📄 s_asinhl.s
字号:
FR_Q3 = f64 FR_Q2 = f65 FR_1LN10_hi = f66 FR_Q1 = f67 FR_1LN10_lo = f68 FR_P5 = f69 FR_rcub = f70 FR_Neg_One = f71 FR_Z = f72 FR_AA = f73 FR_BB = f74 FR_S_lo = f75 FR_2_to_minus_N = f76 // Huge & Main path prolog registersFR_Half = f77FR_Two = f78FR_X2 = f79FR_P2 = f80FR_P2L = f81FR_Rcp = f82FR_GG = f83FR_HH = f84FR_EE = f85FR_DD = f86FR_GL = f87FR_A = f88FR_AL = f89FR_B = f90FR_BL = f91FR_Tmp = f92 // Near 0 & Huges path prolog registersFR_C3 = f93FR_C5 = f94FR_C7 = f95FR_C9 = f96FR_X3 = f97FR_X4 = f98FR_P9 = f99FR_P5 = f100FR_P3 = f101// General Purpose Registers // General prolog registersGR_PFS = r32GR_TwoN7 = r40GR_TwoP63 = r41GR_ExpMask = r42GR_ArgExp = r43GR_Half = r44 // Near 0 path prolog registersGR_Poly_C_35 = r45GR_Poly_C_79 = r46 // Special logl registersGR_Index1 = r34 GR_Index2 = r35 GR_signif = r36 GR_X_0 = r37 GR_X_1 = r38 GR_X_2 = r39 GR_Z_1 = r40 GR_Z_2 = r41 GR_N = r42 GR_Bias = r43 GR_M = r44 GR_Index3 = r45 GR_exp_2tom80 = r45 GR_exp_mask = r47 GR_exp_2tom7 = r48 GR_ad_ln10 = r49 GR_ad_tbl_1 = r50GR_ad_tbl_2 = r51GR_ad_tbl_3 = r52GR_ad_q = r53GR_ad_z_1 = r54GR_ad_z_2 = r55GR_ad_z_3 = r56GR_minus_N = r57.section .textGLOBAL_LIBM_ENTRY(asinhl){ .mfi alloc GR_PFS = ar.pfs,0,27,0,0 fma.s1 FR_P2 = FR_Arg, FR_Arg, f1 // p2 = x^2 + 1 mov GR_Half = 0xfffe // 0.5's exp}{ .mfi addl GR_Poly_C_79 = @ltoff(Poly_C_near_0_79), gp // C7, C9 coeffs fma.s1 FR_X2 = FR_Arg, FR_Arg, f0 // Obtain x^2 addl GR_Poly_C_35 = @ltoff(Poly_C_near_0_35), gp // C3, C5 coeffs};;{ .mfi getf.exp GR_ArgExp = FR_Arg // get arument's exponent fabs FR_AX = FR_Arg // absolute value of argument mov GR_TwoN7 = 0xfff8 // 2^-7 exp}{ .mfi ld8 GR_Poly_C_79 = [GR_Poly_C_79] // get actual coeff table address fma.s0 FR_Two = f1, f1, f1 // construct 2.0 mov GR_ExpMask = 0x1ffff // mask for exp};;{ .mfi ld8 GR_Poly_C_35 = [GR_Poly_C_35] // get actual coeff table address fclass.m p6,p0 = FR_Arg, 0xe7 // if arg NaN inf zero mov GR_TwoP63 = 0x1003e // 2^63 exp}{ .mfi addl GR_ad_z_1 = @ltoff(Constants_Z_1#),gp nop.f 0 nop.i 0};;{ .mfi setf.exp FR_Half = GR_Half // construct 0.5 fclass.m p7,p0 = FR_Arg, 0x09 // if