c1c38e164a2f, 0x3f56c1c38e164a2f, 0x3f56c1c38e164a2f, 0x3f56c1c38e164a2f, 0x3f56c1c38e164a2f, 0x3f56c1c38e164a2f, 0x3f56c1c38e164a2f, 0x3f56c1c38e164a2f + /*== poly_coeff5 ==*/ + .align 64 + .quad 0x3f81111110865214, 0x3f81111110865214, 0x3f81111110865214, 0x3f81111110865214, 0x3f81111110865214, 0x3f81111110865214, 0x3f81111110865214, 0x3f81111110865214 + /*== poly_coeff4 ==*/ + .align 64 + .quad 0x3fa5555554ad3d06, 0x3fa5555554ad3d06, 0x3fa5555554ad3d06, 0x3fa5555554ad3d06, 0x3fa5555554ad3d06, 0x3fa5555554ad3d06, 0x3fa5555554ad3d06, 0x3fa5555554ad3d06 + /*== poly_coeff3 ==*/ + .align 64 + .quad 0x3fc5555555555656, 0x3fc5555555555656, 0x3fc5555555555656, 0x3fc5555555555656, 0x3fc5555555555656, 0x3fc5555555555656, 0x3fc5555555555656, 0x3fc5555555555656 + /*== poly_coeff2 ==*/ + .align 64 + .quad 0x3fe00000000000a2, 0x3fe00000000000a2, 0x3fe00000000000a2, 0x3fe00000000000a2, 0x3fe00000000000a2, 0x3fe00000000000a2, 0x3fe00000000000a2, 0x3fe00000000000a2 + /*== One ==*/ + .align 64 + .quad 0x3ff0000000000000, 0x3ff0000000000000, 0x3ff0000000000000, 0x3ff0000000000000, 0x3ff0000000000000, 0x3ff0000000000000, 0x3ff0000000000000, 0x3ff0000000000000 + .align 64 + .type __svml_dexpm1_data_internal_avx512,@object + .size __svml_dexpm1_data_internal_avx512,.-__svml_dexpm1_data_internal_avx512 diff --git a/sysdeps/x86_64/fpu/multiarch/svml_s_expm1f16_core-avx2.S b/sysdeps/x86_64/fpu/multiarch/svml_s_expm1f16_core-avx2.S new file mode 100644 index 0000000..a2a8699 --- /dev/null +++ b/sysdeps/x86_64/fpu/multiarch/svml_s_expm1f16_core-avx2.S @@ -0,0 +1,20 @@ +/* AVX2 version of vectorized expm1f. + Copyright (C) 2021 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + <https://www.gnu.org/licenses/>. */ + +#define _ZGVeN16v_expm1f _ZGVeN16v_expm1f_avx2_wrapper +#include "../svml_s_expm1f16_core.S" diff --git a/sysdeps/x86_64/fpu/multiarch/svml_s_expm1f16_core.c b/sysdeps/x86_64/fpu/multiarch/svml_s_expm1f16_core.c new file mode 100644 index 0000000..8007d1e --- /dev/null +++ b/sysdeps/x86_64/fpu/multiarch/svml_s_expm1f16_core.c @@ -0,0 +1,28 @@ +/* Multiple versions of vectorized expm1f, vector length is 16. + Copyright (C) 2021 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + <https://www.gnu.org/licenses/>. */ + +#define SYMBOL_NAME _ZGVeN16v_expm1f +#include "ifunc-mathvec-avx512-skx.h" + +libc_ifunc_redirected (REDIRECT_NAME, SYMBOL_NAME, IFUNC_SELECTOR ()); + +#ifdef SHARED +__hidden_ver1 (_ZGVeN16v_expm1f, __GI__ZGVeN16v_expm1f, + __redirect__ZGVeN16v_expm1f) + __attribute__ ((visibility ("hidden"))); +#endif diff --git a/sysdeps/x86_64/fpu/multiarch/svml_s_expm1f16_core_avx512.S b/sysdeps/x86_64/fpu/multiarch/svml_s_expm1f16_core_avx512.S new file mode 100644 index 0000000..5b0dcde --- /dev/null +++ b/sysdeps/x86_64/fpu/multiarch/svml_s_expm1f16_core_avx512.S @@ -0,0 +1,281 @@ +/* Function expm1f vectorized with AVX-512. + Copyright (C) 2021 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + https://www.gnu.org/licenses/. */ + +/* + * ALGORITHM DESCRIPTION: + * After computing exp(x) in high-low parts, an accurate computation is performed to obtain exp(x)-1 + * Typical exp() implementation, except that: + * - tables are small (32 elements), allowing for fast gathers + * - all arguments processed in the main path + * - final VSCALEF assists branch-free design (correct overflow/underflow and special case responses) + * - a VAND is used to ensure the reduced argument |R|<2, even for large inputs + * - RZ mode used to avoid oveflow to +/-Inf for x*log2(e); helps with special case handling + * + * + */ + +/* Offsets for data table __svml_sexpm1_data_internal_avx512 + */ +#define Exp_tbl_H 0 +#define Exp_tbl_L 128 +#define L2E 256 +#define Shifter 320 +#define Threshold 384 +#define SgnMask 448 +#define L2H 512 +#define L2L 576 +#define EMask 640 +#define poly_coeff3 704 +#define poly_coeff2 768 +#define One 832 + +#include <sysdep.h> + + .text + .section .text.exex512,"ax",@progbits +ENTRY(_ZGVeN16v_expm1f_skx) + pushq %rbp + cfi_def_cfa_offset(16) + movq %rsp, %rbp + cfi_def_cfa(6, 16) + cfi_offset(6, -16) + andq $-64, %rsp + subq $192, %rsp + vmovups L2E+__svml_sexpm1_data_internal_avx512(%rip), %zmm5 + vmovups Shifter+__svml_sexpm1_data_internal_avx512(%rip), %zmm3 + vmovups L2H+__svml_sexpm1_data_internal_avx512(%rip), %zmm8 + vmovups L2L+__svml_sexpm1_data_internal_avx512(%rip), %zmm4 + vmovups __svml_sexpm1_data_internal_avx512(%rip), %zmm6 + +/* polynomial */ + vmovups poly_coeff3+__svml_sexpm1_data_internal_avx512(%rip), %zmm9 + vmovups poly_coeff2+__svml_sexpm1_data_internal_avx512(%rip), %zmm12 + vmovups Exp_tbl_L+__svml_sexpm1_data_internal_avx512(%rip), %zmm11 + vmovups Threshold+__svml_sexpm1_data_internal_avx512(%rip), %zmm2 + +/* Th - 1 */ + vmovups One+__svml_sexpm1_data_internal_avx512(%rip), %zmm14 + vmovaps %zmm0, %zmm1 + +/* 2^(52-5)*1.5 + x * log2(e) */ + vfmadd213ps {rn-sae}, %zmm3, %zmm1, %zmm5 + vcmpps $29, {sae}, %zmm2, %zmm1, %k0 + +/* Z0 ~ x*log2(e), rounded to 5 fractional bits */ + vsubps {rn-sae}, %zmm3, %zmm5, %zmm7 + vpermt2ps Exp_tbl_H+64+__svml_sexpm1_data_internal_avx512(%rip), %zmm5, %zmm6 + vpermt2ps Exp_tbl_L+64+__svml_sexpm1_data_internal_avx512(%rip), %zmm5, %zmm11 + vandps SgnMask+__svml_sexpm1_data_internal_avx512(%rip), %zmm1, %zmm0 + +/* R = x - Z0*log(2) */ + vfnmadd213ps {rn-sae}, %zmm1, %zmm7, %zmm8 + +/* scale Th */ + vscalefps {rn-sae}, %zmm7, %zmm6, %zmm2 + vfnmadd231ps {rn-sae}, %zmm7, %zmm4, %zmm8 + kmovw %k0, %edx + +/* ensure |R|<2 even for special cases */ + vandps EMask+__svml_sexpm1_data_internal_avx512(%rip), %zmm8, %zmm13 + vsubps {rn-sae}, %zmm14, %zmm2, %zmm8 + vmulps {rn-sae}, %zmm13, %zmm13, %zmm10 + vfmadd231ps {rn-sae}, %zmm13, %zmm9, %zmm12 + +/* Tlr + R+ R2*Poly */ + vfmadd213ps {rn-sae}, %zmm11, %zmm10, %zmm12 + vaddps {rn-sae}, %zmm13, %zmm12, %zmm15 + +/* (Th-1)+Th*(Tlr + R+ R*Poly) */ + vfmadd213ps {rn-sae}, %zmm8, %zmm15, %zmm2 + vorps %zmm0, %zmm2, %zmm0 + testl %edx, %edx + +/* Go to special inputs processing branch */ + jne L(SPECIAL_VALUES_BRANCH) + # LOE