diff options
Diffstat (limited to 'REORG.TODO/sysdeps/x86_64/fpu/multiarch/svml_s_sincosf8_core_avx2.S')
| -rw-r--r-- | REORG.TODO/sysdeps/x86_64/fpu/multiarch/svml_s_sincosf8_core_avx2.S | 389 |
1 files changed, 389 insertions, 0 deletions
diff --git a/REORG.TODO/sysdeps/x86_64/fpu/multiarch/svml_s_sincosf8_core_avx2.S b/REORG.TODO/sysdeps/x86_64/fpu/multiarch/svml_s_sincosf8_core_avx2.S new file mode 100644 index 0000000000..9e4e2c71c5 --- /dev/null +++ b/REORG.TODO/sysdeps/x86_64/fpu/multiarch/svml_s_sincosf8_core_avx2.S @@ -0,0 +1,389 @@ +/* Function sincosf vectorized with AVX2. + Copyright (C) 2014-2017 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 + <http://www.gnu.org/licenses/>. */ + +#include <sysdep.h> +#include "svml_s_trig_data.h" + + .text +ENTRY (_ZGVdN8vl4l4_sincosf_avx2) +/* + ALGORITHM DESCRIPTION: + + 1) Range reduction to [-Pi/4; +Pi/4] interval + a) Grab sign from source argument and save it. + b) Remove sign using AND operation + c) Getting octant Y by 2/Pi multiplication + d) Add "Right Shifter" value + e) Treat obtained value as integer S for destination sign setting. + SS = ((S-S&1)&2)<<30; For sin part + SC = ((S+S&1)&2)<<30; For cos part + f) Change destination sign if source sign is negative + using XOR operation. + g) Subtract "Right Shifter" (0x4B000000) value + h) Subtract Y*(PI/2) from X argument, where PI/2 divided to 4 parts: + X = X - Y*PI1 - Y*PI2 - Y*PI3 - Y*PI4; + 2) Polynomial (minimax for sin within [-Pi/4; +Pi/4] interval) + a) Calculate X^2 = X * X + b) Calculate 2 polynomials for sin and cos: + RS = X * ( A0 + X^2 * (A1 + x^2 * (A2 + x^2 * (A3)))); + RC = B0 + X^2 * (B1 + x^2 * (B2 + x^2 * (B3 + x^2 * (B4)))); + c) Swap RS & RC if if first bit of obtained value after + Right Shifting is set to 1. Using And, Andnot & Or operations. + 3) Destination sign setting + a) Set shifted destination sign using XOR operation: + R1 = XOR( RS, SS ); + R2 = XOR( RC, SC ). */ + + pushq %rbp + cfi_adjust_cfa_offset (8) + cfi_rel_offset (%rbp, 0) + movq %rsp, %rbp + cfi_def_cfa_register (%rbp) + andq $-64, %rsp + subq $448, %rsp + movq __svml_s_trig_data@GOTPCREL(%rip), %rax + vmovdqa %ymm0, %ymm5 + vmovups %ymm13, 352(%rsp) + vmovups __sAbsMask(%rax), %ymm2 + vmovups __sInvPI(%rax), %ymm1 + vmovups __sPI1_FMA(%rax), %ymm13 + vmovups %ymm15, 288(%rsp) + +/* Absolute argument computation */ + vandps %ymm2, %ymm5, %ymm4 + +/* c) Getting octant Y by 2/Pi multiplication + d) Add "Right Shifter" value */ + vfmadd213ps __sRShifter(%rax), %ymm4, %ymm1 + +/* e) Treat obtained value as integer S for destination sign setting */ + vpslld $31, %ymm1, %ymm0 + +/* g) Subtract "Right Shifter" (0x4B000000) value */ + vsubps __sRShifter(%rax), %ymm1, %ymm1 + +/* h) Subtract Y*(PI/2) from X argument, where PI/2 divided to 3 parts: + X = X - Y*PI1 - Y*PI2 - Y*PI3 */ + vmovdqa %ymm4, %ymm7 + vfnmadd231ps %ymm1, %ymm13, %ymm7 + vfnmadd231ps __sPI2_FMA(%rax), %ymm1, %ymm7 + vandps __sSignMask(%rax), %ymm7, %ymm15 + vxorps __sOneHalf(%rax), %ymm15, %ymm6 + +/* Add correction term 0.5 for cos() part */ + vaddps %ymm6, %ymm1, %ymm6 + vmovdqa %ymm4, %ymm3 + vfnmadd231ps %ymm6, %ymm13, %ymm3 + vmovups __sPI3_FMA(%rax), %ymm13 + vcmpnle_uqps __sRangeReductionVal(%rax), %ymm4, %ymm4 + vfnmadd231ps __sPI2_FMA(%rax), %ymm6, %ymm3 + vfnmadd213ps %ymm7, %ymm13, %ymm1 + vfnmadd213ps %ymm3, %ymm13, %ymm6 + +/* Result sign calculations */ + vxorps __sSignMask(%rax), %ymm15, %ymm3 + vxorps %ymm0, %ymm3, %ymm7 + vxorps %ymm7, %ymm6, %ymm3 + vxorps %ymm0, %ymm1, %ymm15 + vandnps %ymm5, %ymm2, %ymm6 + vmovups __sA7_FMA(%rax), %ymm2 + vmulps %ymm15, %ymm15, %ymm13 + vmovups __sA9_FMA(%rax), %ymm7 + vmulps %ymm3, %ymm3, %ymm1 + +/* 2) Polynomial (minimax for sin within [-Pi/4; +Pi/4] interval) + a) Calculate X^2 = X * X + b) Calculate 2 polynomials for sin and cos: + RS = X * ( A0 + X^2 * (A1 + x^2 * (A2 + x^2 * (A3)))); + RC = B0 + X^2 * (B1 + x^2 * (B2 + x^2 * (B3 + x^2 * (B4)))) */ + vmovdqa %ymm2, %ymm0 + vfmadd231ps __sA9_FMA(%rax), %ymm13, %ymm0 + vfmadd213ps %ymm2, %ymm1, %ymm7 + vfmadd213ps __sA5_FMA(%rax), %ymm13, %ymm0 + vfmadd213ps __sA5_FMA(%rax), %ymm1, %ymm7 + vfmadd213ps __sA3(%rax), %ymm13, %ymm0 + vfmadd213ps __sA3(%rax), %ymm1, %ymm7 + vmulps %ymm13, %ymm0, %ymm13 + vmulps %ymm1, %ymm7, %ymm1 + vfmadd213ps %ymm15, %ymm15, %ymm13 + vfmadd213ps %ymm3, %ymm3, %ymm1 + vmovmskps %ymm4, %ecx + vxorps %ymm6, %ymm13, %ymm0 + testl %ecx, %ecx + jne .LBL_1_3 + +.LBL_1_2: + cfi_remember_state + vmovups 352(%rsp), %ymm13 + vmovups 288(%rsp), %ymm15 + vmovups %ymm0, (%rdi) + vmovups %ymm1, (%rsi) + movq %rbp, %rsp + cfi_def_cfa_register (%rsp) + popq %rbp + cfi_adjust_cfa_offset (-8) + cfi_restore (%rbp) + ret + +.LBL_1_3: + cfi_restore_state + vmovups %ymm5, 256(%rsp) + vmovups %ymm0, 320(%rsp) + vmovups %ymm1, 384(%rsp) + je .LBL_1_2 + + xorb %dl, %dl + xorl %eax, %eax + vmovups %ymm8, 160(%rsp) + vmovups %ymm9, 128(%rsp) + vmovups %ymm10, 96(%rsp) + vmovups %ymm11, 64(%rsp) + vmovups %ymm12, 32(%rsp) + vmovups %ymm14, (%rsp) + movq %rsi, 192(%rsp) + movq %r12, 232(%rsp) + cfi_offset_rel_rsp (12, 232) + movb %dl, %r12b + movq %r13, 224(%rsp) + cfi_offset_rel_rsp (13, 224) + movl %eax, %r13d + movq %r14, 216(%rsp) + cfi_offset_rel_rsp (14, 216) + movl %ecx, %r14d + movq %r15, 208(%rsp) + cfi_offset_rel_rsp (14, 208) + movq %rbx, 200(%rsp) + movq %rdi, %rbx + cfi_remember_state + +.LBL_1_6: + btl %r13d, %r14d + jc .LBL_1_13 + +.LBL_1_7: + lea 1(%r13), %esi + btl %esi, %r14d + jc .LBL_1_10 + +.LBL_1_8: + incb %r12b + addl $2, %r13d + cmpb $16, %r12b + jb .LBL_1_6 + + vmovups 160(%rsp), %ymm8 + movq %rbx, %rdi + vmovups 128(%rsp), %ymm9 + vmovups 96(%rsp), %ymm10 + vmovups 64(%rsp), %ymm11 + vmovups 32(%rsp), %ymm12 + vmovups (%rsp), %ymm14 + vmovups 320(%rsp), %ymm0 + vmovups 384(%rsp), %ymm1 + movq 192(%rsp), %rsi + movq 232(%rsp), %r12 + cfi_restore (%r12) + movq 224(%rsp), %r13 + cfi_restore (%r13) + movq 216(%rsp), %r14 + cfi_restore (%r14) + movq 208(%rsp), %r15 + cfi_restore (%r15) + movq 200(%rsp), %rbx + jmp .LBL_1_2 + +.LBL_1_10: + cfi_restore_state + movzbl %r12b, %r15d + vmovss 260(%rsp,%r15,8), %xmm0 + vzeroupper + + call JUMPTARGET(sinf) + + vmovss %xmm0, 324(%rsp,%r15,8) + vmovss 260(%rsp,%r15,8), %xmm0 + + call JUMPTARGET(cosf) + + vmovss %xmm0, 388(%rsp,%r15,8) + jmp .LBL_1_8 + +.LBL_1_13: + movzbl %r12b, %r15d + vmovss 256(%rsp,%r15,8), %xmm0 + vzeroupper + + call JUMPTARGET(sinf) + + vmovss %xmm0, 320(%rsp,%r15,8) + vmovss 256(%rsp,%r15,8), %xmm0 + + call JUMPTARGET(cosf) + + vmovss %xmm0, 384(%rsp,%r15,8) + jmp .LBL_1_7 + +END (_ZGVdN8vl4l4_sincosf_avx2) +libmvec_hidden_def(_ZGVdN8vl4l4_sincosf_avx2) + +/* vvv version implemented with wrapper to vl4l4 variant. */ +ENTRY (_ZGVdN8vvv_sincosf_avx2) +#ifndef __ILP32__ + pushq %rbp + cfi_adjust_cfa_offset (8) + cfi_rel_offset (%rbp, 0) + movq %rsp, %rbp + cfi_def_cfa_register (%rbp) + andq $-32, %rsp + subq $192, %rsp + vmovdqu %ymm1, 64(%rsp) + lea (%rsp), %rdi + vmovdqu %ymm2, 96(%rdi) + vmovdqu %ymm3, 128(%rdi) + vmovdqu %ymm4, 160(%rdi) + lea 32(%rsp), %rsi + call HIDDEN_JUMPTARGET(_ZGVdN8vl4l4_sincosf_avx2) + movq 64(%rsp), %rdx + movq 72(%rsp), %rsi + movq 80(%rsp), %r8 + movq 88(%rsp), %r10 + movl (%rsp), %eax + movl 4(%rsp), %ecx + movl 8(%rsp), %edi + movl 12(%rsp), %r9d + movl %eax, (%rdx) + movl %ecx, (%rsi) + movq 96(%rsp), %rax + movq 104(%rsp), %rcx + movl %edi, (%r8) + movl %r9d, (%r10) + movq 112(%rsp), %rdi + movq 120(%rsp), %r9 + movl 