arg + denorm and GR_ArgExp = GR_ExpMask, GR_ArgExp // select exp}{ .mfb ld8 GR_ad_z_1 = [GR_ad_z_1] // Get pointer to Constants_Z_1 nop.f 0 nop.b 0};;{ .mfi ldfe FR_C9 = [GR_Poly_C_79],16 // load C9 fclass.m p10,p0 = FR_Arg, 0x0a // if arg - denorm cmp.gt p8, p0 = GR_TwoN7, GR_ArgExp // if arg < 2^-7 ('near 0')}{ .mfb cmp.le p9, p0 = GR_TwoP63, GR_ArgExp // if arg > 2^63 ('huges')(p6) fma.s0 FR_Res = FR_Arg,f1,FR_Arg // r = a + a(p6) br.ret.spnt b0 // return };;// (X^2 + 1) computation{ .mfi(p8) ldfe FR_C5 = [GR_Poly_C_35],16 // load C5 fms.s1 FR_Tmp = f1, f1, FR_P2 // Tmp = 1 - p2 add GR_ad_tbl_1 = 0x040, GR_ad_z_1 // Point to Constants_G_H_h1}{ .mfb(p8) ldfe FR_C7 = [GR_Poly_C_79],16 // load C7(p7) fnma.s0 FR_Res = FR_Arg,FR_Arg,FR_Arg // r = a - a*a(p7) br.ret.spnt b0 // return};;{ .mfi(p8) ldfe FR_C3 = [GR_Poly_C_35],16 // load C3 fcmp.lt.s1 p11, p12 = FR_Arg, f0 // if arg is negative add GR_ad_q = -0x60, GR_ad_z_1 // Point to Constants_P}{ .mfb add GR_ad_z_2 = 0x140, GR_ad_z_1 // Point to Constants_Z_2(p10) fma.s0 FR_Res = FR_Arg,FR_Arg,FR_Arg // r = a + a*a(p10) br.ret.spnt b0 // return};;{ .mfi add GR_ad_tbl_2 = 0x180, GR_ad_z_1 // Point to Constants_G_H_h2 frsqrta.s1 FR_Rcp, p0 = FR_P2 // Rcp = 1/p2 reciprocal appr. add GR_ad_tbl_3 = 0x280, GR_ad_z_1 // Point to Constants_G_H_h3}{ .mfi nop.m 0 fms.s1 FR_P2L = FR_AX, FR_AX, FR_X2 //low part of p2=fma(X*X-p2) mov GR_Bias = 0x0FFFF // Create exponent bias};;{ .mfb nop.m 0(p9) fms.s1 FR_XLog_Hi = FR_Two, FR_AX, f0 // Hi of log1p arg = 2*X - 1(p9) br.cond.spnt huges_logl // special version of log1p};;{ .mfb ldfe FR_log2_hi = [GR_ad_q],16 // Load log2_hi(p8) fma.s1 FR_X3 = FR_X2, FR_Arg, f0 // x^3 = x^2 * x(p8) br.cond.spnt near_0 // Go to near 0 branch};;{ .mfi ldfe FR_log2_lo = [GR_ad_q],16 // Load log2_lo nop.f 0 nop.i 0};;{ .mfi ldfe FR_Q4 = [GR_ad_q],16 // Load Q4 fma.s1 FR_Tmp = FR_Tmp, f1, FR_X2 // Tmp = Tmp + x^2 mov GR_exp_mask = 0x1FFFF // Create exponent mask};;{ .mfi ldfe FR_Q3 = [GR_ad_q],16 // Load Q3 fma.s1 FR_GG = FR_Rcp, FR_P2, f0 // g = Rcp * p2 // 8 bit Newton Raphson iteration nop.i 0}{ .mfi nop.m 0 fma.s1 FR_HH = FR_Half, FR_Rcp, f0 // h = 0.5 * Rcp nop.i 0};;{ .mfi ldfe FR_Q2 = [GR_ad_q],16 // Load Q2 fnma.s1 FR_EE = FR_GG, FR_HH, FR_Half // e = 0.5 - g * h nop.i 0}{ .mfi nop.m 0 fma.s1 FR_P2L = FR_Tmp, f1, FR_P2L // low part of p2 = Tmp + p2l nop.i 0};;{ .mfi ldfe FR_Q1 = [GR_ad_q] // Load Q1 fma.s1 FR_GG = FR_GG, FR_EE, FR_GG // g = g * e + g // 16 bit Newton Raphson iteration nop.i 0}{ .mfi nop.m 0 fma.s1 FR_HH = FR_HH, FR_EE, FR_HH // h = h * e + h nop.i 0};;{ .mfi nop.m 0 fnma.s1 FR_EE = FR_GG, FR_HH, FR_Half // e = 0.5 - g * h nop.i 0};;{ .mfi nop.m 0 fma.s1 FR_GG = FR_GG, FR_EE, FR_GG // g = g * e + g // 32 bit Newton Raphson iteration nop.i 0}{ .mfi nop.m 0 fma.s1 FR_HH = FR_HH, FR_EE, FR_HH // h = h * e + h nop.i 0};;{ .mfi nop.m 0 fnma.s1 FR_EE = FR_GG, FR_HH, FR_Half // e = 0.5 - g * h nop.i 0};;{ .mfi nop.m 0 fma.s1 FR_GG = FR_GG, FR_EE, FR_GG // g = g * e + g // 64 bit Newton Raphson iteration nop.i 0}{ .mfi nop.m 0 fma.s1 FR_HH = FR_HH, FR_EE, FR_HH // h = h * e + h nop.i 0};;{ .mfi nop.m 0 fnma.s1 FR_DD = FR_GG, FR_GG, FR_P2 // Remainder d = g * g - p2 nop.i 0}{ .mfi nop.m 0 fma.s1 FR_XLog_Hi = FR_AX, f1, FR_GG // bh = z + gh nop.i 0};;{ .mfi nop.m 0 fma.s1 FR_DD = FR_DD, f1, FR_P2L // add p2l: d = d + p2l nop.i 0};;{ .mfi getf.sig GR_signif = FR_XLog_Hi // Get significand of x+1 fmerge.ns FR_Neg_One = f1, f1 // Form -1.0 mov GR_exp_2tom7 = 0x0fff8 // Exponent of 2^-7};;{ .mfi nop.m 0 fma.s1 FR_GL = FR_DD, FR_HH, f0 // gl = d * h extr.u GR_Index1 = GR_signif, 59, 4 // Get high 4 bits of signif}{ .mfi nop.m 0 fma.s1 FR_XLog_Hi = FR_DD, FR_HH, FR_XLog_Hi // bh = bh + gl nop.i 0};;{ .mmi shladd GR_ad_z_1 = GR_Index1, 2, GR_ad_z_1 // Point to Z_1 shladd GR_ad_tbl_1 = GR_Index1, 4, GR_ad_tbl_1 // Point to G_1 extr.u GR_X_0 = GR_signif, 49, 15 // Get high 15 bits of signif.};;{ .mmi ld4 GR_Z_1 = [GR_ad_z_1] // Load Z_1 nop.m 0 nop.i 0};;{ .mmi ldfps FR_G, FR_H = [GR_ad_tbl_1],8 // Load G_1, H_1 nop.m 0 nop.i 0};;{ .mfi nop.m 0 fms.s1 FR_XLog_Lo = FR_GG, f1, FR_XLog_Hi // bl = gh - bh pmpyshr2.u GR_X_1 = GR_X_0,GR_Z_1,15 // Get bits 30-15 of X_0 * Z_1};;// WE CANNOT USE GR_X_1 IN NEXT 3 CYCLES BECAUSE OF POSSIBLE 10 CLOCKS STALL!