rbx r12 r13 r14 r15 edx zmm0 zmm1 + +/* Restore registers + * and exit the function + */ + +L(EXIT): + movq %rbp, %rsp + popq %rbp + cfi_def_cfa(7, 8) + cfi_restore(6) + ret + cfi_def_cfa(6, 16) + cfi_offset(6, -16) + +/* Branch to process + * special inputs + */ + +L(SPECIAL_VALUES_BRANCH): + vmovups %zmm1, 64(%rsp) + vmovups %zmm0, 128(%rsp) + # LOE rbx r12 r13 r14 r15 edx zmm0 + + xorl %eax, %eax + # LOE rbx r12 r13 r14 r15 eax edx + + vzeroupper + movq %r12, 16(%rsp) + /* DW_CFA_expression: r12 (r12) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -64; DW_OP_and; DW_OP_const4s: -176; DW_OP_plus) */ + .cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x50, 0xff, 0xff, 0xff, 0x22 + movl %eax, %r12d + movq %r13, 8(%rsp) + /* DW_CFA_expression: r13 (r13) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -64; DW_OP_and; DW_OP_const4s: -184; DW_OP_plus) */ + .cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x48, 0xff, 0xff, 0xff, 0x22 + movl %edx, %r13d + movq %r14, (%rsp) + /* DW_CFA_expression: r14 (r14) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -64; DW_OP_and; DW_OP_const4s: -192; DW_OP_plus) */ + .cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x40, 0xff, 0xff, 0xff, 0x22 + # LOE rbx r15 r12d r13d + +/* Range mask + * bits check + */ + +L(RANGEMASK_CHECK): + btl %r12d, %r13d + +/* Call scalar math function */ + jc L(SCALAR_MATH_CALL) + # LOE rbx r15 r12d r13d + +/* Special inputs + * processing loop + */ + +L(SPECIAL_VALUES_LOOP): + incl %r12d + cmpl $16, %r12d + +/* Check bits in range mask */ + jl L(RANGEMASK_CHECK) + # LOE rbx r15 r12d r13d + + movq 16(%rsp), %r12 + cfi_restore(12) + movq 8(%rsp), %r13 + cfi_restore(13) + movq (%rsp), %r14 + cfi_restore(14) + vmovups 128(%rsp), %zmm0 + +/* Go to exit */ + jmp L(EXIT) + /* DW_CFA_expression: r12 (r12) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -64; DW_OP_and; DW_OP_const4s: -176; DW_OP_plus) */ + .cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x50, 0xff, 0xff, 0xff, 0x22 + /* DW_CFA_expression: r13 (r13) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -64; DW_OP_and; DW_OP_const4s: -184; DW_OP_plus) */ + .cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x48, 0xff, 0xff, 0xff, 0x22 + /* DW_CFA_expression: r14 (r14) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -64; DW_OP_and; DW_OP_const4s: -192; DW_OP_plus) */ + .cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0x40, 0xff, 0xff, 0xff, 0x22 + # LOE rbx r12 r13 r14 r15 zmm0 + +/* Scalar math fucntion call + * to process special input + */ + +L(SCALAR_MATH_CALL): + movl %r12d, %r14d + movss 64(%rsp,%r14,4), %xmm0 + call expm1f@PLT + # LOE rbx r14 r15 r12d r13d xmm0 + + movss %xmm0, 128(%rsp,%r14,4) + +/* Process special inputs in loop */ + jmp L(SPECIAL_VALUES_LOOP) + # LOE rbx r15 r12d r13d +END(_ZGVeN16v_expm1f_skx) + + .section .rodata, "a" + .align 64 + +#ifdef __svml_sexpm1_data_internal_avx512_typedef +typedef unsigned int VUINT32; +typedef struct { + __declspec(align(64)) VUINT32 Exp_tbl_H[32][1]; + __declspec(align(64)) VUINT32 Exp_tbl_L[32][1]; + __declspec(align(64)) VUINT32 L2E[16][1]; + __declspec(align(64)) VUINT32 Shifter[16][1]; + __declspec(align(64)) VUINT32 Threshold[16][1]; + __declspec(align(64)) VUINT32 SgnMask[16][1]; + __declspec(align(64)) VUINT32 L2H[16][1]; + __declspec(align(64)) VUINT32 L2L[16][1]; + __declspec(align(64)) VUINT32 EMask[16][1]; + __declspec(align(64)) VUINT32 poly_coeff3[16][1]; + __declspec(align(64)) VUINT32 poly_coeff2[16][1]; + __declspec(align(64)) VUINT32 One[16][1]; + } __svml_sexpm1_data_internal_avx512; +#endif +__svml_sexpm1_data_internal_avx512: + /*== Exp_tbl_H ==*/ + .long 0x3f800000, 0x3f82cd87, 0x3f85aac3, 0x3f88980f + .long 0x3f8b95c2, 0x3f8ea43a, 0x3f91c3d3, 0x3f94f4f0 + .long 0x3f9837f0, 0x3f9b8d3a, 0x3f9ef532, 0x3fa27043 + .long 0x3fa5fed7, 0x3fa9a15b, 0x3fad583f, 0x3fb123f6 + .long 0x3fb504f3, 0x3fb8fbaf, 0x3fbd08a4, 0x3fc12c4d + .long 0x3fc5672a, 0x3fc9b9be, 0x3fce248c, 0x3fd2a81e + .long 0x3fd744fd, 0x3fdbfbb8, 0x3fe0ccdf, 0x3fe5b907 + .long 0x3feac0c7, 0x3fefe4ba, 0x3ff5257d, 0x3ffa83b3 + /*== Exp_tbl_L ==*/ + .align 64 + .long 0x00000000, 0xb34a3a0a, 0x3346cb6a, 0xb36ed17e + .long 0xb24e0611, 0xb3517dd9, 0x334b2482, 0xb31586de + .long 0x33092801, 0xb2e6f467, 0x331b85f2, 0x3099b6f1 + .long 0xb3051aa8, 0xb2e2a0da, 0xb2006c56, 0xb3365942 + .long 0x329302ae, 0x32c595dc, 0xb302e5a2, 0xb28e10a1 + .long 0x31b3d0e5, 0xb31a472b, 0x31d1daf2, 0xb305bf64 + .long 0xb27ce182, 0xb2f26443, 0xb1b4b0da, 0xb1da8a8f + .long 0xb1d290be, 0xb2d5b899, 0x31b0a147, 0xb2156afc + /*== log2(e) ==*/ + .align 64 + .long 0x3fB8AA3B, 0x3fB8AA3B, 0x3fB8AA3B, 0x3fB8AA3B, 0x3fB8AA3B, 0x3fB8AA3B, 0x3fB8AA3B, 0x3fB8AA3B, 0x3fB8AA3B, 0x3fB8AA3B, 0x3fB8AA3B, 0x3fB8AA3B, 0x3fB8AA3B, 0x3fB8AA3B, 0x3fB8AA3B, 0x3fB8AA3B + /*== Shifter=2^(23-5)*1.5 ==*/ + .align 64 + .long 0x48c00000, 0x48c00000, 0x48c00000, 0x48c00000, 0x48c00000, 0x48c00000, 0x48c00000, 0x48c00000, 0x48c00000, 0x48c00000, 0x48c00000, 0x48c00000, 0x48c00000, 0x48c00000, 0x48c00000, 0x48c00000 + /*== Threshold ==*/ + .align 64 + .long 0x42AD496B, 0x42AD496B, 0x42AD496B, 0x42AD496B, 0x42AD496B, 0x42AD496B, 0x42AD496B, 0x42AD496B, 0x42AD496B, 0x42AD496B, 0x42AD496B, 0x42AD496B, 0x42AD496B, 0x42AD496B, 0x42AD496B, 0x42AD496B + /*== Sgn ==*/ + .align 64 + .long 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000, 0x80000000 + /*== L2H = log(2)_high ==*/ + .align 64 + .long 0x3f317218, 0x3f317218, 0x3f317218, 0x3f317218, 0x3f317218, 0x3f317218, 0x3f317218, 0x3f317218, 0x3f317218, 0x3f317218, 0x3f317218, 0x3f317218, 0x3f317218, 0x3f317218, 0x3f317218, 0x3f317218 + /*== L2L = log(2)_low ==*/ + .align 64 + .long 0xb102e308, 0xb102e308, 0xb102e308, 0xb102e308, 0xb102e308, 0xb102e308, 0xb102e308, 0xb102e308, 0xb102e308, 0xb102e308, 0xb102e308, 0xb102e308, 0xb102e308, 0xb102e308, 0xb102e308, 0xb102e308 + /*== EMask ==*/ + .align 64 + .long 0xbfffffff, 0xbfffffff, 0xbfffffff, 0xbfffffff, 0xbfffffff, 0xbfffffff, 0xbfffffff, 0xbfffffff, 0xbfffffff, 0xbfffffff, 0xbfffffff, 0xbfffffff, 0xbfffffff, 0xbfffffff, 0xbfffffff, 0xbfffffff + /*== poly_coeff3 ==*/ + .align 64 + .long 0x3e2AABF3, 0x3e2AABF3, 0x3e2AABF3, 0x3e2AABF3, 0x3e2AABF3, 0x3e2AABF3, 0x3e2AABF3, 0x3e2AABF3, 0x3e2AABF3, 0x3e2AABF3, 0x3e2AABF3, 0x3e2AABF3, 0x3e2AABF3, 0x3e2AABF3, 0x3e2AABF3, 0x3e2AABF3 + /*== poly_coeff2 ==*/ + .align 64 + .long 0x3f0000F6, 0x3f0000F6, 0x3f0000F6, 0x3f0000F6, 0x3f0000F6, 0x3f0000F6, 0x3f0000F6, 0x3f0000F6, 0x3f0000F6, 0x3f0000F6, 0x3f0000F6, 0x3f0000F6, 0x3f0000F6, 0x3f0000F6, 0x3f0000F6, 0x3f0000F6 + /*== One ==*/ + .align 64 + .long 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000, 0x3f800000 + .align 64 + .type __svml_sexpm1_data_internal_avx512,@object + .size __svml_sexpm1_data_internal_avx512,.