16(%rsp), %r11d + movl 20(%rsp), %edx + movl 24(%rsp), %esi + movl 28(%rsp), %r8d + movl %r11d, (%rax) + movl %edx, (%rcx) + movq 128(%rsp), %r11 + movq 136(%rsp), %rdx + movl %esi, (%rdi) + movl %r8d, (%r9) + movq 144(%rsp), %rsi + movq 152(%rsp), %r8 + movl 32(%rsp), %r10d + movl 36(%rsp), %eax + movl 40(%rsp), %ecx + movl 44(%rsp), %edi + movl %r10d, (%r11) + movl %eax, (%rdx) + movq 160(%rsp), %r10 + movq 168(%rsp), %rax + movl %ecx, (%rsi) + movl %edi, (%r8) + movq 176(%rsp), %rcx + movq 184(%rsp), %rdi + movl 48(%rsp), %r9d + movl 52(%rsp), %r11d + movl 56(%rsp), %edx + movl 60(%rsp), %esi + movl %r9d, (%r10) + movl %r11d, (%rax) + movl %edx, (%rcx) + movl %esi, (%rdi) + movq %rbp, %rsp + cfi_def_cfa_register (%rsp) + popq %rbp + cfi_adjust_cfa_offset (-8) + cfi_restore (%rbp) + ret +#else + leal 8(%rsp), %r10d + .cfi_def_cfa 10, 0 + andl $-32, %esp + pushq -8(%r10d) + pushq %rbp + .cfi_escape 0x10,0x6,0x2,0x76,0 + movl %esp, %ebp + pushq %r10 + .cfi_escape 0xf,0x3,0x76,0x78,0x6 + leal -48(%rbp), %esi + leal -80(%rbp), %edi + subl $136, %esp + vmovdqa %ymm1, -112(%ebp) + vmovdqa %ymm2, -144(%ebp) + call HIDDEN_JUMPTARGET(_ZGVdN8vl4l4_sincosf_avx2) + vmovdqa -112(%ebp), %xmm0 + vmovq %xmm0, %rax + vmovss -80(%ebp), %xmm0 + vmovss %xmm0, (%eax) + vmovss -76(%ebp), %xmm0 + shrq $32, %rax + vmovss %xmm0, (%eax) + movq -104(%ebp), %rax + vmovss -72(%ebp), %xmm0 + vmovss %xmm0, (%eax) + vmovss -68(%ebp), %xmm0 + shrq $32, %rax + vmovss %xmm0, (%eax) + movq -96(%ebp), %rax + vmovss -64(%ebp), %xmm0 + vmovss %xmm0, (%eax) + vmovss -60(%ebp), %xmm0 + shrq $32, %rax + vmovss %xmm0, (%eax) + movq -88(%ebp), %rax + vmovss -56(%ebp), %xmm0 + vmovss %xmm0, (%eax) + vmovss -52(%ebp), %xmm0 + shrq $32, %rax + vmovss %xmm0, (%eax) + vmovdqa -144(%ebp), %xmm0 + vmovq %xmm0, %rax + vmovss -48(%ebp), %xmm0 + vmovss %xmm0, (%eax) + vmovss -44(%ebp), %xmm0 + shrq $32, %rax + vmovss %xmm0, (%eax) + movq -136(%ebp), %rax + vmovss -40(%ebp), %xmm0 + vmovss %xmm0, (%eax) + vmovss -36(%ebp), %xmm0 + shrq $32, %rax + vmovss %xmm0, (%eax) + movq -128(%ebp), %rax + vmovss -32(%ebp), %xmm0 + vmovss %xmm0, (%eax) + vmovss -28(%ebp), %xmm0 + shrq $32, %rax + vmovss %xmm0, (%eax) + movq -120(%ebp), %rax + vmovss -24(%ebp), %xmm0 + vmovss %xmm0, (%eax) + vmovss -20(%ebp), %xmm0 + shrq $32, %rax + vmovss %xmm0, (%eax) + addl $136, %esp + popq %r10 + .cfi_def_cfa 10, 0 + popq %rbp + leal -8(%r10), %esp + .cfi_def_cfa 7, 8 + ret +#endif +END (_ZGVdN8vvv_sincosf_avx2) |