// "DEAD" ZONE!{ .mfi nop.m 0 nop.f 0 nop.i 0};;{ .mfi nop.m 0 fmerge.se FR_S_hi = f1,FR_XLog_Hi // Form |x+1| nop.i 0};;{ .mmi getf.exp GR_N = FR_XLog_Hi // Get N = exponent of x+1 ldfd FR_h = [GR_ad_tbl_1] // Load h_1 nop.i 0};;{ .mfi nop.m 0 nop.f 0 extr.u GR_Index2 = GR_X_1, 6, 4 // Extract bits 6-9 of X_1 };;{ .mfi shladd GR_ad_tbl_2 = GR_Index2, 4, GR_ad_tbl_2 // Point to G_2 fma.s1 FR_XLog_Lo = FR_XLog_Lo, f1, FR_AX // bl = bl + x mov GR_exp_2tom80 = 0x0ffaf // Exponent of 2^-80}{ .mfi shladd GR_ad_z_2 = GR_Index2, 2, GR_ad_z_2 // Point to Z_2 nop.f 0 sub GR_N = GR_N, GR_Bias // sub bias from exp};;{ .mmi ldfps FR_G2, FR_H2 = [GR_ad_tbl_2],8 // Load G_2, H_2 ld4 GR_Z_2 = [GR_ad_z_2] // Load Z_2 sub GR_minus_N = GR_Bias, GR_N // Form exponent of 2^(-N)};;{ .mmi ldfd FR_h2 = [GR_ad_tbl_2] // Load h_2 nop.m 0 nop.i 0};;{ .mmi setf.sig FR_float_N = GR_N // Put integer N into rightmost sign setf.exp FR_2_to_minus_N = GR_minus_N // Form 2^(-N) pmpyshr2.u GR_X_2 = GR_X_1,GR_Z_2,15 // Get bits 30-15 of X_1 * Z_2};;// WE CANNOT USE GR_X_2 IN NEXT 3 CYCLES ("DEAD" ZONE!) // BECAUSE OF POSSIBLE 10 CLOCKS STALL!// So we can negate Q coefficients there for negative values{ .mfi nop.m 0(p11) fma.s1 FR_Q1 = FR_Q1, FR_Neg_One, f0 // Negate Q1 nop.i 0}{ .mfi nop.m 0 fma.s1 FR_XLog_Lo = FR_XLog_Lo, f1, FR_GL // bl = bl + gl nop.i 0};;{ .mfi nop.m 0(p11) fma.s1 FR_Q2 = FR_Q2, FR_Neg_One, f0 // Negate Q2 nop.i 0};;{ .mfi nop.m 0(p11) fma.s1 FR_Q3 = FR_Q3, FR_Neg_One, f0 // Negate Q3 nop.i 0};;{ .mfi nop.m 0(p11) fma.s1 FR_Q4 = FR_Q4, FR_Neg_One, f0 // Negate Q4 extr.u GR_Index3 = GR_X_2, 1, 5 // Extract bits 1-5 of X_2};;{ .mfi shladd GR_ad_tbl_3 = GR_Index3, 4, GR_ad_tbl_3 // Point to G_3 nop.f 0 nop.i 0};;{ .mfi ldfps FR_G3, FR_H3 = [GR_ad_tbl_3],8 // Load G_3, H_3 nop.f 0 nop.i 0};;{ .mfi ldfd FR_h3 = [GR_ad_tbl_3] // Load h_3 fcvt.xf FR_float_N = FR_float_N nop.i 0};;{ .mfi nop.m 0 fmpy.s1 FR_G = FR_G, FR_G2 // G = G_1 * G_2 nop.i 0}{ .mfi nop.m 0 fadd.s1 FR_H = FR_H, FR_H2 // H = H_1 + H_2⌨️ 快捷键说明
复制代码
Ctrl + C
搜索代码
Ctrl + F
全屏模式
F11
切换主题
Ctrl + Shift + D
显示快捷键
?
增大字号
Ctrl + =
减小字号
Ctrl + -