-__svml_sexpm1_data_internal_avx512 diff --git a/sysdeps/x86_64/fpu/multiarch/svml_s_expm1f4_core-sse2.S b/sysdeps/x86_64/fpu/multiarch/svml_s_expm1f4_core-sse2.S new file mode 100644 index 0000000..b4dbb77 --- /dev/null +++ b/sysdeps/x86_64/fpu/multiarch/svml_s_expm1f4_core-sse2.S @@ -0,0 +1,20 @@ +/* SSE2 version of vectorized expm1f, vector length is 4. + Copyright (C) 2021 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + <https://www.gnu.org/licenses/>. */ + +#define _ZGVbN4v_expm1f _ZGVbN4v_expm1f_sse2 +#include "../svml_s_expm1f4_core.S" diff --git a/sysdeps/x86_64/fpu/multiarch/svml_s_expm1f4_core.c b/sysdeps/x86_64/fpu/multiarch/svml_s_expm1f4_core.c new file mode 100644 index 0000000..f8ef125 --- /dev/null +++ b/sysdeps/x86_64/fpu/multiarch/svml_s_expm1f4_core.c @@ -0,0 +1,28 @@ +/* Multiple versions of vectorized expm1f, vector length is 4. + Copyright (C) 2021 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + <https://www.gnu.org/licenses/>. */ + +#define SYMBOL_NAME _ZGVbN4v_expm1f +#include "ifunc-mathvec-sse4_1.h" + +libc_ifunc_redirected (REDIRECT_NAME, SYMBOL_NAME, IFUNC_SELECTOR ()); + +#ifdef SHARED +__hidden_ver1 (_ZGVbN4v_expm1f, __GI__ZGVbN4v_expm1f, + __redirect__ZGVbN4v_expm1f) + __attribute__ ((visibility ("hidden"))); +#endif diff --git a/sysdeps/x86_64/fpu/multiarch/svml_s_expm1f4_core_sse4.S b/sysdeps/x86_64/fpu/multiarch/svml_s_expm1f4_core_sse4.S new file mode 100644 index 0000000..18770f6 --- /dev/null +++ b/sysdeps/x86_64/fpu/multiarch/svml_s_expm1f4_core_sse4.S @@ -0,0 +1,358 @@ +/* Function expm1f vectorized with SSE4. + Copyright (C) 2021 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + https://www.gnu.org/licenses/. */ + +/* + * ALGORITHM DESCRIPTION: + * + * N = (int)(x*2^k/log(2.0)), R = x - N*log(2)/2^k + * exp(x) = 2^(N/2^k) * poly(R) is computed in high-low parts + * expm1(x) = exp(x)-1 is then obtained via multi-precision computation + * + * + */ + +/* Offsets for data table __svml_sexpm1_data_internal + */ +#define Expm1_HA_table 0 +#define poly_coeff 512 +#define Log2e 576 +#define L2H 592 +#define L2L 608 +#define ExpAddConst 624 +#define IndexMask 640 +#define ExpMask 656 +#define MOne 672 +#define AbsMask 688 +#define Threshold 704 +#define L2 720 + +#include <sysdep.h> + + .text + .section .text.sse4,"ax",@progbits +ENTRY(_ZGVbN4v_expm1f_sse4) + pushq %rbp + cfi_def_cfa_offset(16) + movq %rsp, %rbp + cfi_def_cfa(6, 16) + cfi_offset(6, -16) + andq $-32, %rsp + subq $64, %rsp + movaps %xmm0, %xmm4 + movups Log2e+__svml_sexpm1_data_internal(%rip), %xmm9 + lea __svml_sexpm1_data_internal(%rip), %r8 + mulps %xmm0, %xmm9 + movups .FLT_10(%rip), %xmm5 + movups ExpAddConst+__svml_sexpm1_data_internal(%rip), %xmm2 + addps %xmm5, %xmm9 + +/* argument reduction */ + movups L2H+__svml_sexpm1_data_internal(%rip), %xmm6 + subps %xmm5, %xmm9 + mulps %xmm9, %xmm6 + addps %xmm9, %xmm2 + +/* table lookup */ + movdqu IndexMask+__svml_sexpm1_data_internal(%rip), %xmm12 + subps %xmm6, %xmm4 + pand %xmm2, %xmm12 + movups L2L+__svml_sexpm1_data_internal(%rip), %xmm7 + movups AbsMask+__svml_sexpm1_data_internal(%rip), %xmm3 + pshufd $1, %xmm12, %xmm10 + movaps %xmm3, %xmm8 + mulps %xmm9, %xmm7 + andps %xmm0, %xmm8 + cmpnleps Threshold+__svml_sexpm1_data_internal(%rip), %xmm8 + movd %xmm12, %edx + subps %xmm7, %xmm4 + movd %xmm10, %ecx + movmskps %xmm8, %eax + pshufd $2, %xmm12, %xmm11 + movaps %xmm4, %xmm7 + pshufd $3, %xmm12, %xmm13 + andnps %xmm0, %xmm3 + movd %xmm11, %esi + movd %xmm13, %edi + +/* polynomial */ + movups poly_coeff+__svml_sexpm1_data_internal(%rip), %xmm8 + movdqu ExpMask+__svml_sexpm1_data_internal(%rip), %xmm6 + movslq %edx, %rdx + pand %xmm6, %xmm2 + movslq %ecx, %rcx + pslld $14, %xmm2 + movslq %esi, %rsi + movslq %edi, %rdi + movq (%r8,%rdx), %xmm1 + movq (%r8,%rcx), %xmm14 + movq (%r8,%rsi), %xmm5 + movq (%r8,%rdi), %xmm15 + unpcklps %xmm14, %xmm1 + mulps %xmm4, %xmm8 + movaps %xmm1, %xmm10 + mulps %xmm4, %xmm7 + addps poly_coeff+16+__svml_sexpm1_data_internal(%rip), %xmm8 + unpcklps %xmm15, %xmm5 + movlhps %xmm5, %xmm10 + shufps $238, %xmm5, %xmm1 + orps %xmm2, %xmm10 + +/* T-1 */ + movups MOne+__svml_sexpm1_data_internal(%rip), %xmm9 + mulps %xmm2, %xmm1 + addps %xmm9, %xmm10 + mulps %xmm7, %xmm8 + addps %xmm1, %xmm10 + addps %xmm8, %xmm4 + movaps %xmm10, %xmm1 + subps %xmm9, %xmm1 + mulps %xmm1, %xmm4 + addps %xmm4, %xmm10 + orps %xmm3, %xmm10 + testl %eax, %eax + +/* Go to special inputs processing branch */ + jne L(SPECIAL_VALUES_BRANCH) + # LOE rbx r12 r13 r14 r15 eax xmm0 xmm10 + +/* Restore registers + * and exit the function + */ + +L(EXIT): + movaps %xmm10, %xmm0 + movq %rbp, %rsp + popq %rbp + cfi_def_cfa(7, 8) + cfi_restore(6) + ret + cfi_def_cfa(6, 16) + cfi_offset(6, -16) + +/* Branch to process + * special inputs + */ + +L(SPECIAL_VALUES_BRANCH): + movups %xmm0, 32(%rsp) + movups %xmm10, 48(%rsp) + # LOE rbx r12 r13 r14 r15 eax + + xorl %edx, %edx + movq %r12, 16(%rsp) + /* DW_CFA_expression: r12 (r12) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -48; DW_OP_plus) */ + .cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xd0, 0xff, 0xff, 0xff, 0x22 + movl %edx, %r12d + movq %r13, 8(%rsp) + /* DW_CFA_expression: r13 (r13) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -56; DW_OP_plus) */ + .cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xc8, 0xff, 0xff, 0xff, 0x22 + movl %eax, %r13d + movq %r14, (%rsp) + /* DW_CFA_expression: r14 (r14) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -64; DW_OP_plus) */ + .cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x22 + # LOE rbx r15 r12d r13d + +/* Range mask + * bits check + */ + +L(RANGEMASK_CHECK): + btl %r12d, %r13d + +/* Call scalar math function */ + jc L(SCALAR_MATH_CALL) + # LOE rbx r15 r12d r13d + +/* Special inputs + * processing loop + */ + +L(SPECIAL_VALUES_LOOP): + incl %r12d + cmpl $4, %r12d + +/* Check bits in range mask */ + jl L(RANGEMASK_CHECK) + # LOE rbx r15 r12d r13d + + movq 16(%rsp), %r12 + cfi_restore(12) + movq 8(%rsp), %r13 + cfi_restore(13) + movq (%rsp), %r14 + cfi_restore(14) + movups 48(%rsp), %xmm10 + +/* Go to exit */ + jmp L(EXIT) + /* DW_CFA_expression: r12 (r12) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -48; DW_OP_plus) */ + .cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xd0, 0xff, 0xff, 0xff, 0x22 + /* DW_CFA_expression: r13 (r13) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -56; DW_OP_plus) */ + .cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xc8, 0xff, 0xff, 0xff, 0x22 + /* DW_CFA_expression: r14 (r14) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -64; DW_OP_plus) */ + .cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xc0, 0xff, 0xff, 0xff, 0x22 + # LOE rbx r12 r13 r14 r15 xmm10 + +/* Scalar math fucntion call + * to process special input + */ + +L(SCALAR_MATH_CALL): + movl %r12d, %r14d + movss 32(%rsp,%r14,4), %xmm0 + call expm1f@PLT + # LOE rbx r14 r15 r12d r13d xmm0 + + movss %xmm0, 48(%rsp,%r14,4) + +/* Process special inputs in loop */ + jmp L(SPECIAL_VALUES_LOOP) + # LOE rbx r15 r12d r13d +END(_ZGVbN4v_expm1f_sse4) + + .section .rodata, "a" + .align 16 + +#ifdef __svml_sexpm1_data_internal_typedef +typedef unsigned int VUINT32; +typedef struct { + __declspec(align(16)) VUINT32 Expm1_HA_table[(1<<7)][1]; + __declspec(align(16)) VUINT32 poly_coeff[4][4][1]; + __declspec(align(16)) VUINT32 Log2e[4][1]; + __declspec(align(16)) VUINT32 L2H[4][1]; + __declspec(align(16)) VUINT32 L2L[4][1]; + __declspec(align(16)) VUINT32 ExpAddConst[4][1]; + __declspec(align(16)) VUINT32 IndexMask[4][1]; + __declspec(align(16)) VUINT32 ExpMask[4][1]; + __declspec(align(16)) VUINT32 MOne[4][1]; + __declspec(align(16)) VUINT32 AbsMask[4][1]; + __declspec(align(16)) VUINT32 Threshold[4][1]; + __declspec(align(16)) VUINT32 L2[4][1]; +} __svml_sexpm1_data_internal; +#endif +__svml_sexpm1_data_internal: + /* Expm1_HA_table */ + .long 0x00000000, 0x00000000 + .long 0x00016000, 0x391a3e78 + .long 0x0002d000, 0xb89e59d5 + .long 0x00044000, 0xb93ae78a + .long 0x0005b000, 0xb9279306 + .long 0x00072000, 0xb79e6961 + .long 0x0008a000, 0xb97e2fee + .long 0x000a1000, 0x391aaea9 + .long 0x000b9000, 0x39383c7d + .long 0x000d2000, 0xb9241490 + .long 0x000ea000, 0x39073169 + .long 0x00103000, 0x386e218a + .long 0x0011c000, 0x38f4dceb + .long 0x00136000, 0xb93a9a1e + .long 0x0014f000, 0x391df520 + .long 0x00169000, 0x3905a6e4 + .long 0x00183000, 0x397e0a32 + .long 0x0019e000, 0x370b2641 + .long 0x001b9000, 0xb8b1918b + .long 0x001d4000, 0xb8132c6a + .long 0x001ef000, 0x39264c12 + .long 0x0020b000, 0x37221f73 + .long 0x00227000, 0x37060619 + .long 0x00243000, 0x3922b5c1 + .long 0x00260000, 0xb814ab27 + .long 0x0027d000, 0xb89b12c6 + .long 0x0029a000, 0x382d5a75 + .long 0x002b8000, 0xb938c94b + .long 0x002d6000, 0xb97822b8 + .long 0x002f4000, 0xb910ea53 + .long 0x00312000, 0x38fd6075 + .long 0x00331000, 0x38620955 + .long 0x00350000, 0x391e667f + .long 0x00370000, 0xb89b8736 + .long 0x00390000, 0xb90a1714 + .long 0x003b0000, 0xb7a54ded + .long 0x003d1000, 0xb96b8c15 + .long 0x003f1000, 0x397336cf + .long 0x00413000, 0xb8eccd66 + .long 0x00434000, 0x39599b45 + .long 0x00456000, 0x3965422b + .long 0x00479000, 0xb8a2cdd5 + .long 0x0049c000, 0xb9484f32 + .long 0x004bf000, 0xb8fac043 + .long 0x004e2000, 0x391182a4 + .long 0x00506000, 0x38ccf6bc + .long 0x0052b000, 0xb97c4dc2 + .long 0x0054f000, 0x38d6aaf4 + .long 0x00574000, 0x391f995b + .long 0x0059a000, 0xb8ba8f62 + .long 0x005c0000, 0xb9090d05 + .long 0x005e6000, 0x37f4825e + .long 0x0060d000, 0xb8c844f5 + .long 0x00634000, 0xb76d1a83 + .long 0x0065c000, 0xb95f2310 + .long 0x00684000, 0xb952b5f8 + .long 0x006ac000, 0x37c6e7dd + .long 0x006d5000, 0xb7cfe126 + .long 0x006fe000, 0x3917337c + .long 0x00728000, 0x383b9e2d + .long 0x00752000, 0x392fa2a5 + .long 0x0077d000, 0x37df730b + .long 0x007a8000, 0x38ecb6dd + .long 0x007d4000, 0xb879f986 + /*== poly_coeff[4] ==*/ + .align 16 + .long 0x3e2AAABF, 0x3e2AAABF, 0x3e2AAABF, 0x3e2AAABF /* coeff3 */ + .long 0x3f00000F, 0x3f00000F, 0x3f00000F, 0x3f00000F /* coeff2 */ + /* 32 Byte Padding */ + .zero 32 + /*== Log2e ==*/ + .align 16 + .long 0x42B8AA3B, 0x42B8AA3B, 0x42B8AA3B, 0x42B8AA3B + /*== L2H ==*/ + .align 16 + .long 0x3c318000, 0x3c318000, 0x3c318000, 0x3c318000 + /*== L2L ==*/ + .align 16 + .long 0xb65e8083, 0xb65e8083, 0xb65e8083, 0xb65e8083 + /*== ExpAddConst ==*/ + .align 16 + .long 0x49f0fe00, 0x49f0fe00, 0x49f0fe00, 0x49f0fe00 + /*== IndexMask ==*/ + .align 16 + .long 0x000001f8, 0x000001f8, 0x000001f8, 0x000001f8 + /*== ExpMask ==*/ + .align 16 + .long 0x0001fe00, 0x0001fe00, 0x0001fe00, 0x0001fe00 + /*== MOne ==*/ + .align 16 + .long 0xbf800000, 0xbf800000, 0xbf800000, 0xbf800000 + /*== AbsMask ==*/ + .align 16 + .long 0x7fffffff, 0x7fffffff, 0x7fffffff, 0x7fffffff + /*== Threshold ==*/ + .align 16 + .long 0x42AD496B, 0x42AD496B, 0x42AD496B, 0x42AD496B // 86.643394 + /*== L2 ==*/ + .align 16 + .long 0x3cb17218, 0x3cb17218, 0x3cb17218, 0x3cb17218 + .align 16 + .type __svml_sexpm1_data_internal,@object + .size __svml_sexpm1_data_internal,.-__svml_sexpm1_data_internal + .align 16 + +.FLT_10: + .long 0x4b400000,0x4b400000,0x4b400000,0x4b400000 + .type .FLT_10,@object + .size .FLT_10,16 diff --git a/sysdeps/x86_64/fpu/multiarch/svml_s_expm1f8_core-sse.S b/sysdeps/x86_64/fpu/multiarch/svml_s_expm1f8_core-sse.S new file mode 100644 index 0000000..e34e4eb --- /dev/null +++ b/sysdeps/x86_64/fpu/multiarch/svml_s_expm1f8_core-sse.S @@ -0,0 +1,20 @@ +/* SSE version of vectorized expm1f, vector length is 8. + Copyright (C) 2021 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + <https://www.gnu.org/licenses/>. */ + +#define _ZGVdN8v_expm1f _ZGVdN8v_expm1f_sse_wrapper +#include "../svml_s_expm1f8_core.S" diff --git a/sysdeps/x86_64/fpu/multiarch/svml_s_expm1f8_core.c b/sysdeps/x86_64/fpu/multiarch/svml_s_expm1f8_core.c new file mode 100644 index 0000000..7e8b57d --- /dev/null +++ b/sysdeps/x86_64/fpu/multiarch/svml_s_expm1f8_core.c @@ -0,0 +1,28 @@ +/* Multiple versions of vectorized expm1f, vector length is 8. + Copyright (C) 2021 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + <https://www.gnu.org/licenses/>. */ + +#define SYMBOL_NAME _ZGVdN8v_expm1f +#include "ifunc-mathvec-avx2.h" + +libc_ifunc_redirected (REDIRECT_NAME, SYMBOL_NAME, IFUNC_SELECTOR ()); + +#ifdef SHARED +__hidden_ver1 (_ZGVdN8v_expm1f, __GI__ZGVdN8v_expm1f, + __redirect__ZGVdN8v_expm1f) + __attribute__ ((visibility ("hidden"))); +#endif diff --git a/sysdeps/x86_64/fpu/multiarch/svml_s_expm1f8_core_avx2.S b/sysdeps/x86_64/fpu/multiarch/svml_s_expm1f8_core_avx2.S new file mode 100644 index 0000000..8e65d69 --- /dev/null +++ b/sysdeps/x86_64/fpu/multiarch/svml_s_expm1f8_core_avx2.S @@ -0,0 +1,351 @@ +/* Function expm1f vectorized with AVX2. + Copyright (C) 2021 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + https://www.gnu.org/licenses/. */ + +/* + * ALGORITHM DESCRIPTION: + * + * N = (int)(x*2^k/log(2.0)), R = x - N*log(2)/2^k + * exp(x) = 2^(N/2^k) * poly(R) is computed in high-low parts + * expm1(x) = exp(x)-1 is then obtained via multi-precision computation + * + * + */ + +/* Offsets for data table __svml_sexpm1_data_internal + */ +#define Expm1_HA_table 0 +#define poly_coeff 512 +#define Log2e 640 +#define L2H 672 +#define L2L 704 +#define ExpAddConst 736 +#define IndexMask 768 +#define ExpMask 800 +#define MOne 832 +#define AbsMask 864 +#define Threshold 896 +#define L2 928 + +#include <sysdep.h> + + .text + .section .text.avx2,"ax",@progbits +ENTRY(_ZGVdN8v_expm1f_avx2) + pushq %rbp + cfi_def_cfa_offset(16) + movq %rsp, %rbp + cfi_def_cfa(6, 16) + cfi_offset(6, -16) + andq $-32, %rsp + subq $96, %rsp + lea __svml_sexpm1_data_internal(%rip), %rax + vmovaps %ymm0, %ymm3 + vmulps Log2e+__svml_sexpm1_data_internal(%rip), %ymm3, %ymm4 + +/* argument reduction */ + vmovups L2H+__svml_sexpm1_data_internal(%rip), %ymm2 + vmovups AbsMask+__svml_sexpm1_data_internal(%rip), %ymm5 + vroundps $0, %ymm4, %ymm8 + vaddps ExpAddConst+__svml_sexpm1_data_internal(%rip), %ymm8, %ymm0 + vfnmadd213ps %ymm3, %ymm8, %ymm2 + +/* table lookup */ + vandps IndexMask+__svml_sexpm1_data_internal(%rip), %ymm0, %ymm9 + vandps %ymm5, %ymm3, %ymm6 + vcmpnle_uqps Threshold+__svml_sexpm1_data_internal(%rip), %ymm6, %ymm7 + vfnmadd231ps L2L+__svml_sexpm1_data_internal(%rip), %ymm8, %ymm2 + vandps ExpMask+__svml_sexpm1_data_internal(%rip), %ymm0, %ymm0 + vandnps %ymm3, %ymm5, %ymm1 + vpslld $14, %ymm0, %ymm0 + vmovmskps %ymm7, %edx + vmovd %xmm9, %ecx + vextractf128 $1, %ymm9, %xmm10 + movslq %ecx, %rcx + vmovd %xmm10, %r9d + vpextrd $1, %xmm9, %esi + vpextrd $2, %xmm9, %edi + vpextrd $3, %xmm9, %r8d + vmovq (%rax,%rcx), %xmm11 + vpextrd $1, %xmm10, %r10d + vpextrd $2, %xmm10, %r11d + vpextrd $3, %xmm10, %ecx + movslq %esi, %rsi + movslq %edi, %rdi + movslq %r8d, %r8 + movslq %r9d, %r9 + movslq %r10d, %r10 + movslq %r11d, %r11 + movslq %ecx, %rcx + vmovq (%rax,%rsi), %xmm13 + vmovq (%rax,%rdi), %xmm12 + vmovq (%rax,%r8), %xmm14 + vmovq (%rax,%r9), %xmm15 + vmovq (%rax,%r10), %xmm5 + vmovq (%rax,%r11), %xmm4 + vmovq (%rax,%rcx), %xmm6 + vunpcklps %xmm12, %xmm11, %xmm7 + vunpcklps %xmm14, %xmm13, %xmm8 + vunpcklps %xmm4, %xmm15, %xmm15 + vunpcklps %xmm6, %xmm5, %xmm9 + vmulps %ymm2, %ymm2, %ymm13 + vinsertf128 $1, %xmm15, %ymm7, %ymm10 + vinsertf128 $1, %xmm9, %ymm8, %ymm11 + vunpcklps %ymm11, %ymm10, %ymm12 + vorps %ymm0, %ymm12, %ymm14 + +/* polynomial */ + vmovups poly_coeff+__svml_sexpm1_data_internal(%rip), %ymm12 + vfmadd213ps poly_coeff+32+__svml_sexpm1_data_internal(%rip), %ymm2, %ymm12 + vfmadd213ps %ymm2, %ymm13, %ymm12 + +/* T-1 */ + vmovups MOne+__svml_sexpm1_data_internal(%rip), %ymm13 + vaddps %ymm13, %ymm14, %ymm2 + vunpckhps %ymm11, %ymm10, %ymm4 + vfmadd213ps %ymm2, %ymm0, %ymm4 + vsubps %ymm13, %ymm4, %ymm0 + vfmadd213ps %ymm4, %ymm12, %ymm0 + vorps %ymm1, %ymm0, %ymm0 + testl %edx, %edx + +/* Go to special inputs processing branch */ + jne L(SPECIAL_VALUES_BRANCH) + # LOE rbx r12 r13 r14 r15 edx ymm0 ymm3 + +/* Restore registers + * and exit the function + */ + +L(EXIT): + movq %rbp, %rsp + popq %rbp + cfi_def_cfa(7, 8) + cfi_restore(6) + ret + cfi_def_cfa(6, 16) + cfi_offset(6, -16) + +/* Branch to process + * special inputs + */ + +L(SPECIAL_VALUES_BRANCH): + vmovups %ymm3, 32(%rsp) + vmovups %ymm0, 64(%rsp) + # LOE rbx r12 r13 r14 r15 edx ymm0 + + xorl %eax, %eax + # LOE rbx r12 r13 r14 r15 eax edx + + vzeroupper + movq %r12, 16(%rsp) + /* DW_CFA_expression: r12 (r12) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -80; DW_OP_plus) */ + .cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xb0, 0xff, 0xff, 0xff, 0x22 + movl %eax, %r12d + movq %r13, 8(%rsp) + /* DW_CFA_expression: r13 (r13) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -88; DW_OP_plus) */ + .cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xa8, 0xff, 0xff, 0xff, 0x22 + movl %edx, %r13d + movq %r14, (%rsp) + /* DW_CFA_expression: r14 (r14) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -96; DW_OP_plus) */ + .cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xa0, 0xff, 0xff, 0xff, 0x22 + # LOE rbx r15 r12d r13d + +/* Range mask + * bits check + */ + +L(RANGEMASK_CHECK): + btl %r12d, %r13d + +/* Call scalar math function */ + jc L(SCALAR_MATH_CALL) + # LOE rbx r15 r12d r13d + +/* Special inputs + * processing loop + */ + +L(SPECIAL_VALUES_LOOP): + incl %r12d + cmpl $8, %r12d + +/* Check bits in range mask */ + jl L(RANGEMASK_CHECK) + # LOE rbx r15 r12d r13d + + movq 16(%rsp), %r12 + cfi_restore(12) + movq 8(%rsp), %r13 + cfi_restore(13) + movq (%rsp), %r14 + cfi_restore(14) + vmovups 64(%rsp), %ymm0 + +/* Go to exit */ + jmp L(EXIT) + /* DW_CFA_expression: r12 (r12) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -80; DW_OP_plus) */ + .cfi_escape 0x10, 0x0c, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xb0, 0xff, 0xff, 0xff, 0x22 + /* DW_CFA_expression: r13 (r13) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -88; DW_OP_plus) */ + .cfi_escape 0x10, 0x0d, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xa8, 0xff, 0xff, 0xff, 0x22 + /* DW_CFA_expression: r14 (r14) (DW_OP_lit8; DW_OP_minus; DW_OP_const4s: -32; DW_OP_and; DW_OP_const4s: -96; DW_OP_plus) */ + .cfi_escape 0x10, 0x0e, 0x0e, 0x38, 0x1c, 0x0d, 0xe0, 0xff, 0xff, 0xff, 0x1a, 0x0d, 0xa0, 0xff, 0xff, 0xff, 0x22 + # LOE rbx r12 r13 r14 r15 ymm0 + +/* Scalar math fucntion call + * to process special input + */ + +L(SCALAR_MATH_CALL): + movl %r12d, %r14d + movss 32(%rsp,%r14,4), %xmm0 + call expm1f@PLT + # LOE rbx r14 r15 r12d r13d xmm0 + + movss %xmm0, 64(%rsp,%r14,4) + +/* Process special inputs in loop */ + jmp L(SPECIAL_VALUES_LOOP) + # LOE rbx r15 r12d r13d +END(_ZGVdN8v_expm1f_avx2) + + .section .rodata, "a" + .align 32 + +#ifdef __svml_sexpm1_data_internal_typedef +typedef unsigned int VUINT32; +typedef struct { + __declspec(align(32)) VUINT32 Expm1_HA_table[(1<<7)][1]; + __declspec(align(32)) VUINT32 poly_coeff[4][8][1]; + __declspec(align(32)) VUINT32 Log2e[8][1]; + __declspec(align(32)) VUINT32 L2H[8][1]; + __declspec(align(32)) VUINT32 L2L[8][1]; + __declspec(align(32)) VUINT32 ExpAddConst[8][1]; + __declspec(align(32)) VUINT32 IndexMask[8][1]; + __declspec(align(32)) VUINT32 ExpMask[8][1]; + __declspec(align(32)) VUINT32 MOne[8][1]; + __declspec(align(32)) VUINT32 AbsMask[8][1]; + __declspec(align(32)) VUINT32 Threshold[8][1]; + __declspec(align(32)) VUINT32 L2[8][1]; +} __svml_sexpm1_data_internal; +#endif +__svml_sexpm1_data_internal: + /* Expm1_HA_table */ + .long 0x00000000, 0x00000000 + .long 0x00016000, 0x391a3e78 + .long 0x0002d000, 0xb89e59d5 + .long 0x00044000, 0xb93ae78a + .long 0x0005b000, 0xb9279306 + .long 0x00072000, 0xb79e6961 + .long 0x0008a000, 0xb97e2fee + .long 0x000a1000, 0x391aaea9 + .long 0x000b9000, 0x39383c7d + .long 0x000d2000, 0xb9241490 + .long 0x000ea000, 0x39073169 + .long 0x00103000, 0x386e218a + .long 0x0011c000, 0x38f4dceb + .long 0x00136000, 0xb93a9a1e + .long 0x0014f000, 0x391df520 + .long 0x00169000, 0x3905a6e4 + .long 0x00183000, 0x397e0a32 + .long 0x0019e000, 0x370b2641 + .long 0x001b9000, 0xb8b1918b + .long 0x001d4000, 0xb8132c6a + .long 0x001ef000, 0x39264c12 + .long 0x0020b000, 0x37221f73 + .long 0x00227000, 0x37060619 + .long 0x00243000, 0x3922b5c1 + .long 0x00260000, 0xb814ab27 + .long 0x0027d000, 0xb89b12c6 + .long 0x0029a000, 0x382d5a75 + .long 0x002b8000, 0xb938c94b + .long 0x002d6000, 0xb97822b8 + .long 0x002f4000, 0xb910ea53 + .long 0x00312000, 0x38fd6075 + .long 0x00331000, 0x38620955 + .long 0x00350000, 0x391e667f + .long 0x00370000, 0xb89b8736 + .long 0x00390000, 0xb90a1714 + .long 0x003b0000, 0xb7a54ded + .long 0x003d1000, 0xb96b8c15 + .long 0x003f1000, 0x397336cf + .long 0x00413000, 0xb8eccd66 + .long 0x00434000, 0x39599b45 + .long 0x00456000, 0x3965422b + .long 0x00479000, 0xb8a2cdd5 + .long 0x0049c000, 0xb9484f32 + .long 0x004bf000, 0xb8fac043 + .long 0x004e2000, 0x391182a4 + .long 0x00506000, 0x38ccf6bc + .long 0x0052b000, 0xb97c4dc2 + .long 0x0054f000, 0x38d6aaf4 + .long 0x00574000, 0x391f995b + .long 0x0059a000, 0xb8ba8f62 + .long 0x005c0000, 0xb9090d05 + .long 0x005e6000, 0x37f4825e + .long 0x0060d000, 0xb8c844f5 + .long 0x00634000, 0xb76d1a83 + .long 0x0065c000, 0xb95f2310 + .long 0x00684000, 0xb952b5f8 + .long 0x006ac000, 0x37c6e7dd + .long 0x006d5000, 0xb7cfe126 + .long 0x006fe000, 0x3917337c + .long 0x00728000, 0x383b9e2d + .long 0x00752000, 0x392fa2a5 + .long 0x0077d000, 0x37df730b + .long 0x007a8000, 0x38ecb6dd + .long 0x007d4000, 0xb879f986 + /*== poly_coeff[4] ==*/ + .align 32 + .long 0x3e2AAABF, 0x3e2AAABF, 0x3e2AAABF, 0x3e2AAABF, 0x3e2AAABF, 0x3e2AAABF, 0x3e2AAABF, 0x3e2AAABF /* coeff3 */ + .long 0x3f00000F, 0x3f00000F, 0x3f00000F, 0x3f00000F, 0x3f00000F, 0x3f00000F, 0x3f00000F, 0x3f00000F /* coeff2 */ + /* 64 Byte Padding */ + .zero 64 + /*== Log2e ==*/ + .align 32 + .long 0x42B8AA3B, 0x42B8AA3B, 0x42B8AA3B, 0x42B8AA3B, 0x42B8AA3B, 0x42B8AA3B, 0x42B8AA3B, 0x42B8AA3B + /*== L2H ==*/ + .align 32 + .long 0x3c318000, 0x3c318000, 0x3c318000, 0x3c318000, 0x3c318000, 0x3c318000, 0x3c318000, 0x3c318000 + /*== L2L ==*/ + .align 32 + .long 0xb65e8083, 0xb65e8083, 0xb65e8083, 0xb65e8083, 0xb65e8083, 0xb65e8083, 0xb65e8083, 0xb65e8083 + /*== ExpAddConst ==*/ + .align 32 + .long 0x49f0fe00, 0x49f0fe00, 0x49f0fe00, 0x49f0fe00, 0x49f0fe00, 0x49f0fe00, 0x49f0fe00, 0x49f0fe00 + /*== IndexMask ==*/ + .align 32 + .long 0x000001f8, 0x000001f8, 0x000001f8, 0x000001f8, 0x000001f8, 0x000001f8, 0x000001f8, 0x000001f8 + /*== ExpMask ==*/ + .align 32 + .long 0x0001fe00, 0x0001fe00, 0x0001fe00, 0x0001fe00, 0x0001fe00, 0x0001fe00, 0x0001fe00, 0x0001fe00 + /*== MOne ==*/ + .align 32 + .long 0xbf800000, 0xbf800000, 0xbf800000, 0xbf800000, 0xbf800000, 0xbf800000, 0xbf800000, 0xbf800000 + /*== AbsMask ==*/ + .align 32 + .long 0x7fffffff, 0x7fffffff, 0x7fffffff, 0x7fffffff, 0x7fffffff, 0x7fffffff, 0x7fffffff, 0x7fffffff + /*== Threshold ==*/ + .align 32 + .long 0x42AD496B, 0x42AD496B, 0x42AD496B, 0x42AD496B, 0x42AD496B, 0x42AD496B, 0x42AD496B, 0x42AD496B // 86.643394 + /*== L2 ==*/ + .align 32 + .long 0x3cb17218, 0x3cb17218, 0x3cb17218, 0x3cb17218, 0x3cb17218, 0x3cb17218, 0x3cb17218, 0x3cb17218 + .align 32 + .type __svml_sexpm1_data_internal,@object + .size __svml_sexpm1_data_internal,.-__svml_sexpm1_data_internal diff --git a/sysdeps/x86_64/fpu/svml_d_expm12_core.S b/sysdeps/x86_64/fpu/svml_d_expm12_core.S new file mode 100644 index 0000000..a725d61 --- /dev/null +++ b/sysdeps/x86_64/fpu/svml_d_expm12_core.S @@ -0,0 +1,29 @@ +/* Function expm1 vectorized with SSE2. + Copyright (C) 2021 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + <https://www.gnu.org/licenses/>. */ + +#include <sysdep.h> +#include "svml_d_wrapper_impl.h" + + .text +ENTRY (_ZGVbN2v_expm1) +WRAPPER_IMPL_SSE2 expm1 +END (_ZGVbN2v_expm1) + +#ifndef USE_MULTIARCH + libmvec_hidden_def (_ZGVbN2v_expm1) +#endif diff --git a/sysdeps/x86_64/fpu/svml_d_expm14_core.S b/sysdeps/x86_64/fpu/svml_d_expm14_core.S new file mode 100644 index 0000000..1027def --- /dev/null +++ b/sysdeps/x86_64/fpu/svml_d_expm14_core.S @@ -0,0 +1,29 @@ +/* Function expm1 vectorized with AVX2, wrapper version. + Copyright (C) 2021 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + <https://www.gnu.org/licenses/>. */ + +#include <sysdep.h> +#include "svml_d_wrapper_impl.h" + + .text +ENTRY (_ZGVdN4v_expm1) +WRAPPER_IMPL_AVX _ZGVbN2v_expm1 +END (_ZGVdN4v_expm1) + +#ifndef USE_MULTIARCH + libmvec_hidden_def (_ZGVdN4v_expm1) +#endif diff --git a/sysdeps/x86_64/fpu/svml_d_expm14_core_avx.S b/sysdeps/x86_64/fpu/svml_d_expm14_core_avx.S new file mode 100644 index 0000000..3a34262 --- /dev/null +++ b/sysdeps/x86_64/fpu/svml_d_expm14_core_avx.S @@ -0,0 +1,25 @@ +/* Function expm1 vectorized in AVX ISA as wrapper to SSE4 ISA version. + Copyright (C) 2021 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + <https://www.gnu.org/licenses/>. */ + +#include <sysdep.h> +#include "svml_d_wrapper_impl.h" + + .text +ENTRY (_ZGVcN4v_expm1) +WRAPPER_IMPL_AVX _ZGVbN2v_expm1 +END (_ZGVcN4v_expm1) diff --git a/sysdeps/x86_64/fpu/svml_d_expm18_core.S b/sysdeps/x86_64/fpu/svml_d_expm18_core.S new file mode 100644 index 0000000..fa97595 --- /dev/null +++ b/sysdeps/x86_64/fpu/svml_d_expm18_core.S @@ -0,0 +1,25 @@ +/* Function expm1 vectorized with AVX-512, wrapper to AVX2. + Copyright (C) 2021 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + <https://www.gnu.org/licenses/>. */ + +#include <sysdep.h> +#include "svml_d_wrapper_impl.h" + + .text +ENTRY (_ZGVeN8v_expm1) +WRAPPER_IMPL_AVX512 _ZGVdN4v_expm1 +END (_ZGVeN8v_expm1) diff --git a/sysdeps/x86_64/fpu/svml_s_expm1f16_core.S b/sysdeps/x86_64/fpu/svml_s_expm1f16_core.S new file mode 100644 index 0000000..b742363 --- /dev/null +++ b/sysdeps/x86_64/fpu/svml_s_expm1f16_core.S @@ -0,0 +1,25 @@ +/* Function expm1f vectorized with AVX-512. Wrapper to AVX2 version. + Copyright (C) 2021 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + <https://www.gnu.org/licenses/>. */ + +#include <sysdep.h> +#include "svml_s_wrapper_impl.h" + + .text +ENTRY (_ZGVeN16v_expm1f) +WRAPPER_IMPL_AVX512 _ZGVdN8v_expm1f +END (_ZGVeN16v_expm1f) diff --git a/sysdeps/x86_64/fpu/svml_s_expm1f4_core.S b/sysdeps/x86_64/fpu/svml_s_expm1f4_core.S new file mode 100644 index 0000000..334a491 --- /dev/null +++ b/sysdeps/x86_64/fpu/svml_s_expm1f4_core.S @@ -0,0 +1,29 @@ +/* Function expm1f vectorized with SSE2, wrapper version. + Copyright (C) 2021 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + <https://www.gnu.org/licenses/>. */ + +#include <sysdep.h> +#include "svml_s_wrapper_impl.h" + + .text +ENTRY (_ZGVbN4v_expm1f) +WRAPPER_IMPL_SSE2 expm1f +END (_ZGVbN4v_expm1f) + +#ifndef USE_MULTIARCH + libmvec_hidden_def (_ZGVbN4v_expm1f) +#endif diff --git a/sysdeps/x86_64/fpu/svml_s_expm1f8_core.S b/sysdeps/x86_64/fpu/svml_s_expm1f8_core.S new file mode 100644 index 0000000..1058957 --- /dev/null +++ b/sysdeps/x86_64/fpu/svml_s_expm1f8_core.S @@ -0,0 +1,29 @@ +/* Function expm1f vectorized with AVX2, wrapper version. + Copyright (C) 2021 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + <https://www.gnu.org/licenses/>. */ + +#include <sysdep.h> +#include "svml_s_wrapper_impl.h" + + .text +ENTRY (_ZGVdN8v_expm1f) +WRAPPER_IMPL_AVX _ZGVbN4v_expm1f +END (_ZGVdN8v_expm1f) + +#ifndef USE_MULTIARCH + libmvec_hidden_def (_ZGVdN8v_expm1f) +#endif diff --git a/sysdeps/x86_64/fpu/svml_s_expm1f8_core_avx.S b/sysdeps/x86_64/fpu/svml_s_expm1f8_core_avx.S new file mode 100644 index 0000000..4161113 --- /dev/null +++ b/sysdeps/x86_64/fpu/svml_s_expm1f8_core_avx.S @@ -0,0 +1,25 @@ +/* Function expm1f vectorized in AVX ISA as wrapper to SSE4 ISA version. + Copyright (C) 2021 Free Software Foundation, Inc. + This file is part of the GNU C Library. + + The GNU C Library is free software; you can redistribute it and/or + modify it under the terms of the GNU Lesser General Public + License as published by the Free Software Foundation; either + version 2.1 of the License, or (at your option) any later version. + + The GNU C Library is distributed in the hope that it will be useful, + but WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + Lesser General Public License for more details. + + You should have received a copy of the GNU Lesser General Public + License along with the GNU C Library; if not, see + <https://www.gnu.org/licenses/>. */ + +#include <sysdep.h> +#include "svml_s_wrapper_impl.h" + + .text +ENTRY (_ZGVcN8v_expm1f) +WRAPPER_IMPL_AVX _ZGVbN4v_expm1f +END (_ZGVcN8v_expm1f) diff --git a/sysdeps/x86_64/fpu/test-double-libmvec-expm1-avx.c b/sysdeps/x86_64/fpu/test-double-libmvec-expm1-avx.c new file mode 100644 index 0000000..3e59cb7 --- /dev/null +++ b/sysdeps/x86_64/fpu/test-double-libmvec-expm1-avx.c @@ -0,0 +1 @@ +#include "test-double-libmvec-expm1.c" diff --git a/sysdeps/x86_64/fpu/test-double-libmvec-expm1-avx2.c b/sysdeps/x86_64/fpu/test-double-libmvec-expm1-avx2.c new file mode 100644 index 0000000..3e59cb7 --- /dev/null +++ b/sysdeps/x86_64/fpu/test-double-libmvec-expm1-avx2.c @@ -0,0 +1 @@ +#include "test-double-libmvec-expm1.c" diff --git a/sysdeps/x86_64/fpu/test-double-libmvec-expm1-avx512f.c b/sysdeps/x86_64/fpu/test-double-libmvec-expm1-avx512f.c new file mode 100644 index 0000000..3e59cb7 --- /dev/null +++ b/sysdeps/x86_64/fpu/test-double-libmvec-expm1-avx512f.c @@ -0,0 +1 @@ +#include "test-double-libmvec-expm1.c" diff --git a/sysdeps/x86_64/fpu/test-double-libmvec-expm1.c b/sysdeps/x86_64/fpu/test-double-libmvec-expm1.c new file mode 100644 index 0000000..33806a7 --- /dev/null +++ b/sysdeps/x86_64/fpu/test-double-libmvec-expm1.c @@ -0,0 +1,3 @@ +#define LIBMVEC_TYPE double +#define LIBMVEC_FUNC expm1 +#include "test-vector-abi-arg1.h" diff --git a/sysdeps/x86_64/fpu/test-double-vlen2-wrappers.c b/sysdeps/x86_64/fpu/test-double-vlen2-wrappers.c index 68c449e..0222f9f 100644 --- a/sysdeps/x86_64/fpu/test-double-vlen2-wrappers.c +++ b/sysdeps/x86_64/fpu/test-double-vlen2-wrappers.c @@ -34,6 +34,7 @@ VECTOR_WRAPPER_ff (WRAPPER_NAME (hypot), _ZGVbN2vv_hypot) VECTOR_WRAPPER (WRAPPER_NAME (exp2), _ZGVbN2v_exp2) VECTOR_WRAPPER (WRAPPER_NAME (exp10), _ZGVbN2v_exp10) VECTOR_WRAPPER (WRAPPER_NAME (cosh), _ZGVbN2v_cosh) +VECTOR_WRAPPER (WRAPPER_NAME (expm1), _ZGVbN2v_expm1) #define VEC_INT_TYPE __m128i diff --git a/sysdeps/x86_64/fpu/test-double-vlen4-avx2-wrappers.c b/sysdeps/x86_64/fpu/test-double-vlen4-avx2-wrappers.c index df67306..1aad9fa 100644 --- a/sysdeps/x86_64/fpu/test-double-vlen4-avx2-wrappers.c +++ b/sysdeps/x86_64/fpu/test-double-vlen4-avx2-wrappers.c @@ -37,6 +37,7 @@ VECTOR_WRAPPER_ff (WRAPPER_NAME (hypot), _ZGVdN4vv_hypot) VECTOR_WRAPPER (WRAPPER_NAME (exp2), _ZGVdN4v_exp2) VECTOR_WRAPPER (WRAPPER_NAME (exp10), _ZGVdN4v_exp10) VECTOR_WRAPPER (WRAPPER_NAME (cosh), _ZGVdN4v_cosh) +VECTOR_WRAPPER (WRAPPER_NAME (expm1), _ZGVdN4v_expm1) #ifndef __ILP32__ # define VEC_INT_TYPE __m256i diff --git a/sysdeps/x86_64/fpu/test-double-vlen4-wrappers.c b/sysdeps/x86_64/fpu/test-double-vlen4-wrappers.c index 1a67310..e404bf8 100644 --- a/sysdeps/x86_64/fpu/test-double-vlen4-wrappers.c +++ b/sysdeps/x86_64/fpu/test-double-vlen4-wrappers.c @@ -34,6 +34,7 @@ VECTOR_WRAPPER_ff (WRAPPER_NAME (hypot), _ZGVcN4vv_hypot) VECTOR_WRAPPER (WRAPPER_NAME (exp2), _ZGVcN4v_exp2) VECTOR_WRAPPER (WRAPPER_NAME (exp10), _ZGVcN4v_exp10) VECTOR_WRAPPER (WRAPPER_NAME (cosh), _ZGVcN4v_cosh) +VECTOR_WRAPPER (WRAPPER_NAME (expm1), _ZGVcN4v_expm1) #define VEC_INT_TYPE __m128i diff --git a/sysdeps/x86_64/fpu/test-double-vlen8-wrappers.c b/sysdeps/x86_64/fpu/test-double-vlen8-wrappers.c index 4cdfa91..2b4de59 100644 --- a/sysdeps/x86_64/fpu/test-double-vlen8-wrappers.c +++ b/sysdeps/x86_64/fpu/test-double-vlen8-wrappers.c @@ -34,6 +34,7 @@ VECTOR_WRAPPER_ff (WRAPPER_NAME (hypot), _ZGVeN8vv_hypot) VECTOR_WRAPPER (WRAPPER_NAME (exp2), _ZGVeN8v_exp2) VECTOR_WRAPPER (WRAPPER_NAME (exp10), _ZGVeN8v_exp10) VECTOR_WRAPPER (WRAPPER_NAME (cosh), _ZGVeN8v_cosh) +VECTOR_WRAPPER (WRAPPER_NAME (expm1), _ZGVeN8v_expm1) #ifndef __ILP32__ # define VEC_INT_TYPE __m512i diff --git a/sysdeps/x86_64/fpu/test-float-libmvec-expm1f-avx.c b/sysdeps/x86_64/fpu/test-float-libmvec-expm1f-avx.c new file mode 100644 index 0000000..67e31f9 --- /dev/null +++ b/sysdeps/x86_64/fpu/test-float-libmvec-expm1f-avx.c @@ -0,0 +1 @@ +#include "test-float-libmvec-expm1f.c" diff --git a/sysdeps/x86_64/fpu/test-float-libmvec-expm1f-avx2.c b/sysdeps/x86_64/fpu/test-float-libmvec-expm1f-avx2.c new file mode 100644 index 0000000..67e31f9 --- /dev/null +++ b/sysdeps/x86_64/fpu/test-float-libmvec-expm1f-avx2.c @@ -0,0 +1 @@ +#include "test-float-libmvec-expm1f.c" diff --git a/sysdeps/x86_64/fpu/test-float-libmvec-expm1f-avx512f.c b/sysdeps/x86_64/fpu/test-float-libmvec-expm1f-avx512f.c new file mode 100644 index 0000000..67e31f9 --- /dev/null +++ b/sysdeps/x86_64/fpu/test-float-libmvec-expm1f-avx512f.c @@ -0,0 +1 @@ +#include "test-float-libmvec-expm1f.c" diff --git a/sysdeps/x86_64/fpu/test-float-libmvec-expm1f.c b/sysdeps/x86_64/fpu/test-float-libmvec-expm1f.c new file mode 100644 index 0000000..aa9871a --- /dev/null +++ b/sysdeps/x86_64/fpu/test-float-libmvec-expm1f.c @@ -0,0 +1,3 @@ +#define LIBMVEC_TYPE float +#define LIBMVEC_FUNC expm1f +#include "test-vector-abi-arg1.h" diff --git a/sysdeps/x86_64/fpu/test-float-vlen16-wrappers.c b/sysdeps/x86_64/fpu/test-float-vlen16-wrappers.c index 47a9862..9a4a1b8 100644 --- a/sysdeps/x86_64/fpu/test-float-vlen16-wrappers.c +++ b/sysdeps/x86_64/fpu/test-float-vlen16-wrappers.c @@ -34,6 +34,7 @@ VECTOR_WRAPPER_ff (WRAPPER_NAME (hypotf), _ZGVeN16vv_hypotf) VECTOR_WRAPPER (WRAPPER_NAME (exp2f), _ZGVeN16v_exp2f) VECTOR_WRAPPER (WRAPPER_NAME (exp10f), _ZGVeN16v_exp10f) VECTOR_WRAPPER (WRAPPER_NAME (coshf), _ZGVeN16v_coshf) +VECTOR_WRAPPER (WRAPPER_NAME (expm1f), _ZGVeN16v_expm1f) #define VEC_INT_TYPE __m512i diff --git a/sysdeps/x86_64/fpu/test-float-vlen4-wrappers.c b/sysdeps/x86_64/fpu/test-float-vlen4-wrappers.c index e7c5410..eb4e36d 100644 --- a/sysdeps/x86_64/fpu/test-float-vlen4-wrappers.c +++ b/sysdeps/x86_64/fpu/test-float-vlen4-wrappers.c @@ -34,6 +34,7 @@ VECTOR_WRAPPER_ff (WRAPPER_NAME (hypotf), _ZGVbN4vv_hypotf) VECTOR_WRAPPER (WRAPPER_NAME (exp2f), _ZGVbN4v_exp2f) VECTOR_WRAPPER (WRAPPER_NAME (exp10f), _ZGVbN4v_exp10f) VECTOR_WRAPPER (WRAPPER_NAME (coshf), _ZGVbN4v_coshf) +VECTOR_WRAPPER (WRAPPER_NAME (expm1f), _ZGVbN4v_expm1f) #define VEC_INT_TYPE __m128i diff --git a/sysdeps/x86_64/fpu/test-float-vlen8-avx2-wrappers.c b/sysdeps/x86_64/fpu/test-float-vlen8-avx2-wrappers.c index b8e9d48..d8adab5 100644 --- a/sysdeps/x86_64/fpu/test-float-vlen8-avx2-wrappers.c +++ b/sysdeps/x86_64/fpu/test-float-vlen8-avx2-wrappers.c @@ -37,6 +37,7 @@ VECTOR_WRAPPER_ff (WRAPPER_NAME (hypotf), _ZGVdN8vv_hypotf) VECTOR_WRAPPER (WRAPPER_NAME (exp2f), _ZGVdN8v_exp2f) VECTOR_WRAPPER (WRAPPER_NAME (exp10f), _ZGVdN8v_exp10f) VECTOR_WRAPPER (WRAPPER_NAME (coshf), _ZGVdN8v_coshf) +VECTOR_WRAPPER (WRAPPER_NAME (expm1f), _ZGVdN8v_expm1f) /* Redefinition of wrapper to be compatible with _ZGVdN8vvv_sincosf. */ #undef VECTOR_WRAPPER_fFF diff --git a/sysdeps/x86_64/fpu/test-float-vlen8-wrappers.c b/sysdeps/x86_64/fpu/test-float-vlen8-wrappers.c index 328c827..e6e1a90 100644 --- a/sysdeps/x86_64/fpu/test-float-vlen8-wrappers.c +++ b/sysdeps/x86_64/fpu/test-float-vlen8-wrappers.c @@ -34,6 +34,7 @@ VECTOR_WRAPPER_ff (WRAPPER_NAME (hypotf), _ZGVcN8vv_hypotf) VECTOR_WRAPPER (WRAPPER_NAME (exp2f), _ZGVcN8v_exp2f) VECTOR_WRAPPER (WRAPPER_NAME (exp10f), _ZGVcN8v_exp10f) VECTOR_WRAPPER (WRAPPER_NAME (coshf), _ZGVcN8v_coshf) +VECTOR_WRAPPER (WRAPPER_NAME (expm1f), _ZGVcN8v_expm1f) #define VEC_INT_TYPE __m128i |